<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Vikram Vishal on Medium]]></title>
        <description><![CDATA[Stories by Vikram Vishal on Medium]]></description>
        <link>https://medium.com/@vikram.vishal?source=rss-ebf8285a9567------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*kQVbfHxVbXFQGa5-0K-qQw.jpeg</url>
            <title>Stories by Vikram Vishal on Medium</title>
            <link>https://medium.com/@vikram.vishal?source=rss-ebf8285a9567------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 07:48:29 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@vikram.vishal/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Architecture Nobody Owns: Context Is Becoming an Enterprise Supply Chain]]></title>
            <link>https://medium.com/@vikram.vishal/the-architecture-nobody-owns-context-is-becoming-an-enterprise-supply-chain-8c8136e3559e?source=rss-ebf8285a9567------2</link>
            <guid isPermaLink="false">https://medium.com/p/8c8136e3559e</guid>
            <category><![CDATA[ai-architecture]]></category>
            <category><![CDATA[context-engineering]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[rags]]></category>
            <dc:creator><![CDATA[Vikram Vishal]]></dc:creator>
            <pubDate>Fri, 18 Sep 2026 18:00:05 GMT</pubDate>
            <atom:updated>2026-09-18T18:00:05.627Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Enterprise AI will become more reliable when we stop treating context as prompt material and start treating it as a governed, observable flow of information.</em></p><p>When an enterprise AI assistant gives the wrong answer, the model is usually the first thing blamed. Sometimes that’s fair. But imagine a credit manager asks an internal assistant:</p><blockquote><em>“Can this customer receive the requested credit extension?”</em></blockquote><p>The answer is wrong. The natural reaction is to inspect the LLM, adjust the system prompt, perhaps swap the model, and run the evaluation again. Yet the model may have reasoned perfectly over what it was given.</p><p>The customer status could have been stale in CRM. An ERP synchronization might have been delayed. A policy PDF may have failed extraction. Entity resolution might have attached the wrong subsidiary. The current policy could have missed the latest indexing cycle. Retrieval might have found it but reranking pushed it below the context cutoff. Or a summarization step may have removed the exception clause that actually mattered.</p><p>The visible failure happens at the end. The causal failure can exist almost anywhere upstream.</p><p>That is why I think enterprises need a different mental model for context.</p><h3><strong>Context is becoming a supply chain.</strong></h3><p>Not as a metaphor for everything involving AI, but as a precise architectural observation: information moves through a sequence of transformations, custody changes, quality boundaries and policy decisions before a model ever sees it. And every one of those boundaries can fail differently.</p><h3>The Mental Model We Are Missing</h3><p>Much of enterprise AI architecture is still drawn around a surprisingly small box: User — RAG — LLM — Answer.</p><p>It is useful for explaining a prototype. It becomes dangerously compressed when explaining a production system. Even “RAG pipeline” hides more than it reveals.</p><p>The original retrieval-augmented generation work explicitly combined parametric model memory with external, non-parametric memory, partly addressing limitations around updating knowledge and providing provenance. Production enterprise implementations have since added far more machinery around that basic idea: document processors, ACL propagation, metadata enrichment, hybrid search, rerankers, caches, query rewriting, entity resolution, summarizers, policy engines and orchestration.</p><p>So the actual path looks closer to this:</p><blockquote>Raw enterprise information<br>Ingestion<br>Extraction<br>Normalization<br>Enrichment<br>Segmentation<br>Indexing<br>Retrieval<br>Ranking<br>Assembly<br>Compression<br>Prompt construction<br>Model consumption<br>Response or action.</blockquote><p>Every stage matters. Some stages preserve information. Others transform it. Some deliberately discard information. Others infer metadata, derive relationships or reorder evidence. And almost none of them are guaranteed to have the same owner.</p><p>A physical supply chain is useful here because it forces four questions that AI architecture often avoids:</p><blockquote>Where did something originate?</blockquote><blockquote>What happened to it?</blockquote><blockquote>Who had custody?</blockquote><blockquote>Where did quality deteriorate?</blockquote><p>For enterprise AI, I would add a fifth:</p><blockquote>Was that information even permitted to reach this consumer?</blockquote><p>That is the beginning of the enterprise context supply chain.</p><h3>Context Does Not Arrive. It Is Manufactured.</h3><p>We often speak about “giving the model context” as if context were waiting somewhere, fully formed, for an application to retrieve. It isn’t.</p><p>Context is manufactured. A 70-page credit policy begins as a document in a content-management system. A connector acquires it. A parser extracts text. Tables may be reconstructed. Headers and footers may disappear. Sections get segmented. Metadata is attached. Entities are recognized. Embeddings are generated. An index is updated.</p><p>Later, a user asks a question.</p><p>That question may be rewritten before search. Several retrieval mechanisms may execute. Security filters reduce the candidate set. A reranker changes the ordering. Evidence is selected to fit a token budget. Long passages may be summarized. System instructions, conversation history and tool results compete for context-window space. Only then does an LLM receive something it can reason over.</p><p>Consider what can happen along the way.</p><blockquote>Fresh information can become stale because an index refresh missed its SLA.</blockquote><blockquote>Authoritative information can lose authority because source metadata was stripped during extraction.</blockquote><blockquote>A paragraph can remain textually correct but become semantically misleading because chunking separated a rule from its exception.</blockquote><blockquote>A highly relevant document can be correctly retrieved and then removed because identity propagation used an obsolete entitlement.</blockquote><blockquote>Two individually correct records can create an incorrect picture because entity resolution assumed they represented the same customer.</blockquote><blockquote>Compression can preserve the broad meaning of a policy while dropping exactly the clause required for a particular decision.</blockquote><p>These are not primarily prompt-engineering problems. They are transformation problems.</p><p>That distinction matters because remediation follows causality. Better instructions to the model cannot repair evidence that was never indexed. A stronger model cannot reason over a policy that security filtering removed. A more sophisticated reranker cannot recover a document that failed ingestion.</p><p><strong>The model cannot reason over evidence it never received.</strong></p><h3>The Enterprise Context Supply Chain</h3><p>A useful conceptual flow looks like this:</p><blockquote>User question<br>AI recommendation<br>Context supplied to model<br>Policy passages included<br>Retrieved candidate set<br>Ranking decision<br>Index version<br>Extracted document representation<br>Ingestion event<br>Original policy</blockquote><p><em>The important part of this flow is the boundaries between each step.</em></p><p>At the <strong>source layer</strong>, systems of record, SaaS applications, documents, events and databases establish the underlying evidence.</p><p>The <strong>acquisition layer</strong> moves that evidence through CDC, APIs, streams, connectors or crawlers.</p><p>The <strong>transformation layer</strong> changes its representation through parsing, normalization, entity resolution, enrichment and segmentation.</p><p>The <strong>knowledge and index layer</strong> makes it addressable through search indexes, vector stores, knowledge graphs, metadata repositories and caches.</p><p>The <strong>retrieval and selection layer</strong> decides what might matter, using lexical search, semantic retrieval, hybrid techniques, filters, reranking and policy enforcement.</p><p>The <strong>context assembly layer</strong> decides what the model will actually see. Evidence gets selected, ordered, compressed and fitted into a finite context budget.</p><p>The <strong>reasoning layer</strong> interprets that assembled state through models, agents, tools and planning logic.</p><p>Finally, the <strong>action layer</strong> converts reasoning into something consequential: an answer, recommendation, case decision, workflow transition or transaction.</p><p>Each layer therefore deserves five questions:</p><blockquote>What transformation occurred?</blockquote><blockquote>Who owned it?</blockquote><blockquote>What contract governed the handoff?</blockquote><blockquote>How can it fail?</blockquote><blockquote>How would we know?</blockquote><p>That is a much richer architecture conversation than “Which vector database are we using?”</p><h3>Every Transformation Creates a Failure Boundary</h3><p>Return to the credit-extension question. Suppose the assistant recommends approving the extension even though a revised corporate policy prohibits it.</p><p>There are many plausible explanations.</p><p>The CRM may contain an outdated customer risk category.</p><p>The policy repository may contain version 8, while ingestion still exposes version 7.</p><p>The ingestion pipeline may have succeeded technically but skipped the revised attachment.</p><p>The extraction engine may have failed on a table containing the relevant threshold.</p><p>Chunking may have separated “Extensions above €2 million require CFO approval” from a preceding definition restricting which entities qualify.</p><p>The embedding index may be 18 hours behind the source.</p><p>Retrieval may return the correct policy but rank three superficially similar documents above it.</p><p>Authorization filtering may remove the authoritative regional policy.</p><p>Context assembly may include the policy but exclude the governing exception because the token budget was exhausted.</p><p>A summarizer may turn “must not approve without CFO authorization” into “requires additional approval,” subtly changing the operational meaning.</p><p>And yes, the model itself may perform an unsupported inference.</p><p>The point is not that models are innocent.</p><p>The point is that <strong>“the AI was wrong”</strong> is not a useful root-cause category<strong>.</strong></p><p>This is familiar territory in distributed systems. A timeout observed by an application may originate in a database, network, dependency or downstream service. Mature engineering organizations learned to trace the transaction rather than staring only at the final exception.</p><p>Enterprise AI needs the equivalent for context.</p><h3>A Context Failure Trace</h3><p>Imagine tracing our bad credit recommendation backward:</p><blockquote>User question<br>AI recommendation<br>Context supplied to model<br>Policy passages included<br>Retrieved candidate set<br>Ranking decision<br>Index version<br>Extracted document representation<br>Ingestion event<br>Original policy</blockquote><p>The investigation reveals something surprising.</p><p>The LLM did not hallucinate. The retriever actually behaved correctly. The reranker behaved correctly too. The index contained policy version 7 because a connector stopped processing updates 36 hours earlier after an authentication change. Changing the model would have achieved nothing.</p><p>That is the diagnostic payoff of treating context as a supply chain.</p><h3>From Data Lineage to Context Lineage</h3><p>Enterprises already understand data lineage. If a financial-reporting field looks suspicious, lineage can tell us which source column fed it, which transformations modified it and which downstream datasets consumed it.</p><p>Context lineage needs to answer a different question.</p><p>Data lineage asks:</p><blockquote><em>Where did this data come from?</em></blockquote><p>Context lineage asks:</p><blockquote><em>Why did this specific information reach this model, in this representation, at this moment, for this user and task?</em></blockquote><p>That requires more than tracing a document back to SharePoint. It potentially includes source and source version, ingestion time, extraction method, segment derivation, enrichment history, embedding and index versions, the retrieval query, candidate scores, reranking, filters, authorization decisions, assembly rules, compression, token-budget decisions, prompt placement, model version and any downstream action.</p><p>Not every application needs every field. The architectural principle is more important than the schema:</p><p><strong>We need to preserve enough lineage to reconstruct the evidence path of a consequential AI interaction.</strong></p><p>This becomes especially important because context selection itself is computational behavior. A document may exist, be current, be indexed correctly and still not influence the answer because it ranked 12th when only eight passages were admitted.</p><p>That selection decision is part of the effective state of the AI system. Traditional data lineage usually does not capture it. Context lineage must.</p><h3>Context Contracts and Context Observability</h3><p>Once context crosses several teams, another issue appears: ownership.</p><blockquote><strong>Source</strong><br>Owner: Business application team<br>Typical failure: Incorrect or stale record</blockquote><blockquote><strong>Ingestion</strong><br>Owner: Data / platform engineering<br>Typical failure: Missing, delayed, or partially synchronized data</blockquote><blockquote><strong>Extraction</strong><br>Owner: Document / AI platform team<br>Typical failure: Parsing errors, lost structure, or incorrectly interpreted content</blockquote><blockquote><strong>Enrichment</strong><br>Owner: Knowledge / ML team<br>Typical failure: Wrong entity resolution, metadata, classification, or relationship</blockquote><blockquote><strong>Indexing</strong><br>Owner: Search / AI platform team<br>Typical failure: Stale, incomplete, or incorrectly versioned index</blockquote><blockquote><strong>Retrieval</strong><br>Owner: AI application team<br>Typical failure: Relevant evidence is not retrieved</blockquote><blockquote><strong>Ranking</strong><br>Owner: Search / ML team<br>Typical failure: The right evidence exists but is pushed below less relevant content</blockquote><blockquote><strong>Context Assembly</strong><br>Owner: AI platform / application team<br>Typical failure: Critical evidence is excluded, compressed, reordered, or lost to token limits</blockquote><blockquote><strong>Model Reasoning</strong><br>Owner: Model / application layer<br>Typical failure: Unsupported inference, misinterpretation, or reasoning error</blockquote><blockquote><strong>Action</strong><br>Owner: Workflow / application team<br>Typical failure: An incorrect recommendation becomes an incorrect downstream execution</blockquote><p>The important point is what this chain reveals: <strong>the user experiences one AI failure, but the enterprise may be dealing with ten different failure domains owned by ten different parts of the organization.</strong></p><p>That is why “who owns the AI answer?” is often the wrong question. The more useful question is:</p><blockquote><strong><em>Who owns the integrity of context as it moves from source to decision?</em></strong></blockquote><p>This is where I think two architectural abstractions will become increasingly useful: context contracts and context observability.</p><p>A <strong>context contract</strong> would make the handoff between stages explicit. Instead of passing an opaque blob of text downstream, a transformation could preserve information about provenance, semantic identity, freshness, authority, permissions, transformation history, confidence, intended usage, expiration and expected quality.</p><p>The analogy is straightforward. APIs made service boundaries explicit. Data contracts made producer-consumer expectations around data more explicit. Context contracts may become a similar abstraction for AI knowledge flows.</p><p>I am using the term here as an architectural proposition, not suggesting that a universal industry standard already exists.</p><p>The second abstraction is <strong>context observability</strong>: the ability to inspect the provenance, transformations, selection, exclusion, freshness, policy decisions and quality of the context supplied to an AI system.</p><p>Observability matters because an end-to-end evaluation score cannot tell you why quality deteriorated.</p><p>A useful operating model might separately examine context freshness, retrieval coverage, provenance completeness, transformation errors, context precision, policy-filter behavior, assembly latency, stale-context frequency, evidence utilization and trace completeness.</p><p>The point is not to manufacture another dashboard. It is to expose which part of the system is degrading.</p><h3>Agents Make the Supply Chain Dynamic</h3><p>RAG systems make this problem visible. AI agents make it harder.</p><p>An agent can assemble working context from conversation history, enterprise search, databases, APIs, tool responses, workflow state, long-term memory, user identity, policy engines, other agents and real-time events.</p><blockquote>Some of that context exists for milliseconds.</blockquote><blockquote>Some persists across sessions.</blockquote><blockquote>Some authoritative.</blockquote><blockquote>Some merely suggestive.</blockquote><blockquote>Some originates inside the enterprise.</blockquote><blockquote>Some arrives from an external tool during execution.</blockquote><blockquote>And some may itself contain instructions.</blockquote><p>That last point has security implications. OWASP explicitly notes that RAG does not eliminate prompt-injection risk and describes attacks involving manipulated external content and poisoned retrieval sources.</p><p>Once a system can take actions, context provenance is no longer only about answer quality.</p><p>Imagine an agent that receives a tool response saying a supplier bank account changed, combines it with an email requesting an urgent payment, and initiates a workflow.</p><blockquote>Which input was authoritative?</blockquote><blockquote>Which system supplied it?</blockquote><blockquote>What authorization applied?</blockquote><blockquote>Was the tool response current?</blockquote><blockquote>Did retrieved content contain untrusted instructions?</blockquote><blockquote>Which evidence caused the agent to choose that action?</blockquote><p>With agents, context becomes <strong>dynamic, assembled, permissioned and transient</strong>. That makes the supply-chain problem more important, not less.</p><h3>Context Becomes a Governance Domain</h3><p>Enterprises already govern data, models, identities, applications and APIs. Context increasingly sits across all of them. That makes it a governable object in its own right.</p><p>Model governance alone cannot answer whether sensitive data should have entered a prompt. Data governance alone cannot explain why one document was selected and another excluded. IAM can determine whether a user can access a source, but it does not necessarily explain how that entitlement propagated through indexing, retrieval, caching and model consumption.</p><p>NIST’s AI Risk Management Framework deliberately treats AI risk across the lifecycle rather than reducing it to model behavior, organizing practices around Govern, Map, Measure and Manage. Context architecture is one of the places where that lifecycle view becomes operationally concrete.</p><blockquote>Security matters.</blockquote><blockquote>Privacy matters.</blockquote><blockquote>Freshness matters.</blockquote><blockquote>Retention matters.</blockquote><blockquote>Authorization matters.</blockquote><blockquote>Provenance matters.</blockquote><p>But the harder governance question is reconstructability.</p><p><em>If an AI system makes a consequential recommendation, can the enterprise reconstruct what the model actually knew at that moment?</em></p><blockquote>Not what existed in the database.</blockquote><blockquote>Not what should have been retrieved.</blockquote><blockquote>What actually crossed the final context boundary.</blockquote><p>That is a much higher bar. And for regulated, financial, safety-sensitive or highly automated workflows, it may become one of the most important ones.</p><h3>What Enterprise Architects Should Do Now</h3><p>None of this requires creating another giant enterprise architecture program. It does require changing what we draw, instrument and assign ownership to.</p><ol><li><strong>Model context as a first-class architecture domain.</strong> Stop collapsing ingestion, knowledge, retrieval and prompt construction into a single “RAG” component. Show the transformations that materially affect meaning, authority, security or freshness.</li><li><strong>Create explicit ownership across context transformations.</strong> There may never be one owner for end-to-end context quality. That makes stage accountability more important. Define who owns freshness, extraction fidelity, index integrity, retrieval behavior, authorization propagation and assembly logic.</li><li><strong>Design provenance into the pipeline.</strong> Reconstructing context after an incident is painful if intermediate systems discarded source versions, transformation metadata or retrieval decisions. Provenance is much cheaper when designed into the flow.</li><li><strong>Observe context independently from model output.</strong> A good answer does not prove that retrieval is healthy, just as a bad answer does not prove the model failed. Inspect the evidence path directly.</li><li><strong>Evaluate the context-to-decision chain.</strong> Model benchmarks have value, but production systems need tests that traverse source freshness, transformation quality, retrieval, policy enforcement, context assembly, reasoning and action.</li></ol><p>This changes where architects place boundaries. The interesting boundary is no longer just between application and model. It is between <strong>enterprise knowledge and machine reasoning</strong>.</p><h3>Context Is Infrastructure Now</h3><p>Enterprise AI discussions still spend disproportionate energy on models. Models matter enormously.</p><p>But production reliability increasingly depends on everything that determines what a model is allowed to know, able to find and ultimately given to reason over.</p><p>That infrastructure is fragmented today because its pieces grew out of different disciplines: data engineering, search, knowledge management, security, ML platforms, application development and governance.</p><p>AI is forcing those disciplines to meet in the same execution path. We need an architectural model that acknowledges that reality.</p><p>The next durable advantage in enterprise AI may come less from possessing a marginally stronger model and more from building the machinery that can <strong>manufacture the right context, preserve its meaning, enforce its boundaries, trace its journey and deliver it at the precise moment reasoning depends on it.</strong></p><p>When context reaches that level of operational importance, it is no longer prompt material. It is infrastructure. And infrastructure needs a supply chain.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8c8136e3559e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Evaluation Is Fragmented by Design. Accountability Shouldn’t Be.]]></title>
            <link>https://medium.com/@vikram.vishal/ai-evaluation-is-fragmented-by-design-accountability-shouldnt-be-2fb1670cb3dd?source=rss-ebf8285a9567------2</link>
            <guid isPermaLink="false">https://medium.com/p/2fb1670cb3dd</guid>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[technology-leadership]]></category>
            <category><![CDATA[responsible-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <dc:creator><![CDATA[Vikram Vishal]]></dc:creator>
            <pubDate>Tue, 08 Sep 2026 15:57:02 GMT</pubDate>
            <atom:updated>2026-09-08T15:57:02.191Z</atom:updated>
            <content:encoded><![CDATA[<blockquote><strong>Engineering can prove performance, business can prove value, and risk can challenge exposure — but production readiness requires someone to own the decision that joins them.</strong></blockquote><p>An AI customer-service assistant is two weeks from production. Engineering says the system has cleared its benchmark suite. Hallucination rates are within threshold. Latency is acceptable. The prompt and retrieval stack are versioned. The product team points to strong pilot feedback. Users are resolving cases faster. The business sponsor wants to launch before the next peak period. Security has approved the architecture. Legal has documented the remaining concerns. Privacy has signed off the data flow. Risk has classified the use case and recommended additional monitoring.</p><p>Then someone asks the awkward question.</p><blockquote>Who is accountable for saying this system is actually good enough to deploy?</blockquote><p>The room becomes less certain. Each team can defend its part of the decision. But none of that is the same as saying: <strong>we have enough evidence to put this AI system into the real world, under these conditions, with these residual risks.</strong></p><p>That gap is becoming one of the least discussed problems in enterprise AI. AI evaluation is often distributed correctly. AI accountability is often distributed incorrectly.</p><h3>The Problem Is Not Lack of Testing</h3><p>Most serious enterprises already test AI systems from several angles.</p><p>Machine-learning and engineering teams measure accuracy, hallucination rates, robustness, latency, throughput, reproducibility and failure modes. Their question is sensible: <strong>Does the system perform as designed?</strong></p><p>Product and business teams look at adoption, workflow impact, decision quality, customer experience, productivity and economic value. Their question is: <strong>Does this AI create enough real-world value to justify its use?</strong></p><p>Risk, compliance, legal, privacy and security examine another failure surface: regulatory exposure, privacy, explainability, misuse, unfair outcomes, security, reputation and control effectiveness. Their question is: <strong>What can go wrong, and is the remaining exposure acceptable?</strong></p><p>Data teams own critical parts of the evidence base: lineage, provenance, permissions, freshness, representativeness, leakage and quality. A sophisticated evaluation process built on weak evidence is still weak evaluation.</p><p>AI platform, MLOps and operations teams focus on deployment integrity, observability, rollback, resilience, cost, drift and incidents. Their question is: <strong>Can we operate this system safely and consistently after launch?</strong></p><p>Every one of these perspectives is legitimate. None is sufficient on its own.</p><p>The failure occurs when organizations confuse responsibility for measuring part of the system with accountability for deciding whether the system as a whole is ready to operate.</p><h3>Measurement Ownership Is Not Outcome Accountability</h3><p>Engineering can prove that a RAG assistant answers benchmark questions accurately. That does not establish whether it improves customer outcomes.</p><p>A business team can prove that employees like a copilot. That does not establish whether it creates unacceptable privacy or control exposure.</p><p>Risk can identify material concerns and challenge weak controls. That does not automatically make risk the owner of the commercial or operational consequence of deploying the system.</p><p>Operations can prove that an agentic workflow is observable, resilient and reversible. That does not establish that the agent is making decisions it should be permitted to make.</p><p>This is why the enterprise question cannot stop at, “Who tests the model?”</p><p>The harder question is: <strong>Who has the authority and accountability to decide whether the accumulated evidence is sufficient for this AI system to operate?</strong></p><p>That is an operating-model question. And it requires a different kind of framework.</p><h3>The AI Evaluation Accountability Matrix</h3><p>The AI Evaluation Accountability Matrix is not another testing checklist and should not become another RACI spreadsheet that nobody reads after approval. It is a decision-rights model. It asks five recurring questions across the lifecycle.</p><p><strong>Who owns the outcome?</strong><br>The Outcome Owner is accountable for why the system exists, what success means, which users are affected, acceptable trade-offs and real-world failure.</p><p><strong>Who owns the evaluation?</strong><br>The Evaluation Owner is accountable for methodology and evidence: tests, benchmarks, acceptance criteria, datasets, repeatability and evidence quality. In practice, this is often federated.</p><p><strong>Who owns residual risk?</strong><br>The Risk Owner is the person with authority to accept the remaining exposure after controls are applied. Risk functions may identify, assess and challenge that exposure, but they should not automatically inherit ownership of every risk created by a business decision to use AI.</p><p><strong>Who has decision authority?</strong><br>This is the role that can say: ship, ship with constraints, run a limited pilot, hold, escalate or stop.</p><p>It is also the role most frequently missing.</p><p><strong>Who owns monitoring?</strong><br>The Monitoring Owner determines whether release assumptions remain true in production and who must act when they do not.</p><p>The useful part of the matrix is not the five roles themselves. It is how those roles change as the AI system moves through its lifecycle.</p><h3>Stage 1: Problem Framing</h3><p>Evaluation begins before a benchmark exists. At problem framing, the business or product owner defines the outcome, affected users and unacceptable consequences. Engineering challenges whether AI is appropriate. Risk and legal identify constraints early enough to shape the design. The most important decision is often ignored: <strong>Should AI be used here at all?</strong></p><p>A deterministic rule or conventional software feature may be better.</p><p>This is <strong>Gate 1: use-case approval</strong>.</p><p>The organization should not move forward until it knows what outcome is being optimized, who owns that outcome, and what failure would matter. Many later evaluation disputes are simply deferred problem-definition disputes.</p><h3>Stage 2: Data Readiness</h3><p>The evaluation owner now depends heavily on data leadership. Are test datasets representative? Are permissions legitimate? Is lineage known? Has leakage occurred? Does evaluation data resemble the population the system will face?</p><p>Advanced models are still too often tested against evidence bases nobody fully trusts. Before evaluating the intelligence, evaluate the evidence.</p><p>This is <strong>Gate 2: evaluation readiness</strong>. Success criteria, failure thresholds and trustworthy evaluation data should exist before teams start arguing about scores.</p><h3>Stage 3: Model or System Selection</h3><p>The question is no longer merely “Which model performs best?”. For a foundation-model application, the relevant unit may be the full system: model, prompt, retrieval, tools, policies, memory and orchestration. For a copilot, interaction design may matter as much as benchmark quality. For an agentic workflow, the critical variable may be authority: what can the agent observe, decide and write back?</p><p>Selection therefore needs a joint technical, economic and operational judgment. A slightly weaker model may be preferable if it is easier to constrain, cheaper to operate, easier to observe or more stable across versions. Evaluation should reflect the system enterprises will actually run, not the model in isolation.</p><h3>Stage 4: Offline Technical Evaluation</h3><p>Here engineering and ML teams should lead. They own benchmark design, quality testing, robustness checks, adversarial testing where justified, latency, cost, reproducibility and failure analysis. The standards should be demanding. But benchmark success is not a proxy for production readiness.</p><p>A credit-analysis assistant can generate accurate summaries and still encourage analysts to over-trust incomplete evidence. A document-intelligence pipeline can exceed extraction thresholds while downstream users give uncertain fields more authority than they deserve.</p><p>Technical evaluation establishes capability. It does not grant permission to operate.</p><h3>Stage 5: Business Evaluation</h3><p>This stage belongs to people who understand the workflow consequence.</p><blockquote>Does a customer-service copilot improve resolution quality, or simply make agents faster at giving mediocre answers?</blockquote><blockquote>Does an underwriting assistant improve analysis, or mainly reduce reading time?</blockquote><blockquote>Does an employee assistant reduce effort without quietly shifting work into verification?</blockquote><p>Business evaluation should test the human-plus-AI workflow, not just the AI output. Outcome owners therefore need explicit acceptance criteria of their own. Engineering cannot define business usefulness on the business’s behalf.</p><h3>Stage 6: Risk and Control Evaluation</h3><p>Now independent challenge matters. Security, privacy, compliance, legal, model risk and Responsible AI teams should challenge first-line assumptions. Business, product and engineering own the system they are trying to operate. Specialist risk functions challenge whether exposure is understood and controlled. Independent assurance may examine the process where justified.</p><p>The second line should not become the de facto product owner. Its job is to improve the quality of the decision, not to absorb accountability for the decision. Residual risk should be named, bounded and accepted by someone with the authority to own its consequence. “Risk accepted” is not enough. Which risk? By whom? Under which assumptions? With which controls? Until when?</p><h3>Stage 7: Release Is Where Governance Becomes Real</h3><p>This is the point at which fragmented evaluation must converge. By now, technical, business, operational, data and risk evidence should be available. Someone still has to decide.</p><p>That decision should produce one of four outcomes:</p><ul><li><strong>Ship.</strong></li><li><strong>Ship with constraints.</strong></li><li><strong>Hold pending evidence.</strong></li><li><strong>Stop.</strong></li></ul><p>Constraints might include limited users, human approval, read-only operation, reduced agent authority, transaction limits, enhanced monitoring or a time-bound pilot. Unanimity is not always necessary. Explicit authority is.</p><p>This is <strong>Gate 3: pre-production sign-off</strong>, followed by <strong>Gate 4: production release</strong>.</p><p>Some organizations can combine those gates. Higher-risk environments may separate evidence completeness from residual-risk acceptance. Ten reviewers do not create accountability by attendance. If nobody can clearly identify who had the authority to approve the release, the organization has documentation, not governance.</p><h3>Stage 8: Production Monitoring</h3><p>Deployment does not end evaluation. It changes its form. The system now needs to be monitored against the assumptions that justified release. Technical quality may drift. Users change behavior. Retrieval data becomes stale. Costs rise. Prompt changes alter failure patterns. Controls that looked adequate in testing can fail under real workload. Monitoring needs multiple owners, but one integrated escalation path.</p><blockquote>Technical teams watch model and system behavior.</blockquote><blockquote>Business owners watch outcome metrics.</blockquote><blockquote>Data teams watch changes in input quality and provenance.</blockquote><blockquote>Risk teams watch control performance and emerging exposure.</blockquote><blockquote>Operations teams watch resilience, incidents and rollback readiness.</blockquote><p>A useful rule is simple:</p><p><strong>If a metric was important enough to justify launch, someone should be accountable for deciding whether it still matters after launch.</strong></p><h3>Stage 9: Incident Response</h3><p>AI incidents expose ambiguous authority very quickly.</p><blockquote>Who can suspend the system?</blockquote><blockquote>Who can disable a tool used by an agent?</blockquote><blockquote>Who can force a fallback to a deterministic process?</blockquote><blockquote>Who decides whether an apparently minor quality issue is actually a material control failure?</blockquote><p>Stop authority should be defined before the incident. For material AI systems, incident ownership should cover detection, suspension, rollback, communication, remediation and the decision to resume. The person authorized to restart the system should not be inferred during the outage.</p><h3>Stage 10: Change and Re-Evaluation</h3><p>AI systems are unusually fluid. A model version changes. A prompt is rewritten. An agent gains a new tool. The retrieval corpus expands. A new user group gains access. A workflow moves from recommendation to execution.</p><p>Any of these can invalidate original evidence. Annual review is too blunt an instrument for this. Organizations need defined re-evaluation triggers.</p><p>A model change may require technical retesting. A new user population may require business or fairness review. New write-back authority may require security and operational-risk reassessment.</p><p>This is <strong>Gate 5: continuation or re-evaluation</strong>. The question is not, “Did we approve this system last year?”. It is, “Does current evidence still justify current operation?”</p><h3>Stage 11: Retirement</h3><p>There is one final decision that AI governance often neglects.</p><p>Who can decide that the system should no longer exist?</p><p>Retirement may follow declining value, excessive cost, obsolescence, regulatory change, deteriorating performance or a better alternative. Once embedded in workflows, systems can survive long after their business case weakens. Outcome ownership should therefore include the authority to stop funding and using the system, not just the authority to sponsor it.</p><h3>More Metrics Will Not Fix an Accountability Gap</h3><p>When enterprises are uncomfortable with AI, they often respond by producing more evidence. More benchmarks. More dashboards. More red-team reports. More guardrail tests. More assessments. More forms. Some of that is useful.</p><p>But evidence volume is not the same as decision quality. A system can have hundreds of metrics and still leave the key question unanswered: what evidence is sufficient, and who decides?</p><p>A dashboard can illuminate uncertainty. It cannot inherit accountability. The danger is an evaluation bureaucracy that becomes increasingly sophisticated while decision rights remain vague. That is expensive theatre.</p><h3>What This Looks Like in Practice</h3><p>A workable model can be established in 30 to 90 days if the focus stays on material decisions.</p><p>Start with high-impact systems, not every experiment. Prioritize meaningful customer, financial, regulatory, employee, safety or decision-making consequence. A note summarizer should not face the same process as an agent that can alter a customer account. For each material system, name one accountable business owner. Not a steering committee. Not “the AI team.” A person who can own the outcome and consequence of continued operation.</p><p>Then separate evidence owners from decision owners. Engineering may produce technical evidence. Product may produce workflow evidence. Risk may challenge control evidence. Operations may establish recoverability. These responsibilities can remain distributed.</p><p>The release authority should still be explicit.</p><p>Next, define acceptance thresholds before the evaluation begins. Deciding what “good enough” means after seeing the scores invites motivated reasoning. Define stop-ship criteria with the same discipline. A critical security exposure, regulatory breach, severe safety failure, inability to reproduce evaluation, unacceptable business impact or material degradation should have predetermined consequences. Create gates only where a real decision exists. Each should answer a question, assign authority and record an outcome.</p><p>Connect production monitoring back to pre-production evidence. If release depended on a maximum hallucination rate, minimum human-override rate, bounded cost, retrieval freshness or defined business KPI, production monitoring should tell someone whether that assumption still holds.</p><p>Define re-evaluation triggers explicitly. Model replacement, prompt changes, new tools, new data, new markets and changed workflow authority should not depend on someone remembering to notify governance. And record residual risk as an accountable decision.</p><p>“Accepted” is not a control.</p><p>A useful record says what exposure remains, who accepted it, what assumptions support that acceptance, what controls are required, and what event would force reconsideration.</p><h3>The Anti-Patterns Are Easy to Recognize</h3><p><strong>“Engineering signed it off.”</strong><br>Engineering established technical acceptance. That is not organizational approval.</p><p><strong>“Risk approved it.”</strong><br>Risk may have assessed and challenged exposure. That does not necessarily make risk the owner of the business consequence.</p><p><strong>“The business wanted it.”</strong><br>Commercial urgency is not evidence.</p><p><strong>“Everyone was in the meeting.”</strong><br>Participation is not accountability.</p><p><strong>“The governance committee owns it.”</strong><br>Committees can coordinate. They rarely own consequences unless decision rights are explicitly assigned.</p><p><strong>“It passed last quarter.”</strong><br>A fluid AI system can invalidate old evidence without asking permission.</p><p>These statements are not symptoms of weak people. They are symptoms of an incomplete operating model.</p><h3>The Missing Role Is the Call</h3><p>Enterprises have spent significant energy improving how AI is measured. That work matters. Better evaluation datasets matter. Better benchmarks matter. Better red teaming, observability, monitoring and controls matter.</p><p>But the maturity test is not whether an organization can produce more evidence. It is whether evidence converges into an accountable decision.</p><p>The future of AI governance should therefore be less about creating a central function that “owns evaluation” and more about designing an accountability architecture around the lifecycle of the system.</p><blockquote>Distributed expertise.</blockquote><blockquote>Integrated accountability.</blockquote><blockquote>Explicit decision rights.</blockquote><p>That is the model. Because the hardest question in enterprise AI is no longer whether we can test the system. It is whether someone is clearly accountable for deciding what the test results mean.</p><p><strong>When everyone owns a piece of the evidence, someone still has to own the decision.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2fb1670cb3dd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Much Cost Can AI Really Remove From a Bank?]]></title>
            <link>https://medium.com/@vikram.vishal/how-much-cost-can-ai-really-remove-from-a-bank-60087d1f7fc6?source=rss-ebf8285a9567------2</link>
            <guid isPermaLink="false">https://medium.com/p/60087d1f7fc6</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[banking]]></category>
            <dc:creator><![CDATA[Vikram Vishal]]></dc:creator>
            <pubDate>Mon, 24 Aug 2026 02:50:43 GMT</pubDate>
            <atom:updated>2026-08-24T02:50:43.098Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>AI productivity potential is abundant. Converting it into bankable cost takeout is a maturity, operating-model and financial-realization problem.</strong></p><p>Banks do not have an AI opportunity problem. They increasingly have an AI economics problem. Across retail and commercial banking, the list of plausible use cases is already long: contact-centre copilots, voice agents, KYC document processing, credit memo generation, collections, financial-crime investigation, relationship-manager assistants, software engineering copilots and, increasingly, agentic workflows that can execute work rather than simply recommend what a human should do next.</p><p>The headline numbers are equally difficult to ignore. McKinsey estimates that generative AI could create <strong>$200 billion to $340 billion of annual value across global banking</strong>, equivalent to roughly 2.8% to 4.7% of industry revenues, largely through productivity improvements. Accenture has separately estimated <strong>22% to 30% productivity potential</strong> for banking from generative AI.</p><p>Those are significant numbers. They are also surprisingly easy to misuse. A credit analyst becoming 30% more productive does not reduce the credit function’s operating expense by 30%. A contact-centre copilot saving two minutes per call does not automatically remove two minutes of payroll. An AI assistant giving a commercial relationship manager six hours back every week may create substantial business value without taking a single dollar out of the bank’s cost base.</p><p>That distinction matters because CFOs do not report “hours theoretically saved” in the P&amp;L. They report expenses. So the useful question for a bank is no longer: <strong>How productive could AI make us?</strong></p><p>It is: <strong>What proportion of our existing cost base can AI genuinely influence, how much of that potential will survive our process, data, technology and control environment, and how much released capacity can actually become financial benefit?</strong></p><p>That is a much harder question. It is also where the AI business case becomes real.</p><h3>Productivity is not one financial outcome</h3><p>Before discussing use cases, banks need a cleaner vocabulary for AI value. There are at least four economically different outcomes.</p><h3>Hard cost takeout</h3><p>This is expense that genuinely disappears from the run-rate P&amp;L.</p><blockquote>Contractor capacity is reduced.</blockquote><blockquote>BPO volumes fall.</blockquote><blockquote>Vacancies are not filled.</blockquote><blockquote>Natural attrition allows staffing levels to decline.</blockquote><blockquote>A technology platform is retired.</blockquote><blockquote>Licensing or infrastructure expense goes away.</blockquote><p>This is the category most CFOs mean when somebody walks into the room promising “$50 million of AI savings.”</p><h3>Capacity release</h3><p>AI makes people faster, but the people remain. An operations analyst handles more cases. A developer completes more stories. A relationship manager spends less time preparing for client meetings. That released capacity can be extremely valuable. But until something happens to the underlying cost base, it is not hard savings.</p><h3>Cost avoidance</h3><p>This becomes particularly important in growing banks. If transaction volumes increase 20% but AI allows an operation to absorb that growth without a corresponding increase in headcount, the financial benefit is real. The bank has avoided future cost rather than removed today’s cost.</p><h3>Revenue and risk benefits</h3><p>AI can also improve conversion, collections, fraud detection, customer retention, credit decision speed and relationship-manager capacity. Those benefits belong in the broader business case. But they should not quietly be relabelled as cost takeout.</p><p>This distinction is particularly important when external productivity benchmarks enter executive presentations. Accenture’s analysis, for example, found a large proportion of banking work potentially affected by generative AI through a combination of automation and augmentation. That tells us how much work could change. It does not tell us how much banking expense can disappear.</p><p><strong>Work addressability and cost removability are not the same thing.</strong></p><p>The gap between them is where most AI business cases need more discipline.</p><h3>Where AI can genuinely change banking economics</h3><p>The strongest opportunities usually appear where three characteristics coincide: <strong>substantial labour or vendor cost, repetitive information-intensive work, and enough digital workflow to let AI do something useful with the information it processes.</strong></p><p>Let’s consider retail banking first.</p><h3>Customer service</h3><p>Contact centres are attractive because their economics are unusually visible. Banks know interaction volumes, average handling time, after-call work, abandonment, transfer rates, repeat contacts and staffing requirements. Generative AI can retrieve knowledge, recommend responses, summarise calls and complete documentation. Voice agents can go further by resolving selected requests without a human agent at all.</p><p>But there is an important economic distinction. An agent copilot that reduces average handling time is primarily a productivity intervention. A voice agent that prevents 20% of appropriate calls from entering a human queue changes demand for human capacity.</p><p>The second has a clearer path to P&amp;L. That is why the most valuable service transformations will increasingly combine self-service, conversational AI, agent assistance, intelligent routing and workflow redesign rather than deploying a copilot in isolation.</p><h3>Retail onboarding and KYC</h3><p>Onboarding remains rich in document ingestion, data entry, reconciliation, screening, case preparation and exception handling. AI can extract information from identity documents, compare information across sources, prepare screening summaries and assemble cases. Traditional automation can execute deterministic steps. Agents can potentially coordinate document services, screening platforms, customer records and workflow applications.</p><p>But exception rates matter enormously. An onboarding pilot based on clean documents can look highly automated. Production reality may involve incomplete data, poor scans, inconsistent names, old customer records and unusual ownership structures. The percentage of work reaching humans ultimately determines the economics.</p><h3>Lending and credit</h3><p>Retail and small-business lending contain significant information-processing effort: document extraction, policy checks, application analysis, exception review and underwriting preparation. AI can perform much of the preparation. It can support the decision without necessarily owning the decision. That matters because human credit judgment, regulatory requirements and fair-lending controls may appropriately remain.</p><p>The result can still be valuable: faster decisions, more applications handled per analyst and less operational processing. But the financial outcome may initially be capacity release rather than headcount reduction.</p><h3>Collections</h3><p>Collections is particularly interesting because AI can affect both efficiency and outcome. Traditional analytics already support segmentation and next-best-action strategies. Generative AI and voice technology add personalised conversations, agent support and automated workflow execution.</p><p>BCG reported in 2025 that AI-enabled collection agents were achieving improved cure rates while operating at <strong>30% to 40% lower cost</strong> in examples it observed. More recent BCG work continues to identify collections as one of the high-potential domains for end-to-end agentic redesign. These figures should be treated as examples of observed or projected potential, not universal bank benchmarks.</p><p>That qualification matters. The lesson is not that every bank should put 35% into its collections business case. It is that economics change materially when AI removes or orchestrates a unit of work rather than simply helping a collector write better notes.</p><h3>Fraud, AML and financial crime</h3><p>Investigation functions contain large amounts of expensive information gathering and review. AI can prioritise alerts, retrieve evidence, summarise transaction histories, analyse documents, build investigation chronologies and draft narratives. There is genuine potential. There is also a control ceiling.</p><p>Regulatory accountability, evidence preservation, explainability and human approval may deliberately constrain autonomy. A 50% improvement in investigation preparation effort does not necessarily produce a 50% reduction in investigation staffing. The control model determines part of the value model.</p><h3>Commercial banking has a different economic shape</h3><p>Commercial banking should not simply inherit the retail AI playbook. The work is often more judgment-intensive, relationship-driven and document-heavy. That creates substantial AI opportunity, but not always direct cost takeout.</p><h3>Relationship-manager productivity</h3><p>AI can assemble client intelligence, prepare company and industry briefings, summarise meetings, update CRM records, support account planning and draft initial proposals. These are useful interventions.</p><p>But suppose AI releases 15% of a relationship manager’s capacity. Reducing relationship-manager numbers by 15% may be exactly the wrong decision. The bank may create substantially more value by reallocating those hours toward client coverage, cross-sell, prospecting and advisory conversations.</p><p>Here, the correct business case may be revenue capacity rather than cost takeout.</p><h3>Commercial credit</h3><p>Commercial underwriting has a different profile. Financial spreading, borrower-document ingestion, annual reviews, covenant monitoring, industry research, credit-memo preparation and exception identification contain significant operational effort.</p><p>AI can address a meaningful portion of it.</p><p>The potential becomes considerably stronger when the solution can reach financial data, customer information, loan-origination platforms, credit policies and historical decisions rather than requiring an analyst to manually assemble everything first. Without those connections, the AI remains a sophisticated writing assistant. With them, the workflow begins to change.</p><h3>KYB and commercial onboarding</h3><p>Legal-document review, beneficial-ownership analysis, entity reconciliation, screening and exception handling create substantial processing effort. The prize is meaningful. So are the dependencies.</p><p>Data often spans customer masters, external registries, KYC systems, sanctions platforms, document repositories and workflow applications. This is precisely the kind of environment where a brilliant model can deliver a disappointing business case.</p><h3>Loan operations and transaction banking</h3><p>Documentation, booking, maintenance, covenant processing, reconciliations, payment investigations and servicing requests create attractive operational pools. Trade finance is another obvious candidate because the underlying processes remain document-intensive and exception-heavy.</p><blockquote>Here, agentic AI becomes economically interesting. The opportunity is no longer merely generating an answer. It is: <strong>retrieve information → reason within defined boundaries → interact with systems → execute permitted actions → evidence what happened → escalate the exception.</strong></blockquote><p>That is a much more consequential form of automation. It also requires a much more mature bank around it.</p><h3>Not every AI use case has the same economic shape</h3><p>A useful way to think about the banking AI portfolio is to map opportunities against two questions:</p><p><strong>How directly can productivity translate into a lower cost base?</strong></p><p>And:</p><p><strong>How much process, data and technology maturity is required before the value can be realised at scale?</strong></p><p>That produces four different value zones.</p><h3>1. TAKEOUT-READY</h3><p><strong>High P&amp;L potential. Relatively visible cost lever.</strong></p><p>Typical examples include:</p><p><strong>Customer service:</strong> Voice AI, self-service, routing and agent assistance can alter contact volumes, handling time and ultimately agent or BPO capacity.</p><p><strong>High-volume retail KYC:</strong> Document processing, reconciliation and case orchestration can reduce measurable operations effort.</p><p><strong>Collections:</strong> Treatment optimisation, voice AI and workflow automation can influence both staffing requirements and third-party collections expenditure.</p><p>What makes this zone attractive is not that the AI is necessarily more advanced. It is that management can see the cost pool and explain what happens when the workload falls.</p><h3>2. SCALE-TO-SAVE</h3><p><strong>High potential. High ecosystem-maturity dependency.</strong></p><p>This includes:</p><p><strong>Commercial credit</strong> — spreading, reviews, covenant monitoring and credit-memo workflows.</p><p><strong>Commercial onboarding/KYB</strong> — legal-document processing, entity reconciliation and beneficial-ownership analysis.</p><p><strong>Loan operations</strong> — booking, maintenance, servicing, reconciliations and exception handling.</p><p><strong>Financial crime</strong> — alert triage, investigation preparation and evidence assembly.</p><p>The prizes can be large. But these processes cross data sources, systems, controls and organisational boundaries. The AI capability may already exist.</p><p><strong>The bank around it may not yet be ready.</strong></p><h3>3. CAPACITY-FIRST</h3><p><strong>Meaningful productivity. Moderate direct cost conversion.</strong></p><p>Typical examples include underwriting support, risk and compliance copilots, regulatory analysis and software-engineering assistance. These initiatives can release substantial employee capacity.</p><p>But the first-order financial effect may be:</p><blockquote>increased throughput,</blockquote><blockquote>reduced backlog,</blockquote><blockquote>improved control coverage,</blockquote><blockquote>lower contractor dependence,</blockquote><blockquote>or avoidance of future hiring.</blockquote><p>All are economically valuable. None should automatically be presented as immediate headcount reduction.</p><h3>4. GROWTH-REDEPLOY</h3><p><strong>High business value. Low direct cost-takeout potential.</strong></p><p>Commercial relationship-manager AI belongs here. AI can remove administrative work, improve preparation and increase the quality of client insight. But if the bank responds by mechanically removing front-office capacity, it may destroy the value it just created. Some productivity should be redeployed. That is not failure. It is a different economic outcome.</p><p>The portfolio therefore looks something like this:</p><p><strong>Best candidates for near-term cost takeout:</strong><br>Customer service → High-volume onboarding/KYC → Collections</p><p><strong>Large prizes requiring ecosystem maturity:</strong><br>Commercial credit → KYB → Loan operations → Financial crime</p><p><strong>Productivity plays likely to begin as capacity release:</strong><br>Underwriting → Risk and compliance → Technology engineering</p><p><strong>Use cases where redeployment may be worth more than removal:</strong><br>Commercial relationship management</p><p>Two AI initiatives can both show 30% productivity improvement and deserve completely different financial treatment. One may reduce a vendor contract. Another may prevent future hiring. A third may increase throughput. And a fourth may create no credible P&amp;L reduction at all.</p><p><strong>The mistake is not necessarily choosing the wrong AI use case. It is assigning the wrong economic outcome to it.</strong></p><h3>The AI-to-P&amp;L Conversion Framework</h3><p>This leads to a more useful way of evaluating AI economics. Think of realizable savings as a chain. Every weak link reduces the number reaching the P&amp;L.</p><h3>Step 1: Establish the Addressable Cost Base</h3><p>Start with what the bank spends today. Not the entire function. Not the entire bank. Define an <strong>Addressable Cost Base — ACB</strong>.</p><p>It may include: employee cost, contractor expenditure, BPO and processing cost, relevant technology run expense, application licences and infrastructure. It should also include operating volumes: transactions, cases, calls, customers, applications, handling times and unit costs.</p><p>Only include costs associated with activities that AI can materially influence. If a bank cannot establish today’s unit economics, proving tomorrow’s savings becomes surprisingly difficult.</p><h3>Step 2: Decompose work into tasks</h3><p>Jobs are too coarse a unit for AI economics. Tasks are better. A commercial credit analyst does not perform “credit analysis” eight hours a day. The role contains document retrieval, financial spreading, search, reconciliation, analysis, writing, customer interaction, judgment and approval. Estimate what percentage of current effort is genuinely addressable.</p><p>Call this <strong>Task Addressability — TA</strong>.</p><h3>Step 3: Determine the feasible AI intervention</h3><p>For addressable work, ask what AI can realistically do.</p><h3>Assist</h3><p>AI makes a human faster.</p><h3>Automate</h3><p>AI completes a defined activity with limited intervention.</p><h3>Orchestrate</h3><p>AI or an agent coordinates multiple tasks and systems, executes bounded actions and escalates exceptions.</p><p>Estimate the <strong>Feasible AI Intervention Rate — FAIR</strong>.</p><p>The word <em>feasible</em> matters. A model demonstration should not determine this figure. Production constraints should.</p><h3>Step 4: Apply the Maturity Realization Factor</h3><p>This is the factor most AI business cases underestimate. A sophisticated model placed on top of a fragmented operating environment does not magically make the environment sophisticated. Assess maturity across five dimensions.</p><h3>Process and workflow</h3><blockquote>Are processes standardised?</blockquote><blockquote>Are workflows digitised?</blockquote><blockquote>Are exceptions identifiable?</blockquote><blockquote>Do process-level KPIs exist?</blockquote><h3>Data</h3><blockquote>Is the required information available, reliable and accessible?</blockquote><blockquote>Can structured and unstructured data be combined?</blockquote><blockquote>Are lineage, metadata, permissions and quality sufficiently understood?</blockquote><h3>Technology and integration</h3><blockquote>Can AI interact with applications through governed interfaces?</blockquote><blockquote>Are APIs, workflow engines and event architectures available?</blockquote><blockquote>How much work remains trapped inside legacy platforms?</blockquote><blockquote>Can the bank monitor AI applications in production?</blockquote><h3>AI, risk and controls</h3><blockquote>Can outputs be evaluated?</blockquote><blockquote>Can decisions be evidenced?</blockquote><blockquote>Are human-in-the-loop controls appropriately designed?</blockquote><blockquote>Are privacy, security, explainability and model governance established?</blockquote><h3>Operating model and adoption</h3><blockquote>Who owns the outcome?</blockquote><blockquote>Do business and technology redesign the workflow together?</blockquote><blockquote>Are incentives aligned?</blockquote><blockquote>Can released capacity actually be redeployed or removed?</blockquote><p>Score these dimensions and calibrate a bank-specific <strong>Maturity Realization Factor — MRF</strong>.</p><p>It should not be an industry constant. It should come from evidence. Four broad maturity archetypes help make this tangible.</p><h3>Fragmented / Manual</h3><p>Email, spreadsheets, manual hand-offs and inconsistent data dominate. AI primarily assists humans.</p><h3>Digitized but Siloed</h3><p>Workflows exist, but journeys cross system boundaries and data remains fragmented. Task automation is achievable; end-to-end automation remains difficult.</p><h3>Integrated &amp; AI-Ready</h3><p>Governed data, APIs, workflow orchestration and production controls allow AI to interact with the operating environment. Now meaningful orchestration becomes possible.</p><h3>AI-Native / Agentic</h3><p>Agents can retrieve context, invoke authorised tools, execute bounded actions, maintain evidence and escalate exceptions. The unit of transformation begins to move from the individual task toward the entire workflow.</p><p>This maturity question is particularly relevant for APAC banks and banking GCCs in India. Many have sophisticated technology teams and significant AI ambition, while still working through legacy estates, fragmented data ownership and global control structures.</p><p>Regulators are also pushing in the same direction. The RBI’s <strong>FREE-AI framework</strong>, published in August 2025, explicitly addresses responsible and ethical AI adoption in financial services. BIS research in March 2026 similarly emphasised data quality, privacy, security and third-party dependencies as material constraints on wider AI adoption. Data readiness is therefore not an AI plumbing issue. <strong>It is an economic variable.</strong></p><h3>Step 5: Measure adoption, not deployment</h3><p>A bank can purchase 10,000 AI licences and create very little value. Deployment is not adoption. Measure: eligible users, active use, workflow penetration, frequency, output acceptance and exception rates.</p><p>Call this the <strong>Adoption &amp; Utilization Factor — AUF</strong>.</p><p>If AI is technically capable of addressing 50% of a process but is meaningfully used in only 40% of eligible work, the business case needs to reflect that.</p><h3>Step 6: Apply the P&amp;L Conversion Factor</h3><p>This is where many impressive AI business cases collapse. Suppose an initiative releases capacity equivalent to 100 FTEs. What happens next?</p><blockquote>Can contractor numbers fall?</blockquote><blockquote>Can vacancies remain unfilled?</blockquote><blockquote>Will natural attrition reduce staffing?</blockquote><blockquote>Can BPO volumes be renegotiated?</blockquote><blockquote>Can future hiring be avoided?</blockquote><blockquote>Can systems be retired?</blockquote><blockquote>Or will all 100 FTE equivalents simply absorb other work?</blockquote><p>Define the percentage capable of becoming financial benefit as the <strong>P&amp;L Conversion Factor — PCF</strong>. This factor forces Finance, HR, Procurement and Operations into the AI business case. They need to be there.</p><p><strong>The spreadsheet is the easy part. Converting released capacity into a lower cost base is harder.</strong></p><h3>Step 7: Calculate gross realizable cost takeout</h3><p>The equation becomes:</p><p><strong>Gross Realizable Cost Takeout = ACB × TA × FAIR × MRF × AUF × PCF</strong></p><p>Or, in plain English:</p><blockquote>Start with the cost AI can influence.</blockquote><blockquote>Reduce it to the work AI can address.</blockquote><blockquote>Reduce it again for what AI can realistically perform.</blockquote><p>Then adjust for whether the bank is mature enough to support it, whether people actually use it, and whether released capacity can become financial benefit. Do not pretend these variables are precise. Run <strong>Low, Base and High scenarios</strong>. False precision is not sophistication.</p><h3>Step 8: Calculate net AI economics</h3><p>AI is not free automation. Production environments require model consumption, cloud compute, AI platforms, data pipelines, integration, security, evaluation, observability, model-risk controls, human quality assurance and operational support.</p><p>Therefore: <strong>Net Annual Cost Takeout = Gross Realizable Cost Takeout − Incremental Annual AI Run Cost</strong></p><p>And:</p><p><strong>Year-One Net Benefit = Net Annual Cost Takeout − One-Time Implementation and Transformation Cost</strong></p><p>Then look at payback and three-year economics. Now the discussion starts looking like a CFO business case rather than an AI presentation.</p><h3>One illustrative example</h3><p>Consider a hypothetical retail contact centre with a <strong>$20 million annual addressable cost base</strong>. These numbers are deliberately illustrative, not banking benchmarks.</p><p>Assume:</p><blockquote><strong>Task Addressability:</strong> 70%<br><strong>Feasible AI Intervention:</strong> 50%<br><strong>Maturity Realization:</strong> 70%<br><strong>Adoption &amp; Utilization:</strong> 80%<br><strong>P&amp;L Conversion:</strong> 60%</blockquote><p>Then:</p><p><strong>$20m × 70% × 50% × 70% × 80% × 60%</strong></p><p>produces approximately: <strong>$2.35 million of gross realizable annual cost takeout.</strong></p><p>Suppose the recurring AI platform, model, integration, controls and operations cost is $600,000. Net recurring cost takeout becomes approximately:</p><p><strong>$1.75 million annually.</strong></p><p>If the initial transformation costs another $1 million: <strong>Year-one net benefit is approximately $750,000.</strong></p><p>Now move the identical AI capability into a fragmented legacy bank. Leave Task Addressability and AI Intervention Potential unchanged.</p><p>But reduce:</p><blockquote>Maturity Realization to 35%,<br>Adoption to 55%,<br>P&amp;L Conversion to 35%.</blockquote><p>Gross realizable savings fall to roughly <strong>$470,000</strong>.</p><p>If recurring AI expense is still $600,000, the use case is negative on a pure cost-takeout basis. Same models. Same apparent use case. Completely different economics.</p><p><strong>That is why AI cost takeout is as much an ecosystem-maturity problem as an AI problem.</strong></p><h3>Controls belong inside the economics</h3><p>Banking cannot optimise AI economics by pretending risk does not exist. Model risk, privacy, data leakage, bias, fair-lending obligations, explainability, cybersecurity, auditability and human accountability are operating requirements.</p><p>BIS research has highlighted model-risk management, data governance, governance capabilities and third-party AI providers among the areas financial institutions and supervisors need to address. But controls themselves also consume effort. Suppose AI produces an AML investigation narrative in five minutes instead of an investigator spending forty. Excellent!!! If policy then requires the investigator to independently reconstruct every step before approving the result, very little of the theoretical productivity reaches the cost base. The right question is not:</p><blockquote><em>“Can we remove human oversight?”</em></blockquote><p>It is:</p><blockquote><em>“What is the minimum control necessary for this level of operational consequence?”</em></blockquote><p>Good control design therefore sits inside the value equation. Not outside it.</p><h3>Ten ways the business case gets inflated</h3><p>Before approving a major AI cost-takeout number, I would test whether the business case has done any of the following:</p><blockquote>1. Counted every hour saved as P&amp;L savings.</blockquote><blockquote>2. Applied one productivity percentage across an entire function.</blockquote><blockquote>3. Counted the same FTE pool against multiple use cases.</blockquote><blockquote>4. Mixed revenue benefits with cost reduction.</blockquote><blockquote>5. Ignored AI infrastructure and model operating costs.</blockquote><blockquote>6. Ignored human validation and control requirements.</blockquote><blockquote>7. Extrapolated pilot productivity directly to enterprise scale.</blockquote><blockquote>8. Assumed near-perfect adoption.</blockquote><blockquote>9. Ignored the implementation and transition ramp.</blockquote><blockquote>10. Treated a vendor or consulting benchmark as a guaranteed bank outcome.</blockquote><p>A CFO will eventually find these assumptions. Better that the AI team finds them first.</p><h3>A five-question executive diagnostic</h3><p>Before launching another use-case prioritisation exercise, leadership can ask five questions.</p><blockquote><strong>1. Can we identify the current unit cost of the process we intend to change?</strong></blockquote><blockquote><strong>2. Do we have reliable, accessible and permissioned data covering most of that workflow?</strong></blockquote><blockquote><strong>3. Can AI interact with the systems required to complete the work — not just generate an answer?</strong></blockquote><blockquote><strong>4. Can its outputs and actions be governed, evidenced and audited to banking standards?</strong></blockquote><blockquote><strong>5. Do we know precisely how released capacity will turn into financial benefit?</strong></blockquote><p>If leadership cannot answer these questions, the organisation may have an AI experimentation plan. It probably does not yet have an AI cost-takeout plan.</p><h3>Stop building a portfolio of disconnected pilots</h3><p>The next mistake is to score 100 AI ideas and launch the highest-ranking 25. AI value is rarely additive in that way.</p><p>A better prioritisation equation is: <strong>Value × Feasibility × Maturity × P&amp;L Convertibility</strong></p><p>Then identify a small number of <strong>end-to-end value domains</strong>:</p><blockquote>Customer Service.</blockquote><blockquote>Onboarding.</blockquote><blockquote>Lending.</blockquote><blockquote>Commercial Credit.</blockquote><blockquote>Collections.</blockquote><blockquote>Financial Crime.</blockquote><blockquote>Engineering.</blockquote><p>Redesign the domain. Automating fourteen tasks across fourteen different processes frequently creates fourteen integration projects while leaving the underlying operating model untouched. Transforming one complete value stream allows workload, exceptions, systems, controls and capacity to change together. That is when productivity has a chance of becoming structural cost reduction.</p><h3>AI potential is abundant. P&amp;L conversion is scarce.</h3><p>Banking is unlikely to suffer from a shortage of AI ideas. Models will become cheaper. Voice will improve. Document intelligence will become routine. Agents will execute increasingly complex workflows. The technical ceiling will keep moving upward.</p><p>The more interesting constraints will increasingly sit elsewhere.</p><blockquote>Can the bank expose trustworthy data?</blockquote><blockquote>Can its systems be invoked safely rather than worked around?</blockquote><blockquote>Can a process be redesigned instead of merely accelerated?</blockquote><blockquote>Can controls become proportionate rather than duplicative?</blockquote><blockquote>Can employees genuinely adopt the new workflow?</blockquote><p>And when capacity is released, can management make the operating-model decisions required to turn some of it into economics? That is why an AI cost-takeout exercise should not begin with the model.</p><p><strong>It should begin with the cost base.</strong></p><p>Then work through the tasks. Then maturity. Then AI. Then adoption. And only finally, P&amp;L conversion.</p><p>If you are working through the gap between your bank’s AI ambition and the cost that can realistically come out of its operating model, I would be interested in comparing notes — particularly around applying an <strong>AI-to-P&amp;L diagnostic</strong>, maturity assessment or use-case prioritisation exercise to a real banking cost base.</p><p>Because the question is no longer whether AI can make banking more productive.</p><p><strong>The question is how much of that productivity your bank is actually capable of banking.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=60087d1f7fc6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Hidden Risk Is Not Hallucination. It Is Write-Back.]]></title>
            <link>https://medium.com/@vikram.vishal/the-hidden-risk-is-not-hallucination-it-is-write-back-8037e05dc743?source=rss-ebf8285a9567------2</link>
            <guid isPermaLink="false">https://medium.com/p/8037e05dc743</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Vikram Vishal]]></dc:creator>
            <pubDate>Tue, 18 Aug 2026 20:33:33 GMT</pubDate>
            <atom:updated>2026-08-18T20:33:33.203Z</atom:updated>
            <content:encoded><![CDATA[<p>An AI copilot gives a senior executive the wrong answer about quarterly revenue. Uncomfortable. Potentially embarrassing. Usually recoverable.</p><p>Now give the same underlying intelligence permission to update the financial ledger. Or approve a $50,000 customer refund. Or revoke an employee’s privileged access. Or modify a production configuration and restart the service. The reasoning error may be identical. The operational event is not.</p><p>For the past few years, much of the enterprise AI debate has concentrated on whether models hallucinate, whether their answers are accurate, whether prompts can be manipulated, and whether humans remain sufficiently involved. Those are legitimate concerns.</p><p>But agentic AI changes the boundary of the problem. Once an AI system can call tools, invoke APIs and modify enterprise state, the question is no longer simply whether its answer can be trusted. The question becomes: <strong>what authority have we attached to that answer?</strong></p><p>That distinction will matter far more as enterprises move from copilots that inform people to agents that perform work. A model that can hallucinate inside a read-only research environment creates one class of exposure. A less sophisticated model carrying production credentials and permission to alter customer accounts creates another. The second system may actually deserve more executive attention. We have become very interested in what AI can think. We need to become much more precise about what AI can change.</p><h3>Being Wrong Is Not the Same as Doing Wrong</h3><p>There are at least four risk states hiding inside the phrase “AI risk.”</p><p>The first is <strong>inference risk</strong>. The system interprets something incorrectly. Perhaps it misunderstands a policy document, misclassifies a transaction or draws the wrong conclusion from incomplete data.</p><p>Then comes <strong>recommendation risk</strong>. The incorrect inference becomes a proposed action: reject the payment, approve the claim, change the configuration, contact the customer. At this point, a human can still intervene.</p><p>The third state is <strong>execution risk</strong>. The AI can initiate an operational action from its reasoning. It can invoke the workflow, call the API, submit the transaction or trigger another system.</p><p>Then comes the boundary enterprises should pay much more attention to: <strong>write-back risk</strong>. Write-back occurs when the agent alters persistent state in a system, transaction, record, entitlement, workflow, configuration or external environment. A recommendation disappears when someone closes the screen. A posted journal does not. A generated remediation suggestion can be ignored. A modified firewall policy cannot. A proposed piece of code can be reviewed. Code automatically merged, deployed and allowed to change production data has crossed several control boundaries before anyone notices.</p><p>This is why “human in the loop” can also be misleading. A person presented with 300 AI-generated approvals during a busy afternoon may technically remain in the loop while contributing almost no meaningful judgment. The important distinction is not whether a human clicked something. It is where decision authority actually resided.</p><h3>The Write-Back Risk Ladder</h3><p>Enterprises need a simple way of expressing this. I would start with a <strong>Write-Back Risk Ladder</strong>: six levels that classify an AI system according to the operational authority it has been given.</p><blockquote>Not the size of the model.</blockquote><blockquote>Not whether it is called a copilot or an agent.</blockquote><blockquote>Not how impressive the demo looks.</blockquote><blockquote>What can it actually do?</blockquote><h3>Level 0: Observe</h3><p>The system can retrieve, search, analyze and summarize. It cannot initiate a business change. Think enterprise search, policy assistants, document summarization, read-only analytics or an engineering assistant diagnosing telemetry. This is not zero risk. Confidentiality, data entitlements, prompt injection, inappropriate retrieval and information leakage still matter. But the control envelope is primarily about what the system may see. It cannot change the source of truth.</p><h3>Level 1: Advise</h3><p>The AI can recommend an action, but another actor must initiate it. A fraud investigator receives a recommendation to escalate a case. A salesperson receives a next-best-action suggestion. An engineer receives a proposed remediation. The control problem expands from information access to recommendation quality, explainability where appropriate, human competence and decision accountability. The human remains the execution point. That boundary is valuable.</p><h3>Level 2: Prepare</h3><p>Here the AI produces an executable artifact but requires explicit approval before it becomes operational. It drafts the customer email. Prepares the journal entry. Creates the purchase order. Writes the SQL migration. Generates the infrastructure change. Produces the code. This level is more consequential than it first appears because preparation can create <strong><em>automation bias</em></strong>. If humans simply approve machine-generated work at scale, the control may exist in the workflow diagram but not in practice.</p><p>An effective approval gate needs enough context, time and authority for someone to say no. Otherwise “<em>human in the loop</em>” becomes “<em>human near the loop</em>.”</p><h3>Level 3: Execute Bounded, Reversible Actions</h3><p>Now the agent can write. But the authority is deliberately narrow. It may update a low-risk CRM attribute, categorize a service ticket, schedule an internal meeting or restart a non-critical development service. The important properties are bounded impact, strong observability and credible reversibility.</p><p>This is where enterprises can allow useful autonomy without treating every action as a privileged transaction. Rollback matters enormously here. An autonomous action becomes easier to tolerate when the organization can detect it quickly, reconstruct exactly what happened and restore the previous state safely.</p><h3>Level 4: Execute Material Actions</h3><p>At this level the agent can alter systems or business processes with meaningful financial, customer, regulatory or operational consequences. Consider:</p><blockquote>Customer refunds.</blockquote><blockquote>Production configuration changes.</blockquote><blockquote>Procurement transactions.</blockquote><blockquote>Customer entitlements.</blockquote><blockquote>HR case updates.</blockquote><blockquote>Financial postings.</blockquote><blockquote>Code deployment.</blockquote><p>The control architecture must now change materially. The agent needs its own identity. Permissions should be tightly scoped. Transaction ceilings matter. Policy constraints should be enforced outside the model. Actions need durable logs. High-risk conditions may require step-up authorization. Abnormal patterns should trigger intervention. The model cannot be its own control function.</p><h3>Level 5: High-Consequence or Difficult-to-Reverse Authority</h3><p>This should be exceptional. An agent at this level could move substantial funds, modify privileged access, change security policy, deploy into critical production environments, execute legally consequential transactions or make regulated decisions with significant downstream effect. Some organizations will eventually permit forms of Level 5 autonomy. That does not mean Level 5 should become the maturity destination for every agent. Quite the opposite. A mature enterprise may deliberately keep many highly capable systems at Level 1, 2 or 3. Maximum autonomy is not synonymous with good architecture.</p><h3>One API Call Can Have a Very Long Tail</h3><p>Write-back is not binary. A useful mental model is: <strong>Write-Back Risk ≈ Authority × Consequence × Blast Radius × Irreversibility</strong></p><p>It is not a quantitative risk formula. It is a design test. Authority asks how much the agent is permitted to do. Consequence asks what one incorrect action could cost. Blast radius asks how far a failure can propagate. Irreversibility asks how difficult it is to restore the world that existed before the action. Consider something apparently mundane: an AI agent changes one field in a customer record. That field might feed pricing. Pricing might trigger billing. Billing may trigger customer communication.</p><p>The transaction may enter revenue reporting, downstream analytics and another automated workflow before anyone detects the original error. Technically, the agent performed one database update. Operationally, it initiated a chain reaction. This is why rollback cannot be treated as an application-level afterthought. In agentic environments, reversibility becomes part of the authorization decision itself. Before granting an agent permission to write, an architect should know not only <em>whether</em> the write can be reversed but how far its consequences may travel before reversal becomes meaningless.</p><h3>Authority Should Determine the Control Envelope</h3><p>Enterprises already know how to do much of this. Banks do not give every employee unlimited payment authority. Developers should not receive unrestricted production privileges simply because they can write code. Junior employees cannot sign every contract. Financial systems use approval limits. Privileged-access systems constrain administrative authority. Service accounts are supposed to receive only the permissions they require.</p><p>Yet AI is sometimes discussed as though decades of control engineering somehow stopped applying once software began reasoning in natural language. They did not. The controls should simply become proportional to the authority being delegated. A Level 0 knowledge agent may primarily require data-access controls, confidentiality safeguards and retrieval governance. A Level 2 finance agent preparing journal entries needs provenance, segregation of duties and an approval experience designed for genuine review. A Level 3 operational agent needs bounded API scopes, idempotent actions, rate limits, telemetry and reliable rollback. A Level 4 refund agent may need per-transaction limits, cumulative spending ceilings, policy-as-code, anomaly detection, escalation rules, immutable transaction histories and emergency revocation. A production deployment agent requires a different envelope again: environment restrictions, change policies, deployment gates, blast-radius controls and immediate suspension capability.</p><p>The architecture should become stricter as operational consequence rises. That sounds obvious. Many important control failures begin with something obvious that nobody converted into an enforceable design rule.</p><h3>Agent Identity Is Only Half the Answer</h3><p>This is where the emerging discussion around <strong><em>agent identity</em></strong> becomes important. Identity platforms are beginning to treat AI agents as principals that need distinct authentication and authorization rather than allowing every agent to disappear behind a human or generic application credential.</p><p>Microsoft, for example, now describes Entra Agent ID as a framework for authenticating, authorizing, governing and protecting AI-agent identities. Google Cloud has introduced Agent Identity as a distinct IAM construct capable of identifying an agent acting independently or on behalf of a user.</p><p>That direction matters. An autonomous agent operating under a shared human credential creates a forensic and governance problem from the start. Suppose an account entitlement changes at 2:14 a.m. Knowing that the credential ultimately belonged to Alice is no longer sufficient.</p><blockquote>Which agent made the change? Was it acting directly or on Alice’s behalf? Which policy granted the delegation? What resource was it authorized to modify? What tool invocation caused the write? Was the credential valid specifically for that action? What happened next?</blockquote><p>These questions become fundamental once software actors begin participating in enterprise workflows. Google’s current agent-identity architecture explicitly distinguishes an agent’s own credential and supports agent-specific authorization; Microsoft’s agent identity model similarly gives AI agents dedicated identity constructs. This is a useful evolution, but identity alone does not tell us whether the authority granted to that identity is sensible.</p><blockquote><strong>Agent identity answers: Who is acting?</strong></blockquote><blockquote><strong>The Write-Back Risk Ladder answers: What is that actor permitted to change?</strong></blockquote><p>Enterprise agent governance needs both.</p><h3>Classify the AI Estate Again</h3><p>There is a simple exercise I would put in front of any enterprise deploying agentic AI. Take every AI use case in production, every pilot approaching production and every agent on the roadmap. Assign each one a Write-Back Level. The exercise becomes interesting almost immediately. An enterprise knowledge assistant is probably Level 0. A sales system recommending an email is Level 1. Drafting the email is Level 2. Sending it automatically crosses into execution and creates an external write. CRM summarization is Level 0. Autonomous CRM modification may be Level 3 or Level 4 depending on what fields can change and what those fields trigger.</p><p>Code generation is Level 2. Generating code, merging it automatically and deploying it to production is an entirely different risk proposition.</p><p>A fraud recommendation is advisory. Autonomously rejecting a customer payment may be a material operational decision. A finance agent preparing a journal is not equivalent to one posting the journal. A security agent identifying excessive privileges is not equivalent to one revoking privileged access.</p><p>The nouns barely change. The authority does. That is precisely why AI inventories organized only around “use cases,” models or business functions are becoming inadequate. Two implementations carrying the same label can occupy completely different risk classes.</p><h3>Autonomy Gets the Attention. Authority Creates the Exposure.</h3><p>The industry frequently uses autonomy as the shorthand for agentic risk. That is incomplete. <strong><em>Autonomy</em></strong><em> </em>describes how independently a system can reason, plan and decide. <strong><em>Authority</em></strong><em> </em>describes what it can change once it has decided. Those dimensions can move independently.</p><p>A sophisticated research agent may reason through dozens of steps, invoke multiple retrieval tools and produce complex conclusions while remaining completely read-only. A much simpler automation may follow elementary rules and possess permission to transfer money. Which one deserves the stronger operational controls?</p><p>Model capability benchmarks will not answer that question. A less capable model with excessive authority can be more dangerous than a more capable model operating inside carefully engineered boundaries. Security practitioners have already started describing this problem in terms such as excessive agency. OWASP’s guidance for AI agents emphasizes least privilege, constrained tool permissions, human oversight for higher-risk actions and monitoring of agent behavior. The architectural lesson is broader. Do not confuse intelligence with power.</p><h3>Before an Agent Goes Live</h3><p>Every consequential agent should be able to survive a fairly simple executive interrogation.</p><blockquote>What is the agent’s identity, and who owns it?</blockquote><blockquote>On whose behalf is it acting?</blockquote><blockquote>Which systems can it read, and which can it modify?</blockquote><blockquote>What is its maximum permitted Write-Back Level?</blockquote><blockquote>What is the largest consequence of one incorrect action?</blockquote><blockquote>What happens if it repeats that action 500 times before somebody notices?</blockquote><blockquote>Can every action be reversed?</blockquote><blockquote>Which actions require genuine human judgment rather than ceremonial approval?</blockquote><blockquote>What transaction, rate or execution limits exist?</blockquote><blockquote>Can the agent escalate its own permissions?</blockquote><blockquote>How are consequential actions recorded and reconciled?</blockquote><blockquote>How quickly can its authority be revoked?</blockquote><blockquote>And when the system behaves outside expectations, which human remains accountable?</blockquote><p>That is not governance paperwork. It is basic architecture. No production service would be considered well designed if nobody could explain its privileges, failure modes, dependencies and recovery mechanisms. An AI agent should not receive a waiver because its interface happens to speak English.</p><h3>The Next AI Control Plane Will Govern Delegated Authority</h3><p>The most important enterprise AI architecture over the next few years may not be another model gateway. It may be the control layer that connects agent identity, delegated authority, policy enforcement, transaction boundaries, evidence, monitoring and revocation. Identity tells the enterprise which digital actor has arrived. Policy determines what authority it receives. The Write-Back Risk Ladder expresses how consequential that authority is. Runtime controls constrain its actions. Evidence tells us what actually happened. And revocation gives the enterprise a way to stop the actor when assumptions fail. That is a much more useful framing than treating every AI system as a variation of the same model-governance problem.</p><p>Hallucinations still matter. Evaluation still matters. Model quality still matters. But the operational stakes change dramatically when an imperfect reasoning system acquires credentials. For years we have asked whether we can trust what a model tells us. Agentic systems force a harder question: <strong>How much of the enterprise are we prepared to let them change?</strong></p><p>That is the boundary executives should govern. Because model errors become enterprise incidents only when authority gives them somewhere consequential to go. <strong>The next frontier of AI governance is not better answers. It is disciplined digital authority.</strong></p><p><strong>How is your organization handling AI write-back risks?</strong> Let’s discuss in the comments below. Follow my Medium profile for weekly insights on enterprise AI governance.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8037e05dc743" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Models Can Reason. Who Authorizes the Consequences?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-snippet">The next enterprise AI architecture problem sits between governance and execution: turning policies, permissions, context and evaluation&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@vikram.vishal/models-can-reason-who-authorizes-the-consequences-0aa0af106db2?source=rss-ebf8285a9567------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@vikram.vishal/models-can-reason-who-authorizes-the-consequences-0aa0af106db2?source=rss-ebf8285a9567------2</link>
            <guid isPermaLink="false">https://medium.com/p/0aa0af106db2</guid>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[technology-strategy]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[enterprise-architecture]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[Vikram Vishal]]></dc:creator>
            <pubDate>Tue, 11 Aug 2026 18:32:38 GMT</pubDate>
            <atom:updated>2026-08-11T18:32:38.318Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Your AI Adoption Dashboard Is Lying to You]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@vikram.vishal/your-ai-adoption-dashboard-is-lying-to-you-5c108bd62ea9?source=rss-ebf8285a9567------2"><img src="https://cdn-images-1.medium.com/max/1536/1*BYpYNrE2LHgsdN7D9oCO9g.png" width="1536"></a></p><p class="medium-feed-snippet">Enterprise value does not come from more pilots or agents. It appears when leaders redesign important workflows, assign decision rights&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@vikram.vishal/your-ai-adoption-dashboard-is-lying-to-you-5c108bd62ea9?source=rss-ebf8285a9567------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@vikram.vishal/your-ai-adoption-dashboard-is-lying-to-you-5c108bd62ea9?source=rss-ebf8285a9567------2</link>
            <guid isPermaLink="false">https://medium.com/p/5c108bd62ea9</guid>
            <category><![CDATA[ai-strategy]]></category>
            <category><![CDATA[enterprise-transformation]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <dc:creator><![CDATA[Vikram Vishal]]></dc:creator>
            <pubDate>Mon, 03 Aug 2026 20:04:20 GMT</pubDate>
            <atom:updated>2026-08-03T20:04:20.007Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Your Tax Team Does Not Have a Calculation Problem. It Has a Data Problem.]]></title>
            <link>https://medium.com/@vikram.vishal/your-tax-team-does-not-have-a-calculation-problem-it-has-a-data-problem-a4ee3a295a82?source=rss-ebf8285a9567------2</link>
            <guid isPermaLink="false">https://medium.com/p/a4ee3a295a82</guid>
            <category><![CDATA[data-strategy]]></category>
            <category><![CDATA[tax-technology]]></category>
            <category><![CDATA[digital-taxes]]></category>
            <category><![CDATA[tax-transformation]]></category>
            <dc:creator><![CDATA[Vikram Vishal]]></dc:creator>
            <pubDate>Sun, 02 Aug 2026 10:41:41 GMT</pubDate>
            <atom:updated>2026-08-03T09:04:48.268Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Why CFOs need a tax data strategy before e-invoicing, Pillar Two, ERP modernization and AI turn fragmented data into enterprise risk</strong></h3><p>A multinational closes its quarter. The enterprise resource planning system is operational. The general ledger balances. Finance has completed its controls.</p><p>Tax is still waiting.</p><p>One team needs transaction-level invoices from accounts payable. Another is reconciling legal entities to tax registrations. Someone is requesting employee work-location data from payroll. Product descriptions must be converted into tax and customs classifications. Intercompany charges need to be connected to agreements, counterparties and transfer-pricing characterizations. The business has the data. It simply does not have it in a form that tax can reliably use.</p><p>Across tax, finance, data and engineering programs, I have repeatedly seen this contradiction: an organization can run sophisticated global operations and still depend on manual intervention to explain its tax position. That is why most organizations do not primarily have a tax calculation problem. They have a tax data problem.</p><h3><strong>Tax data strategy in plain language</strong></h3><p>A tax data strategy is the enterprise blueprint for identifying, capturing, governing, transforming, controlling and delivering the information required to calculate, report, explain and manage tax outcomes. It answers several practical questions.</p><p>What information does each tax obligation require? Where should that information originate? Who owns the critical attributes? Which definitions and taxonomies should apply? How should the data be validated and reconciled? How will tax rules and transformations be documented? How will lineage, evidence, security and retention be maintained?</p><p>Most importantly, it determines whether the same controlled data can support multiple outcomes, including tax compliance, provisions, planning, audit defense, analytics and appropriately governed AI.</p><p>A tax data strategy is therefore not the same as a tax technology strategy. A technology strategy selects platforms, applications and integration tools. The data strategy defines what those platforms must receive, preserve, transform and produce. It is not an ERP strategy either. ERP decisions matter greatly, but tax information also originates in procurement, payroll, treasury, billing, logistics, customer systems, core banking, trading platforms, tax engines and third-party providers.</p><p>Nor is it simply a tax warehouse. A repository without ownership, controls, common definitions and issue resolution may centralize inconsistency rather than eliminate it.</p><p>The tax operating model defines how tax work is organized. Enterprise data governance establishes wider decision rights and standards. A tax data strategy connects these disciplines around the information required to support tax outcomes.</p><h3><strong>Finance data is not automatically tax-ready</strong></h3><p>Accounting and tax use many of the same transactions, but they often ask different questions of them. Finance may require the posting date. Tax may require the invoice date, supply date, payment date or statutory tax point. A ledger may identify the legal entity. Indirect tax may also require the relevant tax registration, establishment, ship-from location, ship-to location and place-of-supply evidence. A product description may satisfy an invoice-processing team while remaining inadequate for taxability, exemption eligibility or customs classification. A vendor master can support payment while lacking withholding status, beneficial-owner information, tax residency or documentation-expiry dates. A consolidated balance may support financial reporting while tax needs local statutory ledgers, permanent-establishment attribution, tax-basis information and historical book-to-tax adjustments.</p><p>Financial institutions face additional fragmentation. A payment, trade or customer relationship may need to be connected to product classification, legal entity, booking location, counterparty documentation, withholding treatment, information-reporting status and jurisdiction. Those attributes may be distributed across core banking, securities, customer onboarding, payments, treasury, entity-management and finance systems. Tax data therefore tends to demand greater granularity, more jurisdictional context, historical rule versions, non-financial attributes and stronger transaction-level evidence than conventional management reporting.</p><p>Hence, the tax return is an aggregation. The defensible tax position lies underneath it.</p><h3><strong>Why this has become urgent</strong></h3><p>Three forces are converging: regulatory change, enterprise transformation and accelerating use of analytics and AI.</p><p><strong>Regulators are moving closer to the transaction</strong></p><p>The OECD’s <a href="https://www.oecd.org/en/publications/tax-administration-3-0-the-digital-transformation-of-tax-administration_ca274cc5-en.html">Tax Administration 3.0</a> vision describes a future in which tax processes become increasingly embedded within the systems that businesses and individuals use naturally, rather than operating only as periodic and separate interactions with tax authorities.</p><p>That vision is becoming operational. The OECD’s 2026 guidance on <a href="https://www.oecd.org/en/publications/digital-continuous-transactional-reporting-for-value-added-tax_34c88c39-en.html">digital continuous transactional reporting for VAT</a> describes the near-real-time transmission of invoice or transaction information to tax authorities. It also highlights the complexity created by regimes developing differently across jurisdictions. In the European Union, the VAT in the Digital Age package was adopted in March 2025. Digital reporting requirements for specified cross-border business-to-business transactions are scheduled to apply from 1 July 2030. Relevant domestic real-time reporting systems must align with the European framework by January 2035.</p><p>The strategic implication is larger than any individual mandate. Tax authorities increasingly expect structured, granular and machine-processable information rather than periodic summaries assembled after the reporting period.</p><p><strong>International tax requires broader and more connected datasets</strong></p><p>The OECD’s Pillar Two framework applies a 15% minimum effective tax rate to in-scope multinational groups meeting the €750 million consolidated-revenue threshold.</p><p>The calculation operates at jurisdictional level. It draws on financial accounts, entity structures, covered taxes, deferred tax information, payroll, tangible assets and numerous rule-based adjustments.</p><p>The GloBE Information Return formalizes a standardized reporting structure that includes entity identifiers, jurisdictional information, financial accounting income or loss, covered taxes and detailed adjustments.</p><p>Country-by-country reporting already requires qualifying multinational groups to report specified financial, tax, employee and tangible-asset information by jurisdiction.</p><p>These are not merely tax-rule exercises.</p><p>They are data-integration, classification and traceability exercises involving finance, legal-entity management, consolidation, payroll, fixed assets and local tax records.</p><p><strong>Tax authorities and businesses are adopting analytics and AI</strong></p><p>The OECD’s 2025 digitalization study covered the 54 jurisdictions in its Forum on Tax Administration. Seventy-two percent of participating administrations reported using artificial intelligence, while 85% reported assessing the quality of taxpayer-reported data.</p><p>Common AI applications included fraud detection and risk assessment. The United States Internal Revenue Service has similarly described advanced compliance analytics as a means of identifying higher-risk cases involving large corporations, complex partnerships and sophisticated taxpayers. Businesses recognize the pressure but remain constrained by execution.</p><p>EY’s 2025 survey of 1,600 tax and finance executives across 30 jurisdictions and 22 industries found that 45% considered the inability to execute a sustainable data, AI and technology plan their largest barrier. Only 16% were very confident in their ability to execute their data strategy.</p><p>The conclusion is straightforward:</p><blockquote><em>AI does not repair uncontrolled tax data. It increases the value of controlled data and the consequences of using poor data at scale.</em></blockquote><h3><strong>Why ERP alone does not solve the problem</strong></h3><p>ERP platforms are indispensable. A properly designed ERP can capture tax codes, posting logic, legal entities, invoice attributes, account mappings and jurisdictional information while supporting strong transactional controls.</p><p>The issue is not that ERP is inherently inadequate. The issue is that enterprise tax requirements rarely stop at one ERP. Multinationals often operate several ERP instances because of acquisitions, geographic autonomy or incomplete platform consolidation. Global templates can simplify finance while suppressing local tax nuances. Historical information remains in retired systems. Tax logic may be distributed among ERP configurations, external tax engines, custom applications and spreadsheets. Important attributes also sit outside the core finance platform.</p><p>Worker location may sit in a human-capital system. Beneficial-owner and residency information may sit in onboarding platforms. Incoterms and goods-movement details may sit in logistics systems. Foreign-exchange information may come from treasury. Tax depreciation history may remain in specialist fixed-asset tools. Agreements and entity relationships may be maintained by legal teams. Customer-location evidence may sit in digital-commerce platforms. Government acknowledgements may remain in country-specific portals.</p><p>An ERP modernization creates an excellent opportunity to embed tax requirements upstream. But migrating existing fields and configurations without resolving ownership, definitions, master data and controls can reproduce the old problem on a newer platform. ERP is a critical source and processing platform. It is not, by itself, an enterprise tax data strategy.</p><h3><strong>Manual preparation, <em>always</em>, succeeds by hiding the underlying weakness</strong></h3><p>Manual tax processes survive because experienced professionals compensate for weak systems. They know which report is unreliable. They remember that one entity uses a different mapping. They recognize a vendor whose name appears in four forms. They understand why a formula must be overridden this quarter.</p><p>That expertise is valuable. But when professional memory becomes the primary control mechanism, the organization acquires hidden risk. Copy-and-paste transformations create version problems. Undocumented mappings weaken reproducibility. Repeated cleansing consumes highly skilled capacity. Formula errors may remain invisible until reconciliation or audit. Supporting evidence becomes scattered across email, folders and workpapers. Key-person dependency grows with every workaround.</p><p>Spreadsheets remain entirely appropriate for exploratory analysis, controlled exception review, prototyping and small-volume calculations. They become problematic when they serve simultaneously as the integration platform, rules engine, workflow system, evidence repository and control framework for a recurring multi-entity regulatory process.</p><p><strong>The spreadsheet is often not the original defect. It is the footprint left by a data issue the enterprise never assigned anyone to resolve.</strong></p><p>PwC’s 2025 global survey of 1,205 senior buyers of tax advisory services found that 47% viewed improved data quality as a priority for the following three years. The strategic question is therefore not whether the organization uses spreadsheets. It is whether its most material tax outputs depend on transformations that cannot be consistently governed, tested and reproduced.</p><h3><strong>From periodic preparation to continuous readiness</strong></h3><p>The traditional operating model begins after the reporting period. Transactions occur throughout the business. Finance captures them primarily for accounting and operational purposes. Tax receives extracts after month-end, quarter-end or year-end. Teams infer missing attributes, cleanse files, map accounts, reconcile balances, calculate adjustments and assemble evidence.</p><p>The same exercise begins again during the next reporting cycle. The modern model moves tax requirements upstream. Tax-sensitive attributes are captured closer to transaction creation. Data contracts specify mandatory fields, formats and quality thresholds. Automated pipelines integrate information from multiple systems. Rules test completeness, consistency and tax treatment. Reconciliations run throughout the period. Exceptions are directed to accountable owners. Curated tax datasets are preserved with lineage, approval evidence and rule versions.</p><p>This does not mean every tax process must operate in real time. A corporate income-tax provision may tolerate different latency from an e-invoicing clearance process. Transfer-pricing analysis may not require event-stream processing. Some local calculations may remain periodic. The goal is <strong>fit-for-obligation readiness</strong>: data that is sufficiently timely, complete, controlled, explainable and reusable for the relevant requirement. Translating that principle into practice requires more than a reporting database. It requires an end-to-end data chain that preserves source evidence, standardizes enterprise information, applies tax context, executes controls and publishes reusable tax data products.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*hGqvSbymSxqNUr3tTV_qiw.png" /><figcaption><strong>Figure 1: Illustrative enterprise tax data architecture — from source transaction to controlled tax output</strong></figcaption></figure><p><em>An illustrative multinational tax data pipeline connecting operational and financial systems to governed tax data products, regulatory submissions, analytics and controlled AI applications. Enterprise governance, security, lineage, retention and data-quality controls operate across every layer.</em></p><h3><strong>Where reusable tax data creates value</strong></h3><p>The strongest business case rarely comes from automating one return. It comes from reusing the same governed data elements across several obligations.</p><p><strong>Corporate tax provision</strong></p><p>A corporate tax provision may require trial balances, legal-entity hierarchies, permanent and temporary differences, deferred tax information and book-to-tax adjustments. The common failure is not necessarily the calculation logic. It is the late arrival of inconsistent mappings and manually maintained adjustments. A reusable provision dataset can accelerate close activities, improve effective-tax-rate analysis and create a stronger bridge between financial reporting, returns and audit support.</p><p><strong>Corporate income-tax returns</strong></p><p>Return preparation requires local ledgers, tax adjustments, elections, prior-year attributes and jurisdiction-specific information. In many organizations, return data is collected separately from provision data even when the underlying information overlaps substantially. A reusable tax dataset reduces duplicate extraction and creates greater consistency between provision, return and audit positions.</p><p><strong>VAT and GST reporting</strong></p><p>Indirect-tax reporting requires invoice-line information, tax point, tax registrations, tax codes, supply locations, rates, exemptions and recoverability attributes. Typical problems include invalid tax codes, missing jurisdictional data and incorrect registration mappings. A governed VAT or GST transaction product can support filing, reconciliation, anomaly detection, audit requests and recovery analysis from the same controlled foundation.</p><p><strong>E-invoicing and continuous transaction controls</strong></p><p>E-invoicing requires structured invoice fields, valid identifiers, accurate tax treatment and, in some jurisdictions, government clearance or acknowledgement. Missing or invalid fields can cause invoice rejection, payment disruption or late correction. A controlled e-invoicing dataset allows issues to be detected earlier and government responses to be connected back to the originating transaction.</p><p><strong>Withholding tax</strong></p><p>Withholding calculations depend on vendor residency, income type, beneficial ownership, treaty documentation, payment information and documentation-validity dates. Problems often arise because payment can proceed even when tax documentation is incomplete or expired. A reusable withholding dataset improves classification consistency, treaty application, reporting and exception management.</p><p><strong>Pillar Two</strong></p><p>Pillar Two requires entity structures, jurisdictional results, financial accounting income, covered taxes, deferred tax data, payroll and tangible assets. The information is commonly split across consolidation platforms, local finance systems, human-resources systems and tax workpapers. A controlled Pillar Two dataset makes jurisdictional calculations more repeatable and improves the traceability of adjustments and elections.</p><p><strong>Country-by-country reporting</strong></p><p>Country-by-country reporting draws on revenue, profit, taxes, employees and tangible assets by jurisdiction. Legal-entity and jurisdiction mappings are frequent sources of inconsistency, particularly where tax, finance and corporate-secretarial systems maintain different hierarchies. Reusable entity and jurisdiction data improves alignment across country-by-country reporting, Pillar Two and other tax-transparency obligations.</p><p><strong>Transfer pricing</strong></p><p>Transfer-pricing calculations require intercompany transactions, counterparties, agreements, cost pools, service classifications and allocation drivers. Common failures include unclear transaction characterizations, inconsistent counterparty identifiers and poorly governed allocation logic. A curated transfer-pricing dataset supports calculation, monitoring, documentation and audit defense.</p><p><strong>Customs and trade</strong></p><p>Customs treatment depends on product classification, origin, valuation, Incoterms, movement information and supply-chain relationships. A commercial product master may not contain the detail needed for customs classification or preferential-origin analysis. A tax-enriched product dataset can support duty calculations, classification governance and supply-chain scenario modelling.</p><p><strong>Fixed-asset taxation</strong></p><p>Tax depreciation requires asset class, location, acquisition date, disposal information, book basis, tax basis and historical adjustments. Financial fixed-asset registers may not retain the complete tax history. A reusable asset-tax dataset supports tax depreciation, deferred tax, property-related obligations and tax-basis reporting.</p><p><strong>Employment and payroll tax</strong></p><p>Employment-tax obligations depend on payroll, work location, employing entity, mobility data and benefit classifications. Problems emerge when human-resources and payroll systems disagree on where an employee works or which entity bears the cost. A controlled employment-tax dataset improves withholding, mobility compliance and jurisdictional reporting.</p><p><strong>Audit defense</strong></p><p>Audit requests require more than the number reported in a return. They may require underlying transactions, transformation logic, approvals, correspondence, evidence and the precise rule version used at the time of filing. A governed audit dataset allows the organization to reproduce the position rather than reconstruct it under regulatory pressure.</p><p><strong>The architecture should follow the obligation</strong></p><p>There is no universal target architecture for tax data. A centralized tax warehouse can provide strong control and consistent reporting. It may, however, duplicate significant volumes of enterprise data and introduce batch latency.</p><p>A cloud Lakehouse can combine structured and semi-structured information at scale. Its success depends on disciplined metadata, access control, lineage and data-product governance.</p><p>Federated data products can place accountability closer to the business domains that create the information. This model works best where enterprise definitions, interoperability standards and cross-domain controls are already mature. A query-in-place or virtualized architecture can reduce replication. It may create concerns around source availability, performance, point-in-time reproducibility and dependence on operational systems.</p><p>Most large enterprises will require a hybrid approach. Selected transaction data will be copied into controlled tax products. Some information will continue to be accessed from authoritative enterprise platforms. A smaller number of genuine exceptions will require controlled manual upload.</p><p>Architecture should follow regulatory latency, data volume, security, data-residency restrictions, source complexity, existing platforms and the required level of traceability. The architecture decision should not begin with a preferred technology. It should begin with the obligation the enterprise must fulfil and the evidence it must preserve.</p><p><strong>Tax data should be managed as an enterprise product</strong></p><p>Consider a VAT transaction data product. Its purpose is not simply to feed a tax return. It provides a controlled transaction-level record of output and input tax across relevant entities and tax registrations. The product may require invoice identifier, invoice line, supplier, customer, legal entity, tax registration, tax point, supply jurisdiction, product or service classification, taxable amount, tax rate, tax amount, currency and source-system identifier.</p><p>Its controls should test whether invoice-line keys are unique, tax-code and registration combinations are valid, place-of-supply fields are complete, duplicate invoices exist and transactions reconcile to subledgers and the general ledger. It should also connect government acknowledgements or rejections back to the original invoice where relevant. The consumers may include indirect-tax compliance, controllership, accounts payable, accounts receivable, treasury, internal audit and tax analytics teams.</p><p>The product should have a named owner. It should have documented sources, lineage, quality measures, service expectations, retention rules, security classification and a governed process for managing changes. Without these elements, it is merely a dataset. With them, it becomes a dependable enterprise capability.</p><p><strong>Accountability must cross functional boundaries</strong></p><p>Tax data failures frequently occur between functions rather than within a single department. Tax understands the regulatory requirement but does not control procurement, payroll, billing or customer platforms. Technology operates the systems but may not understand why a field is material for tax.</p><p>The data office defines governance but may not include tax in its prioritization model. The finance function reconciles the ledger but may not own the non-financial attributes that determine tax treatment. A practical operating model should divide responsibility clearly. Tax should define tax data requirements, approve tax rules and establish risk tolerances.</p><p>Business and process owners should ensure that required attributes are captured when suppliers, customers, products, employees and transactions are created. Master-data owners should maintain legal entities, tax registrations, accounts, products, vendors, customers and jurisdictional mappings. Technology and data-engineering teams should build pipelines, integrations, controls and monitoring. Data-product owners should track quality, service levels and issue resolution.</p><p>The owner of the originating business process should resolve upstream exceptions. Tax should support the interpretation, but it should not become the permanent repair team for every source-system defect. Tax and the appropriate finance controller should certify regulatory outputs.</p><p>Architecture changes should involve technology, data governance, tax, finance, security and relevant business owners. Tax should define requirements and risk tolerance. It should not be expected to own every upstream system that creates tax-relevant information.</p><p><strong>Measure regulatory fitness, not theoretical perfection</strong></p><p>A useful tax data-quality framework covers completeness, accuracy, validity, consistency, uniqueness, timeliness, traceability, reconciliation and regulatory fitness. These dimensions should be translated into operational measures. For example, an organization may track the percentage of invoices carrying valid tax codes, the proportion of vendors with complete and current withholding attributes, and the percentage of transactions mapped to a valid tax registration.</p><p>It may monitor the number and value of unreconciled records, the value of manual tax journals, the percentage of regulatory fields sourced automatically and the volume of post-close corrections. It should also measure the number and age of open data-quality exceptions. For audit readiness, two measures are particularly revealing: the time required to reproduce a filed return and the percentage of material transformations supported by documented source-to-output lineage. These are not universal industry benchmarks. They are sample organizational measures that should be calibrated to materiality, regulatory exposure and operating complexity.</p><p>The objective is not perfect data everywhere. That would be expensive and usually unrealistic. The objective is controlled data for material tax obligations, supported by transparent exceptions and proportionate risk thresholds.</p><h3><strong>Building the investment case</strong></h3><blockquote>A balanced tax data business case should address four forms of value.</blockquote><p><strong>Risk reduction</strong></p><p>Better tax data can reduce filing errors, strengthen controls, improve audit defense and lower exposure to avoidable penalties and interest. It can also improve management’s confidence that reported positions are supported by consistent source evidence.</p><p><strong>Efficiency</strong></p><p>Reusable data reduces repeated extraction, cleansing, mapping and reconciliation. It can shorten close and filing cycles, reduce dependency on individual knowledge and allow tax professionals to spend more time on review, interpretation and decision support.</p><p><strong>Insight</strong></p><p>Controlled transaction-level information enables cash-tax forecasting, effective-tax-rate analysis, recovery identification, scenario modelling and risk monitoring. It can help management understand not only what tax was reported, but why it changed and which transactions or jurisdictions caused the movement.</p><p><strong>Strategic readiness</strong></p><p>A tax data capability improves the organization’s ability to respond to regulatory change, ERP modernization, acquisitions, divestitures, market expansion and controlled AI adoption. The financial case should be built from the organization’s own evidence. Relevant inputs include the hours spent collecting and cleansing data, the number of manual transformations, the cost of external support, filing corrections, rejected submissions, audit-response effort, reconciliation time and the number of systems and spreadsheets involved.</p><p>The organization should also consider transaction-volume growth, new jurisdictions and decisions delayed because tax information was unavailable. A credible business case does not begin with a generic automation benchmark. It begins with the cost, risk and opportunity loss embedded in the current process.</p><h3><strong>A practical roadmap</strong></h3><p><strong>Phase 1: Diagnose</strong></p><p>Begin by inventorying material obligations, critical datasets, source systems, manual interventions, control weaknesses and unresolved data-quality issues. Prioritize use cases according to materiality, regulatory urgency, reuse potential and implementation feasibility. A 30-day diagnostic can often reveal where the most significant problems sit without launching a large transformation.</p><p><strong>Phase 2: Design</strong></p><p>Define the target tax data model, authoritative sources, ownership, terminology, architecture, controls, data contracts and governance forums. The design should explicitly identify which fields must be captured at source and which tax attributes may be derived later.</p><p><strong>Phase 3: Pilot</strong></p><p>Select one material use case with measurable pain and reusable data. Build a minimum viable tax data product with automated ingestion, reconciliation, lineage and exception management. The pilot should demonstrate a business outcome, not simply technical connectivity.</p><p><strong>Phase 4: Scale</strong></p><p>Reuse shared components such as legal entity, tax registration, account, currency, jurisdiction, vendor and transaction definitions across additional obligations. Integrate the work with ERP, finance transformation and enterprise data programs rather than building a separate tax estate wherever possible.</p><p><strong>Phase 5: Optimize</strong></p><p>Once the underlying datasets are demonstrably fit for purpose, introduce continuous monitoring, forecasting, anomaly detection and narrowly controlled AI applications. Do not begin with a multi-year promise to transform every tax process. Begin with one high-value obligation and prove that the organization can turn fragmented information into a controlled, reusable and defensible data product.</p><h3><strong>How tax data programs fail</strong></h3><blockquote>In my limited experiences — Tax data programs tend to fail in predictable ways.</blockquote><p><strong>Buying technology before defining the problem</strong></p><p>A platform is selected before the organization has agreed which obligations, risks and datasets matter.</p><p><strong>Corrective action:</strong> Begin with material outcomes, tax requirements and current-process evidence.</p><p><strong>Treating tax as a downstream reporting consumer</strong></p><p>Tax receives data only after business and system decisions have already been made.</p><p><strong>Corrective action:</strong> Include tax during process, master-data, ERP and architecture design.</p><p><strong>Ignoring master data</strong></p><p>The program focuses on transactions while legal entities, registrations, products, vendors and accounts remain inconsistent.</p><p><strong>Corrective action:</strong> Assign named owners and controls to tax-critical master data.</p><p><strong>Automating a broken process</strong></p><p>Every manual step is translated into code without questioning whether the step is still necessary.</p><p><strong>Corrective action:</strong> Simplify, standardize and remove avoidable transformations before automating.</p><p><strong>Creating a tax data lake without ownership</strong></p><p>Data is copied into a new platform, but quality, definitions and issue resolution remain unclear.</p><p><strong>Corrective action:</strong> Establish named data products, owners, consumers and service expectations.</p><p><strong>Replicating every available source field</strong></p><p>The program ingests large volumes of data without connecting them to material obligations.</p><p><strong>Corrective action:</strong> Prioritize fields according to tax requirements, control needs and reuse value.</p><p><strong>Omitting reconciliation from the design</strong></p><p>The data pipeline produces output, but no systematic bridge exists from source to ledger to return.</p><p><strong>Corrective action:</strong> Treat reconciliation as a core architecture capability rather than a final review activity.</p><p><strong>Assuming an ERP migration will solve tax automatically</strong></p><p>The organization expects the new platform to resolve tax requirements without converting those requirements into design criteria.</p><p><strong>Corrective action:</strong> Make tax data, rules, controls and acceptance tests explicit within the ERP program.</p><p><strong>Building without tax-user involvement</strong></p><p>Technical teams design outputs without understanding professional judgment, evidence requirements or regulatory risk.</p><p><strong>Corrective action:</strong> Make tax professionals active product owners and control-design participants.</p><p><strong>Underestimating local-country requirements</strong></p><p>A global model suppresses legitimate local attributes, rules and reporting formats.</p><p><strong>Corrective action:</strong> Standardize common data while preserving governed jurisdiction-specific extensions.</p><p><strong>Neglecting security and retention</strong></p><p>Sensitive employee, customer, counterparty and transaction data is consolidated without adequate controls.</p><p><strong>Corrective action:</strong> Design access, privacy, residency, archiving and retention requirements into the architecture.</p><p><strong>Measuring activity rather than outcomes</strong></p><p>The program reports pipelines built and fields ingested but cannot demonstrate lower effort, fewer errors or better audit readiness.</p><p><strong>Corrective action:</strong> Track manual effort, exceptions, corrections, cycle time, reuse and reproducibility.</p><p><strong>Pursuing AI before trusted data exists</strong></p><p>The organization deploys AI on fragmented or poorly controlled datasets.</p><p><strong>Corrective action:</strong> Restrict AI to approved data products with lineage, access controls, quality monitoring and human oversight.</p><h3>Avoiding these failures</h3><p>These failure patterns are rarely caused by a lack of effort. More often, they persist because tax data responsibilities are fragmented across functions, systems and transformation programs, with no executive view of the complete chain. That is why leadership attention matters. Before approving another technology investment, ERP enhancement or automation initiative, executives should first test whether the organization can answer a small set of fundamental questions about ownership, traceability, control and reuse.</p><p><strong>Ten questions executives should ask</strong></p><ol><li>Can we trace every material figure in a filed return to the originating transactions, transformations, rules and approvals?</li><li>How many manual data transformations sit between our ledgers and our most material tax submissions?</li><li>Which tax-critical attributes are missing when transactions are first created, and who is accountable for capturing them?</li><li>Where do tax teams repeatedly cleanse the same entity, account, vendor, customer or product data?</li><li>Can we reproduce a previous filing using the exact data, mappings and rule versions applied at that time?</li><li>Which tax calculations depend on undocumented spreadsheets or knowledge held by one or two individuals?</li><li>Is tax represented when ERP, finance, customer, payroll and enterprise data-platform changes are approved?</li><li>Which tax datasets have sufficient quality, lineage, access control and human oversight to support AI?</li><li>How quickly could we respond to a tax authority requesting transaction-level evidence across several legal entities?</li><li>Are we investing in reusable tax data products, or creating another generation of obligation-specific reports?</li></ol><h3>The Future</h3><p>The most important tax transformation may occur before anybody calculates tax: when a product is classified, a vendor is onboarded, an invoice is issued, an employee changes location or an intercompany transaction is recorded.</p><p><strong>Tomorrow’s tax position will be only as defensible as the data decisions made when today’s transaction enters the enterprise.</strong></p><p><strong>Suggestive actionable starting point</strong></p><p>Within the next 30 days, choose one material tax obligation and document:</p><ol><li>Every source system and spreadsheet used.</li><li>Every manual transformation.</li><li>Every missing attribute inferred by tax.</li><li>Every reconciliation and approval.</li><li>The owner of each critical data element.</li><li>The time required to reproduce the previous submission.</li></ol><p>That exercise usually reveals whether the organization possesses a tax process or a genuine tax data capability. A tightly scoped tax-data readiness diagnostic can also form a credible specialist advisory proposition.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a4ee3a295a82" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>