<?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 Alexander Statnikov on Medium]]></title>
        <description><![CDATA[Stories by Alexander Statnikov on Medium]]></description>
        <link>https://medium.com/@statnikov?source=rss-31a9961fd6bc------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*MN7Vv_Q1EGSLMMXb7y0laA.jpeg</url>
            <title>Stories by Alexander Statnikov on Medium</title>
            <link>https://medium.com/@statnikov?source=rss-31a9961fd6bc------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 22:44:48 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@statnikov/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 Data Is Broken”: The Most Expensive Misdiagnosis in Enterprise AI]]></title>
            <link>https://medium.com/@statnikov/the-data-is-broken-the-most-expensive-misdiagnosis-in-enterprise-ai-1b5aa8f4be7d?source=rss-31a9961fd6bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/1b5aa8f4be7d</guid>
            <category><![CDATA[data]]></category>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Alexander Statnikov]]></dc:creator>
            <pubDate>Thu, 19 Mar 2026 17:36:03 GMT</pubDate>
            <atom:updated>2026-03-19T17:36:03.697Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sMzMu9V6yTzXsHSlz_8zgQ.png" /><figcaption>Image produced by Google Nano Banana 2.</figcaption></figure><p>Over the past 17 years, I have worked with dozens of organizations across financial services, healthcare, and technology. In nearly every one of them, at some point, a senior executive has said some version of the same thing: “The data is broken.”</p><p>It is one of those statements that sounds urgent, specific, and actionable. It implies a clear problem with a clear solution: fix the data, and the business will run better. But in my experience, this statement almost always conflates two fundamentally different problems. The first is an infrastructure problem: pipelines are unreliable, tables are poorly documented, tooling is fragmented, and teams are spending too much time on downstream cleanup rather than building things right at the source. The second is a perception problem: “data is broken” has become a cultural reflex, a default explanation for friction that executives cannot otherwise articulate, and a convenient way to defer accountability for decisions that have nothing to do with data quality.</p><p>The first problem is an engineering challenge. It is hard work, but it is known work, and the modern tooling ecosystem has made it dramatically easier to solve. The second problem is an organizational and cultural challenge. It is harder, slower, and largely invisible to the people who perpetuate it. And it is the one that almost nobody talks about.</p><p>I thought we had moved past this. The data infrastructure landscape has matured enormously. And yet, last year I encountered a modern organization with thousands of tables in Snowflake, no meaningful data lineage, fragmented tooling across teams, and executives who described the situation as “data is broken.” When I asked which specific metrics were unreliable, they could not name them. When I asked whether their weekly and quarterly business reviews were functioning, they confirmed that they were. The numbers they needed to run the business were there. But the narrative persisted.</p><p>This is the misdiagnosis. And in the AI era, it is becoming extraordinarily expensive.</p><h3>The Infrastructure Problem (Mostly Solved)</h3><p>Let me be clear: real data infrastructure problems exist and they are serious. I have walked into organizations where pipelines break constantly, where teams join data on different transaction IDs producing contradictory results, where there is no distinction between a quick analysis someone ran once and an official dataset that the entire company depends on. These are real engineering failures with real business consequences.</p><p>But the playbook for solving them is well established and getting easier every year. Enforce data contracts at the source rather than cleaning up downstream. Establish clear governance around data permanency: make it easy for anyone to create and explore data, but gate the process of publishing a dataset as an official, shared asset that others will depend on. Consolidate on shared semantics and tooling so that every team is working from the same definitions. Invest in lineage and discoverability so that when someone needs a number, they can find the authoritative source rather than guessing among thousands of confusingly named tables.</p><p>The tooling for all of this has matured dramatically. Platforms like Snowflake and Databricks, combined with a rich ecosystem of open-source and enterprise tools for orchestration, cataloging, and governance, mean that a competent team with adequate investment can make meaningful progress in months, not years. And now, AI-powered development tools are further accelerating the work. You can build MCP servers and analytics agents that make thousands of underlying tables navigable and queryable without a massive reengineering effort. The infrastructure problem is not trivial, but it is tractable. A moderate investment and the right engineering culture can get an organization to a solid foundation faster than most executives expect.</p><p>This is not where the money gets wasted.</p><h3>The Perception Problem (Mostly Ignored)</h3><p>The more damaging problem is the one that persists even after the infrastructure is fixed. I have seen this firsthand: a team spends months stabilizing pipelines, implementing contracts, building documentation, and improving data quality across the board. The metrics that matter are reliable. The dashboards are accurate. And then an executive walks into a meeting and says, “The data is still broken.”</p><p>When you press for specifics, the response is vague. It is not “the Q3 retention cohort is miscounted” or “the revenue reconciliation is off by 2%.” It is a general sense that data cannot be trusted. And because the statement is vague, it is unfalsifiable. No amount of infrastructure improvement can disprove a feeling.</p><p>This happens because “the data is broken” has evolved from a technical complaint into something more insidious: a cultural norm. In some organizations, it functions as a universal excuse. A product launch underperformed? The data must have been wrong. A forecast missed? The inputs were unreliable. A strategic bet did not pay off? We did not have the right data to make the decision. It is a remarkably convenient narrative because it shifts accountability away from the decision-maker and onto a team that is rarely in the room to defend itself.</p><p>What makes this especially frustrating is the disconnect between the executive level and the ground level. In my experience, the teams doing the actual work, the analysts, the engineers, the data scientists, are usually collaborating effectively. They know which tables to use, they understand the caveats, and they get things done. The “data is broken” narrative lives primarily in executive meetings, where it serves a political function that has little to do with the actual state of the data.</p><p>I want to be fair here. Not every executive who says “data is broken” is deflecting. Some are genuinely experiencing friction: they ask a question and get conflicting answers, or they cannot get a simple report without a two-week turnaround. These are real problems. But they are usually symptoms of organizational design and prioritization failures, not of broken data. The data team is not the disease. It is often the canary.</p><h3>The Electricity Myth</h3><p>Part of what sustains the perception problem is an unrealistic mental model of what data should be. I have heard executives say, more than once, that “data should be like electricity.” They cite past experiences at other companies where they had the impression that data just worked, invisibly and frictionlessly, and they want to replicate that here.</p><p>This analogy is seductive but deeply misleading. Electricity requires zero interpretation from the consumer. You flip a switch and the light turns on. You do not need to understand voltage, load balancing, or grid architecture. It is a commodity with a binary state: on or off.</p><p>Data is not like that and never will be. Data requires context. A single metric can be calculated multiple ways depending on the business question. Two reports can show different numbers and both be correct, because they are answering slightly different questions or using different time windows or filtering criteria. Revenue recognized is not the same as revenue booked is not the same as revenue collected. Each is “correct” in its own context. The expectation that data should be as frictionless as electricity guarantees permanent disappointment, because even perfect infrastructure cannot eliminate the ambiguity inherent in how businesses measure themselves.</p><p>When executives internalize the electricity metaphor, they interpret any friction, any need for clarification, any instance where two numbers do not match, as evidence that “the data is broken.” In reality, the friction is often evidence that the business is complex and that good decisions require engaging with that complexity rather than expecting it to disappear.</p><h3>Why It Matters Now</h3><p>A few years ago, the cost of the data perception problem was primarily wasted executive attention and demoralized data teams. That was bad enough. But the stakes have changed.</p><p>Data is no longer just fuel for dashboards and quarterly reviews. It is the foundation for AI workflows, agentic automation, and foundational models that are reshaping how enterprises operate. The organizations that will capture the most value from AI are the ones that can move quickly from “we have data” to “our data is powering autonomous decisions at scale.” That transition requires not just clean infrastructure but organizational confidence in data, the willingness to trust data enough to let AI systems act on it.</p><p>An organization stuck in the “data is broken” perception loop cannot make that leap. If executives do not trust the data enough to make their own decisions with it, they will certainly not trust it enough to let an AI agent make decisions autonomously. The perception problem, once a drag on productivity, is now a direct blocker to AI readiness. Every month spent relitigating whether the data is trustworthy is a month not spent building the AI systems that competitors are already deploying.</p><p>Meanwhile, the infrastructure side has gotten dramatically easier. Modern platforms, combined with AI-powered engineering tools, mean that a smaller team can do what used to require a large dedicated organization. The bottleneck has shifted. The hard problem is no longer building the pipes. It is building the organizational trust to use what flows through them.</p><h3>The Practitioner’s Playbook</h3><p>Solving this requires treating the infrastructure problem and the perception problem as two parallel workstreams with different tools, different timelines, and different success metrics. Most organizations invest heavily in the first and ignore the second entirely. That is why they keep hearing “the data is broken” long after the data has been fixed.</p><p>On the infrastructure side, the playbook is relatively straightforward: shift left on quality by enforcing contracts at the source, gate data permanency to prevent the proliferation of ungoverned tables, consolidate shared semantics across teams, and leverage AI tooling to accelerate engineering productivity and make large data estates navigable. This workstream has clear milestones and shows progress in weeks to months.</p><p>The perception side requires a different kind of discipline.</p><p><strong>Force specificity.</strong> Every time someone says “the data is broken,” make it a practice to ask: which metric, in which report, for which decision? Do not accept “data” as the subject of a complaint. Insist on nouns. Which table, which dashboard, which number does not match what you expected? Most of the time, this question reveals either a narrow, fixable issue or a vague sentiment with no technical basis. Both are useful discoveries, but they require completely different responses.</p><p><strong>Establish trust anchors.</strong> Identify the three to five most critical business processes that depend on data, whether that is the weekly business review, the board deck, or the core product metrics, and make them bulletproof. Invest disproportionate effort in ensuring these specific outputs are reliable, well-documented, and understood by the executives who consume them. When leaders experience trustworthy data in the contexts that matter most to them, the broader perception starts to shift. You do not need to fix everything. You need to fix the things that shape executive experience.</p><p><strong>Create transparency rituals.</strong> Publish a regular data health scorecard that shows pipeline reliability, data freshness, known issues, and remediation status. This transforms the conversation from “is the data broken?” (a vibes question with no answer) to “what is the current state of data reliability?” (a measurable question with a specific answer). It also signals that the data team is proactively monitoring quality rather than waiting to be blamed.</p><p><strong>Invest in data literacy.</strong> Not training sessions on SQL. Real education about how metrics are constructed, where legitimate ambiguity lives, and why two reports might show different numbers without either being wrong. Executives who understand that “revenue” can be defined five different ways depending on the context stop interpreting discrepancies as evidence of broken data and start asking better questions.</p><p><strong>Build executive empathy through partnership.</strong> The perception problem often persists because the data team is solving technical problems while executives are experiencing decision-making friction. These are not the same thing. Instead of waiting for executives to articulate what they need, embed with them. Understand their decisions, their cadence, their pain points. The data leader who speaks the language of business decisions rather than pipelines will earn trust faster than the one who delivers a perfect infrastructure no one understands.</p><h3>Conclusion</h3><p>The most important thing I have learned about “the data problem” is that it is almost never one problem. The infrastructure challenge is real but increasingly solvable, and the modern tooling ecosystem has compressed what used to take years into months or even weeks. The perception challenge is real and stubbornly persistent, because it is rooted in organizational culture, executive expectations, and the political convenience of having a universal scapegoat.</p><p>Organizations that invest only in infrastructure will fix their pipelines and still hear “the data is broken” at the next board meeting. The ones that treat the perception problem with the same rigor they apply to their data platforms will build something more valuable than clean tables: they will build organizational confidence in data as a foundation for decision-making and AI-era automation.</p><p>The tools have never been better. The data has never been more important. And the most expensive problem left to solve is not in the warehouse. It is in the room where the decisions are made.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1b5aa8f4be7d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Next Fraud Defense: Why Network Intelligence Is Just the Starting Point]]></title>
            <link>https://medium.com/@statnikov/the-next-fraud-defense-why-network-intelligence-is-just-the-starting-point-c223d342334e?source=rss-31a9961fd6bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/c223d342334e</guid>
            <category><![CDATA[fraud]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[financial-services]]></category>
            <category><![CDATA[risk-management]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Alexander Statnikov]]></dc:creator>
            <pubDate>Fri, 13 Mar 2026 22:20:35 GMT</pubDate>
            <atom:updated>2026-03-13T22:20:35.484Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RPVABi19DiZ3qj2HAZhbIw.png" /><figcaption>Image credits to Google Nano Banana 2.</figcaption></figure><p>A recent <a href="https://plaid.com/new-identity-crisis-ai-fraud-report/">white paper from Plaid</a> makes a compelling case that financial network intelligence represents the next foundation for fraud prevention. The argument is straightforward: as AI degrades traditional identity signals like knowledge-based authentication, document checks, and even biometrics, the most durable fraud signal is a person’s actual financial behavior observed across institutions. Transaction patterns, device relationships, and account activity across banks, fintech apps, and payment platforms form a behavioral footprint that is hard to fabricate and deeply connected to fraud outcomes.</p><p>It is a well-constructed thesis, and directionally correct. Fraud is fundamentally a financial event, and tracking the flow of money across the ecosystem is a better foundation than verifying static identity attributes that can now be synthesized by anyone with access to a generative AI tool. Cross-institutional visibility is a genuine step forward from the siloed, single-platform detection that still dominates the industry.</p><p>But network intelligence, as currently framed, is a starting point rather than a destination. Several of the assumptions underlying this approach are already under pressure, and the emergence of agentic AI will stress them further. If we want fraud defenses that hold up over the next 3 years rather than the next twelve months, we need to examine where the approach falls short and what must be built on top of it.</p><h3>What Has to Be True</h3><p>For financial network intelligence to work as described, four conditions must hold.</p><p>First, the network must have broad coverage. The insights are only as good as the visibility. If the network sees activity across a wide range of institutions and platforms, the signal is strong. If coverage is patchy, the behavioral footprint is incomplete and the patterns become unreliable. In practice, platforms like Plaid have achieved significant coverage of the U.S. financial ecosystem, so this condition is largely met today. It is worth noting, however, that regulatory fragmentation (particularly across jurisdictions), competitive dynamics between banks, and evolving consumer consent requirements will keep coverage uneven globally. The network will always have blind spots.</p><p>Second, behavioral signals must remain difficult to fake. The core premise is that financial footprints are grounded in real banking relationships and real transactional activity, making them significantly harder to fabricate than static identity data. This is true relative to what came before. But “harder” is not “hard enough” when the adversary has access to the same generative tools that are reshaping every other domain.</p><p>Third, correlational pattern detection must be sufficient to catch sophisticated fraud. Network intelligence identifies anomalies: unusual transaction velocity, abrupt behavioral shifts, unexpected account connections. These are statistical signals. They flag activity that deviates from a baseline. This works well for opportunistic fraud, but it is structurally weaker against adversaries who are specifically designing their behavior to look normal.</p><p>Fourth, the approach must scale to a world where AI agents, not humans, are executing financial transactions. The paper briefly mentions agentic commerce but does not grapple with its implications for the behavioral model it proposes.</p><p>The first condition is reasonably well satisfied today. The remaining three are where the approach begins to show its limits.</p><h3>When the Footprint Is Fake</h3><p>The promise of financial behavioral data is that it reflects real activity that is costly and time-consuming to fabricate. That was true when fraud was primarily a manual operation. It becomes less true as AI makes it possible to build convincing financial histories at scale.</p><p>Consider synthetic identity fraud, already one of the fastest-growing fraud categories in the United States. A well-constructed synthetic identity can open real accounts, build a genuine transaction history over months, establish relationships across multiple platforms, and then execute a bust-out. From the perspective of network intelligence, this identity has a consistent, plausible financial footprint. It passes the behavioral test precisely because the adversary invested the time to build the footprint before attacking.</p><p>Now scale that with AI. An adversary can use generative tools to create hundreds of synthetic personas, each with a distinct behavioral profile, and run them through the financial ecosystem in parallel. The footprints are not spoofed in the traditional sense. They are real transactions on real accounts, orchestrated by software that understands what “normal” looks like.</p><p>The challenge deepens with agentic commerce. When AI agents transact autonomously on behalf of users, the behavioral patterns in the network are generated by software, not by humans. The spending cadence, the merchant selection, the transfer patterns will all reflect algorithmic logic rather than individual human habits. In this world, distinguishing a legitimate agent’s financial behavior from a compromised or malicious agent’s behavior is a fundamentally different problem than the one network intelligence was designed to solve. You are no longer authenticating a person through their behavior. You are trying to verify the integrity of a delegation chain: did the user authorize this agent, is the agent operating within its intended scope, and has the agent been tampered with? Behavioral analytics cannot answer these questions.</p><h3>Distilling the Defense</h3><p>There is a deeper structural vulnerability that the network intelligence approach shares with most fraud detection systems: it can be reverse-engineered.</p><p>In machine learning, there is a technique called model distillation. You do not need access to a model’s internal weights to learn its behavior. You simply send it many inputs, observe its outputs, and train a simpler model that approximates its decisions. The original model’s knowledge is effectively extracted through its responses.</p><p>The same logic applies to fraud defenses. An adversary does not need to steal your fraud model. They need to probe it. By submitting hundreds or thousands of synthetic applications or transactions with systematically varied profiles, an attacker can observe which get flagged and which pass through. Over time, this maps the decision boundary of the fraud system. The attacker learns what “good” looks like to the defender and can then manufacture behavior that stays just inside the boundary.</p><p>This is not theoretical. Adversarial attacks on ML systems are well-documented in the research literature, and the economics of fraud make this approach highly attractive. The cost of generating synthetic probes with AI is approaching zero. The reward for successfully mapping a fraud system’s blind spots is enormous.</p><p>The implication is that any fraud defense built on a fixed set of signals, no matter how rich or cross-institutional, has a half-life. The signals themselves become the attack surface. The more consistent and predictable the detection logic, the easier it is to reverse-engineer. This does not mean network intelligence is worthless. It means it must be treated as one layer in a deeper defense, not the foundation by itself.</p><h3>Building a Thicker Defense</h3><p>If network intelligence is necessary but not sufficient, what does a more resilient fraud defense look like? Three principles stand out.</p><p>The <strong>first</strong> is signal diversity across heterogeneous networks. Financial transaction data is one behavioral layer, but fraud leaves traces across many ecosystems simultaneously. Device intelligence networks, IP reputation systems, telecom signals, merchant and marketplace networks, and app usage graphs all carry complementary information. A fraudster who builds a plausible financial footprint may still exhibit anomalies in their device graph, their IP behavior, or their app usage patterns. The key insight is that each additional network dimension increases the cost and complexity of fabrication for the adversary. Faking one behavioral signal is feasible. Faking five simultaneously, across independent networks with different data structures, is exponentially harder.</p><p>Critically, many of these networks are closed ecosystems. The behavioral signals inside Amazon, Shopify, or major telecom carriers are not available through any Plaid-like intermediary. This is precisely why a single-network-hub model, however well-covered, cannot provide complete protection. The most resilient architecture fuses signals from multiple independent networks rather than deepening reliance on any single one.</p><p>The <strong>second principle</strong> is combining first-party and third-party data. Third-party network intelligence provides breadth: visibility across institutions and platforms. First-party data provides depth: the proprietary signals from your own customer relationships, your internal transaction patterns, your product-specific behavioral baselines. Neither alone is sufficient. Third-party data tells you what the customer looks like across the ecosystem. First-party data tells you what the customer looks like in your specific context, and that context contains signals that are invisible to any external network and therefore inaccessible to an adversary probing from the outside.</p><p>Organizations that outsource their fraud intelligence entirely to a third-party network are effectively making their defense fully observable to any attacker who can probe that same network. Proprietary first-party signals create an asymmetry: the adversary cannot map what they cannot see.</p><p>The<strong> third principle</strong> is the shift from behavioral pattern matching to intent verification and causal reasoning. Correlational signals tell you that something looks unusual. Causal reasoning tells you why it is happening. In the context of fraud, this means moving beyond “this transaction deviates from the user’s historical pattern” to “the causal chain from user intent to agent action to transaction execution is intact and authorized.”</p><p>This is where the connection to agentic commerce becomes critical. When an AI agent executes a transaction, the question is not whether the transaction matches the user’s behavioral baseline. The question is whether the user authorized this specific action, whether the agent is operating within its defined scope, and whether the instruction chain has been tampered with. This is an intent verification problem, not a pattern recognition problem. It requires understanding the provenance of a decision, not just its statistical profile.</p><p>Causal approaches are also more durable against adversarial probing. A correlational model can be reverse-engineered by observing its outputs. A causal model that evaluates the logic and provenance of an action is structurally harder to game, because the attacker must not only produce the right statistical signature but also construct a valid causal chain, which requires genuine authorization rather than behavioral mimicry.</p><h3>The Converging Perimeter</h3><p>One final observation worth surfacing, though it deserves a deeper treatment in a future post…</p><p>Historically, cybersecurity and fraud prevention have operated as separate disciplines with separate teams, separate tools, and separate reporting lines. Cybersecurity focused on keeping unauthorized actors out of systems. Fraud focused on abuse by actors who got through the front door. The boundary between “breach” and “fraud” was relatively clear.</p><p>Agentic AI dissolves that boundary. When an AI agent is compromised through prompt injection to redirect a payment, is that a cyber attack or a fraud event? When a malicious agent mimics a legitimate financial assistant, is that a software supply chain attack or identity fraud? When an agent is manipulated to authorize a transaction the user never intended, is that social engineering or unauthorized payment fraud? The answer in each case is: both.</p><p>Organizations that continue to treat cybersecurity and fraud as separate functions will have blind spots at exactly the intersection where agentic threats live. The most resilient defense architectures of the next few years will likely unify these functions, recognizing that in an agent-mediated world, system integrity and transaction integrity are the same problem.</p><h3>Conclusion</h3><p>The financial services industry has correctly identified that static identity signals are no longer sufficient in the AI era. Network intelligence, built on cross-institutional financial behavior, is a meaningful step forward and a necessary component of modern fraud defense.</p><p>But it is a starting point. The adversaries are not standing still. They are using the same generative tools to build convincing behavioral footprints, they are probing detection systems to map their decision boundaries, and they are preparing to operate in an agentic landscape where the concept of a human behavioral baseline becomes increasingly irrelevant.</p><p>The next fraud defense will not be built on any single data source or any single network, no matter how broad. It will be built on heterogeneous signal fusion across independent networks, the combination of first-party and third-party intelligence, the shift from correlational pattern matching to causal intent verification, and the continuous evolution of defenses that assume the adversary is learning just as fast as the defender.</p><p>The tools have changed. The attackers have changed. The defense must change with them.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c223d342334e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Simulation Advantage: Why Synthetic Data Is Becoming the Engine of Modern AI]]></title>
            <link>https://medium.com/@statnikov/the-simulation-advantage-why-synthetic-data-is-becoming-the-engine-of-modern-ai-b1fcc6a15635?source=rss-31a9961fd6bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/b1fcc6a15635</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[causality]]></category>
            <category><![CDATA[simulation]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[reinforcement-learning]]></category>
            <dc:creator><![CDATA[Alexander Statnikov]]></dc:creator>
            <pubDate>Sat, 07 Mar 2026 20:36:40 GMT</pubDate>
            <atom:updated>2026-03-07T20:36:40.837Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KLTboAiPCQBJvbqzAquktA.png" /><figcaption>Image credits to Google Nano Banana 2.</figcaption></figure><p>At its core, machine learning is about learning from experience. A system observes examples, identifies patterns, and uses those patterns to make predictions or decisions it has never encountered before. The quality and quantity of that experience, the training data, has always been the binding constraint on what AI can do.</p><p>For most of the field’s history, that experience came from one source: direct observation of the real world. You collect labeled examples (images, transactions, clinical measurements) and feed them to a model. This is effective but expensive. Labeling data requires human expertise. Collecting it at scale requires infrastructure. And in many high-stakes domains, such as medicine, finance, or robotics, the data you need most is precisely the data that is hardest, riskiest, or most expensive to obtain.</p><p>Over the past two decades, two additional mechanisms for acquiring experience have matured. The first is inherited knowledge through <em>pretraining and transfer learning</em>. Large Language Models, vision transformers, and embedding models are trained on massive corpora and then fine-tuned on smaller, domain-specific datasets. The model arrives at your problem already possessing a rich prior understanding of language, structure, or visual patterns. This dramatically reduces the amount of task-specific data required. The second mechanism, and the focus of this post, is <em>simulation</em>. Instead of collecting data from the real world or inheriting it from a pretrained model, you generate it. You build a synthetic environment, run it millions of times under varying conditions, and use the resulting data to train your model. The experience is manufactured, not observed.</p><p>This idea is not new. But something has changed. Simulation has moved from a niche technique used in narrow settings to a first-class training paradigm powering some of the most impressive AI systems in the world. To understand why, and why it matters for the enterprise, we need to confront a few inconvenient truths about what makes simulation hard.</p><h3>The Elephant in the Room: Generative Models Are Hard</h3><p>There is a long-standing debate in machine learning between discriminative and generative models. A discriminative model learns to draw a boundary: given an input, predict the output. A generative model learns something far more ambitious: the full joint distribution of the data, allowing it to produce new samples that look like they came from the real world.</p><p>For decades, the practical acceleration of ML was driven primarily by discriminative methods. Support Vector Machines in the early 2000s, then gradient-boosted trees, then deep neural networks. All achieved breakthrough results by focusing on the comparatively easier task of classification and regression rather than attempting to model the entire data-generating process.</p><p>This matters for simulation because generating realistic synthetic data is, at its heart, a generative problem. You need a model that can produce plausible samples, samples that capture the meaningful structure and variation of the real world. If the simulation is too simple or misspecified, the model trained on it will learn the wrong lessons and fail in production.</p><p>Some skeptics argue that this makes simulation a circular trap: if you could build a generative model good enough to simulate the data, you would not need the discriminative model in the first place. This is theoretically elegant but does not hold in practice. The key insight is that you do not need a perfect generative model. You need one that is good enough to improve a downstream discriminative model in a specific, targeted way. The simulator does not have to reproduce reality with photographic fidelity; it has to produce training signal that fills the gaps the real data cannot cover.</p><p>There is a second, deeper challenge. Good simulation requires more than statistical fidelity. It requires some understanding of the mechanisms that generate the data, specifically the causal structure. To simulate what happens when you change a variable, you need to answer counterfactual questions: what would the outcome be under a different regime? This is a fundamentally harder problem than pattern recognition, and it is why the quality of the simulator’s assumptions matters as much as its statistical output. We will return to this point, but first, let us trace how the field has evolved.</p><h3>The Evolution of Simulation in ML</h3><h4>Causal Graphical Models: The Rigorous Foundation</h4><p>The earliest principled approach to simulation in ML came from causal discovery and Bayesian network research. The idea was to reconstruct the data-generating process as a directed graphical model, a network of variables connected by causal relationships, and then parameterize that network to generate synthetic data under different interventional scenarios.</p><p>This approach had deep intellectual foundations, rooted in the work of Judea Pearl and others. It offered something no other simulation method could: explicit causal controls. You could intervene on a variable, hold others constant, and generate data reflecting a true counterfactual scenario. The applications were remarkable, particularly in biomedicine and precision medicine, where researchers used causal graphical models to discover gene regulatory networks, identify biomarkers, and simulate treatment effects that could not be tested experimentally.</p><p>However, these methods are fragile in practice. They were highly sensitive to assumptions about the data: the structure of the graph, the functional forms of the relationships, the faithfulness conditions. They did not “learn” deeply from data in the way neural networks do; they inferred structure and then parameterized it. Productionizing them at enterprise scale remained a persistent challenge.</p><h4>GANs and VAEs: Learning to Generate</h4><p>The next wave came with the rise of learned generative models. Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) represented a fundamentally different approach. Instead of explicitly modeling the causal structure, these methods learned to generate realistic samples by training neural networks on observed data.</p><p>GANs worked through an adversarial game: a generator network learned to produce synthetic samples, while a discriminator network learned to distinguish real from fake. The generator improved until its output was indistinguishable from reality. VAEs took a probabilistic approach, learning a compressed latent representation of the data and then sampling from it to generate new examples.</p><p>These methods opened up new possibilities, particularly in image generation and data augmentation. But they had well-known limitations: GANs were notoriously unstable to train and prone to mode collapse (generating only a narrow slice of the true data distribution), while VAEs often produced blurry or overly smoothed outputs. More fundamentally, both approaches learned to mimic the surface statistics of data without necessarily capturing the underlying mechanisms. They were powerful interpolators, not causal reasoners. In the broader arc of simulation for ML, they represented an important waypoint: proof that neural networks could learn to generate plausible data at scale, even if the generation lacked causal grounding.</p><h4>Reinforcement Learning: The Breakout Success</h4><p>The most dramatic demonstration of simulation’s power came from reinforcement learning. RL is, at its core, a simulation-driven paradigm. An agent interacts with an environment, takes actions, observes outcomes, and learns a policy that maximizes long-term reward. The environment is the simulator, and the agent’s training data is generated entirely through interaction with it.</p><p>The results have been extraordinary. DeepMind’s AlphaGo Zero learned to play Go, one of the most complex strategic games ever devised, entirely through self-play, without any human game data. The system started with nothing but the rules of the game and a neural network. It played millions of games against itself, generating its own training data through simulation. Within three days, it defeated the previous version of AlphaGo. Within 40 days, it surpassed every known Go player, human or machine.</p><p>AlphaZero generalized this approach to chess and shogi, achieving superhuman performance in all three games within hours of training, again entirely through self-play simulation. The training data was not collected; it was manufactured by the system’s own interaction with a simulated environment.</p><p>The principle extended far beyond board games. In robotics, the sim-to-real paradigm has become a cornerstone of modern development. Training a physical robot through trial and error in the real world is slow, expensive, and dangerous. Instead, researchers train agents in simulated physics environments, running millions of episodes in parallel, and then transfer the learned policies to physical hardware. Tesla’s Optimus humanoid robot, for example, learned its remarkably human-like walking gait entirely through reinforcement learning in simulation before the policy was deployed to the physical robot in a zero-shot transfer. Nvidia’s research on using video-generation models as “neural physics engines” to create simulated worlds for robot training represents the latest frontier of this approach.</p><p>In each case, the pattern is the same: simulation provides effectively unlimited training data at near-zero marginal cost, enabling the agent to explore vast spaces of possibility that would be infeasible to cover with real-world experience alone.</p><h3>A Critical Distinction: Simulating Artifacts vs. Simulating Processes</h3><p>Before we turn to the modern frontier, it is worth drawing a distinction that is easy to miss in the current AI moment.</p><p>Large Language Models and diffusion models are, technically, generative models. They learn distributions and produce samples (text, images, video) that are often indistinguishable from human-created content. In that narrow sense, they are simulators. But they simulate <em>artifacts</em>: the outputs of processes.</p><p>The simulation paradigm we are discussing here is different. It is about simulating <em>processes and systems</em>: the dynamics, mechanisms, and causal structures that generate observable data. An LLM can produce a realistic-looking financial transaction log. A process simulator models the mechanism by which a borrower’s job loss changes their spending pattern, which affects their utilization, which affects their default probability. One produces plausible-looking data. The other models the dynamics that generate the data.</p><p>That said, there is a fascinating convergence happening. LLMs and diffusion models are increasingly being used as components inside process simulators. For example, they can generate realistic user intents that feed into agentic workflow simulations, or create visual environments for robotics training. World models, which aim to learn how systems evolve over time rather than just what they look like at a single moment, represent the explicit attempt to bridge this gap. The line between artifact generation and process simulation is blurring, and some of the most exciting work sits precisely at this boundary.</p><h3>The Modern Frontier: Simulation as a First-Class Training Strategy</h3><p>What is happening right now is a qualitative shift. Simulation is no longer a supplement to real data or a workaround for data scarcity. It is becoming the primary training strategy for state-of-the-art foundation models in domains that have nothing to do with games or robotics.</p><p>Consider Amazon’s Chronos-2, a foundation model for time series forecasting released in late 2025. Chronos-2 achieves state-of-the-art performance across multiple comprehensive benchmarks, and its most remarkable capabilities rely almost entirely on synthetic data. The model’s ability to handle multivariate forecasting and incorporate covariates was enabled by generating synthetic time series data with imposed multivariate structures. The researchers trained a version of Chronos-2 using only synthetic data and found that it performed only slightly below the version trained on a mix of real and synthetic data. The paper’s own conclusion is striking: with further research, real data may not even be required for effective pretraining of time series models.</p><p>A similar pattern is emerging in relational data. KumoRFM, the first relational foundation model for enterprise data, was trained solely on synthetic enterprise-like data. It learns to make predictions like churn, fraud detection, and recommendations directly from structured relational databases, without task-specific training. The fact that it can generalize to unseen database schemas and tasks it was never explicitly trained for, all while being trained on synthetic data, challenges the assumption that real-world data is always necessary for building enterprise AI.</p><p>These are not isolated cases. They represent a broader pattern: as models become more architecturally sophisticated, incorporating attention mechanisms, in-context learning, and graph-based representations, they become better at extracting structural knowledge from synthetic data. The simulator does not need to perfectly replicate reality. It needs to expose the model to the right diversity of structures, relationships, and edge cases. The model’s architecture does the rest.</p><h3>The Causality Thread: What Separates Good Simulation from Noise</h3><p>Running through all of this is a quiet but essential thread: causality. The quality of a simulator is ultimately determined by how well it captures the causal structure of the domain it is modeling.</p><p>In reinforcement learning, the environment itself encodes causality. The rules of Go, the physics of a simulated warehouse, the dynamics of a financial market. The agent learns causal relationships implicitly through interaction: action A in state S leads to outcome O.</p><p>In generative approaches like GANs or diffusion models, causality is largely absent. The model learns statistical patterns without understanding the mechanisms behind them. This is why synthetic data generated by purely statistical methods can introduce subtle biases or fail to generalize to out-of-distribution scenarios.</p><p>The most promising modern approaches sit between these extremes. Chronos-2’s synthetic data generation imposes structural relationships like multivariate dependencies and covariate effects that encode a form of lightweight causal structure. KumoRFM’s training data reflects the relational and temporal patterns of real enterprise databases, even though the specific records are synthetic.</p><p>This is the practical frontier: simulation frameworks that encode enough causal structure to produce meaningful training signal, without requiring the full rigor of a formal causal graphical model. The generative versus discriminative debate has not been resolved in theory, but it has been largely bypassed in practice by this pragmatic middle ground.</p><h3>Why This Matters for the Enterprise</h3><p>For organizations that rely on models to predict and influence customer behavior in financial services, e-commerce, healthcare, and insurance, the simulation paradigm represents an enormous and largely untapped opportunity.</p><p>These industries face a structural data problem. The most valuable training data is often the data you cannot collect: outcomes for customers you rejected, behavior under credit limits you never assigned, responses to prices you never charged, results of treatments you never administered. These are counterfactual scenarios, and they are precisely where simulation can fill the gap.</p><p>Moreover, real-world experimentation in these domains is expensive and sometimes impossible. You cannot randomly approve risky loans to build a training set. You cannot randomly withhold medical treatment to create a control group. Simulation offers a path to generating the training data you need without the cost, risk, or ethical constraints of real-world experimentation.</p><p>The enterprises that invest in simulation infrastructure, not just generative AI for content creation but structured simulators for their core business processes, will build a compounding advantage. Every simulation run generates training data. Every model trained on that data makes better decisions. Every better decision generates better real-world outcomes that can be used to improve the simulator. It is a flywheel, and the organizations that spin it first will be difficult to catch.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4caDV5PuPR8XxlgwBvPojw.png" /></figure><h3>Conclusion</h3><p>We are witnessing a paradigm shift in how AI systems acquire experience. Direct observation of the real world, once the only viable path, is now one of three, alongside inherited knowledge from pretrained models and manufactured experience from simulation.</p><p>The generative versus discriminative tension that once made simulation seem impractical has not disappeared, but it is no longer a blocker. Modern systems have demonstrated that imperfect simulators, ones that capture enough structural diversity without modeling every mechanism perfectly, can produce training data sufficient to match or exceed models trained on real-world data alone.</p><p>But we should not lose sight of what separates the best simulation from noise generation: causal grounding. The simulators that encode meaningful structure, the relationships between variables, the dynamics of systems, the logic of interventions, produce models that generalize. The ones that merely reproduce surface statistics produce models that memorize.</p><p>The tools have changed, but the fundamental challenge remains: to build systems that learn not just from what has happened, but from what could happen. Simulation is how we manufacture that experience. And for the industries that depend on understanding cause and effect to drive growth and manage risk, this may be the most important infrastructure investment of the next decade.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b1fcc6a15635" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Prediction Trap: Why Financial Services Keeps Solving the Wrong Problem]]></title>
            <link>https://medium.com/@statnikov/the-prediction-trap-why-financial-services-keeps-solving-the-wrong-problem-33d9f6540ce3?source=rss-31a9961fd6bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/33d9f6540ce3</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[causal-inference]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Alexander Statnikov]]></dc:creator>
            <pubDate>Sat, 28 Feb 2026 23:51:51 GMT</pubDate>
            <atom:updated>2026-02-28T23:51:51.842Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZPIaASwQAVUIEKDKJD9SXw.png" /><figcaption>Image credits to Google Nano Banana 2.</figcaption></figure><p>We have spent the last decade building increasingly sophisticated prediction machines for financial services. The models have gotten larger, the feature sets expanded, the infrastructure more robust. And yet, the most consequential decisions in the industry — who to approve, how much credit to extend, when to intervene, how to price — are not prediction problems at all. They are causal problems. And the gap between what we predict and what we actually need to know is where billions in value silently leak out.</p><p>The fundamental issue is deceptively simple. Prediction answers “what will happen?” Causal inference answers “what would happen <em>if</em>?” In a world where every decision is an intervention — approving a loan changes the borrower’s trajectory, extending a credit line changes spending behavior, sending a collections notice changes repayment likelihood — prediction alone is structurally insufficient. We are applying Rung 1 tools from <a href="https://medium.com/causal-inference/the-ladder-of-causation-climbing-up-in-the-world-of-causal-inference-2-15-7539f92c280d">Judea Pearl’s Ladder of Causation</a> to problems that live firmly on Rung 2 and Rung 3.</p><p>This is not a new observation. Researchers have been saying it for decades. What <em>is</em> new is that the modern ML toolkit has finally caught up to the theory. Libraries like <a href="https://www.microsoft.com/en-us/research/project/econml/">EconML</a>, <a href="https://arxiv.org/abs/2011.04216">DoWhy</a>, and <a href="https://causalml.readthedocs.io/en/latest/about.html">CausalML</a> have industrialized methods that were once confined to papers and custom scripts. <a href="https://arxiv.org/html/2406.00853v1">Doubly robust estimators</a>, <a href="https://arxiv.org/abs/1902.07409">causal forests</a>, and <a href="https://medium.com/@med.hmamouch99/double-machine-learning-for-causal-inference-a-practical-guide-5d85b77aa586">Double ML</a> are no longer exotic — they are production-ready. And with the rise of agentic AI systems that make autonomous decisions at machine speed, the cost of ignoring causality has shifted from “suboptimal” to “dangerous.”</p><p>Let me walk through five problems that every financial services professional will recognize. Each one looks like a prediction problem on the surface. Each one is, at its core, a causal inference problem. And each one illustrates a distinct methodological dimension that the modern toolkit can address.</p><h3>The Approval Paradox: Selection Bias in Underwriting</h3><p>Consider the most classical problem in consumer lending: credit underwriting. You want to build a model that predicts default. Naturally, you construct a binary dependent variable — 1 for defaulted, 0 for did not default — and train your model on historical outcomes.</p><p>The problem is immediately obvious once you see it, yet routinely ignored: you only observe outcomes for borrowers you <em>approved</em>. The rejected applicants — the ones your previous model or policy turned away — are invisible. You have no idea what would have happened if you had said yes.</p><p>This is not a data quality issue. It is a structural counterfactual problem. You are training a model on a biased sample that was filtered by a prior decision, and then using that model to make the <em>next</em> decision. The training data is a product of the very policy you are trying to improve. In Pearl’s framework, this is a Rung 3 problem — “what <em>would have</em> happened?” — hiding inside what looks like a Rung 1 classification task.</p><p>The classical remedy is <a href="https://www.experian.com/blogs/insights/reject-inference/">reject inference</a>: building a secondary model on the entire applicant population, using external signals such as credit bureau performance data to infer likely outcomes for the rejected group. When bureau data is available, this approach is well-trodden. But in contexts where bureau data is thin or unavailable — think underbanked populations, new-to-credit consumers, or markets outside the U.S. — the problem becomes significantly harder. This is precisely where modern causal methods such as <a href="https://www.bmj.com/content/352/bmj.i189">inverse probability weighting</a>, <a href="https://arxiv.org/html/2406.00853v1">doubly robust estimators</a>, and <a href="https://link.springer.com/article/10.1007/s40471-022-00308-6">sensitivity analyses for unmeasured confounding</a> become essential rather than optional.</p><p>Every lending institution that ignores this selection bias is systematically underestimating its true addressable market and overestimating the riskiness of its rejected population. The prediction model says it is performing well. The evaluation metrics look clean. But the model is confident about a world it has never actually seen.</p><h3>The Dosage Problem: Credit Line Assignment</h3><p>Now take a borrower you have already approved. You need to assign a credit limit — for a credit card, a personal loan, a line of credit. In your training data, you observe what happened at <em>one specific limit</em> for each customer: their spending patterns/utilization rates, and default/repayment behavior.</p><p>But the decision you actually need to make is interventional: what would happen if you increased this customer’s limit by $2,000? What if you decreased it? You have exactly one observation per customer at exactly one treatment level. This is the fundamental problem of causal inference in its most textbook form — estimating the effect of a continuous treatment on an outcome, with only observational data.</p><p>Randomized experimentation is the gold standard solution, but it is enormously costly in financial services. Randomly assigning high credit limits to potentially risky borrowers, or low limits to valuable customers, has direct P&amp;L consequences and can trigger regulatory scrutiny.</p><p>This is where the modern causal ML toolkit shines. <a href="https://medium.com/@med.hmamouch99/double-machine-learning-for-causal-inference-a-practical-guide-5d85b77aa586">Double Machine Learning</a> was designed precisely for estimating treatment effects in high-dimensional settings — which is exactly what credit line optimization is. <a href="https://arxiv.org/abs/1902.07409">Causal forests</a> allow you to estimate <em>individualized</em> treatment effects: for this specific customer, with this specific profile, what is the marginal impact of an additional $1,000 in credit on expected value? <a href="https://medium.com/@mandeep0405/causal-inference-intro-to-meta-learners-1e35383a19d2">Meta-learners</a> — the T-learner, S-learner, X-learner, and R-learner family — offer flexible frameworks for estimating heterogeneous effects without imposing strong parametric assumptions.</p><p>The institution that treats credit line assignment as a prediction problem is leaving money on the table. The one that treats it as a causal problem can optimize the entire portfolio along the treatment frontier.</p><h3>The Attribution Mirage: Marketing and Customer Acquisition</h3><p>Every fintech and bank runs multi-channel acquisition campaigns: paid search, direct mail, referral bonuses, app install ads, influencer partnerships. The standard analytical approach is attribution modeling — last-touch, first-touch, or multi-touch — which attempts to assign credit for a conversion to the channels that preceded it. This is purely associational. It answers “which channels were <em>present</em> before conversion?” not “which channels <em>caused</em> conversion?”</p><p>The distinction is crucial. It directly determines how hundreds of millions of dollars in marketing spend are allocated. If your attribution model tells you paid search is driving 40% of conversions, but half of those customers would have converted organically, you are systematically overspending on that channel and underspending on channels with genuine incremental lift.</p><p>The causal framing here is <a href="https://en.wikipedia.org/wiki/Uplift_modelling">uplift modeling</a>, also called incremental response modeling. Instead of predicting who will convert, you are predicting who will convert <em>because of</em> the treatment. This requires a fundamentally different model architecture and a different evaluation framework — you are optimizing for the conditional average treatment effect, not for classification accuracy.</p><p>At the market level, <a href="https://en.wikipedia.org/wiki/Synthetic_control_method">synthetic control methods</a> offer another powerful tool. By holding out entire geographic markets as controls and constructing a synthetic “counterfactual market” from weighted combinations of control regions, you can estimate the true causal impact of a campaign at scale without needing user-level randomization. Geo-based causal inference has become increasingly sophisticated, and for large-scale brand and awareness campaigns where individual attribution is impossible, it is often the only credible approach.</p><p>The broader point is this: every dollar spent based on correlational attribution is a dollar allocated on faith rather than evidence. The prediction model says the campaign is working. The causal model tells you whether it actually is.</p><h3>The Timing Trap: Collections Strategy and Dynamic Interventions</h3><p>When a borrower starts falling behind on payments, the institution faces a sequence of decisions: when to send a reminder, when to offer a payment plan, when to escalate to a hardship program, when to involve external collections. Each of these is an intervention, and each intervention changes the borrower’s state in ways that affect the optimal <em>next</em> intervention.</p><p>The conventional approach is to build a propensity-to-pay model, segment borrowers into risk tiers, and apply rule-based strategies. This treats the problem as static: predict the probability of repayment, then act. But the reality is dynamic. Calling a borrower at day 15 past due changes whether they pay at day 30, which changes whether you should offer a restructuring plan at day 60. The intervention itself is a confounder for future outcomes.</p><p>This is the problem of time-varying confounding, and it is one of the most challenging areas in causal inference. Standard regression adjustment fails here because conditioning on intermediate outcomes (like whether the borrower responded to the first call) introduces collider bias. The methodological toolkit for this includes <a href="https://en.wikipedia.org/wiki/Marginal_structural_model">marginal structural models</a>, <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC6074945/">G-estimation</a>, and more recently,<a href="https://causalai.net/r48.pdf"> approaches that combine reinforcement learning with causal frameworks to learn optimal dynamic treatment regimes from observational data</a>.</p><p>What makes this problem particularly compelling in the current moment is that collections is one of the first areas where agentic AI is being deployed operationally. Agents are already making real-time decisions about when and how to contact borrowers. If those agents are optimizing on naive predictive models that ignore the causal dynamics of sequential intervention, they will confidently scale suboptimal strategies at machine speed — calling too early, escalating too aggressively, or missing the narrow window where a well-timed offer could have recovered the relationship.</p><h3>The Elasticity Illusion: Pricing and Fee Sensitivity</h3><p>Banks and fintechs constantly adjust prices: interest rate spreads, subscription fees, interchange tiers, late fees, foreign transaction charges. The standard analytical approach is to estimate price elasticity from historical data: when we raised the monthly fee from $5 to $7, attrition increased by X%.</p><p>The problem is <a href="https://en.wikipedia.org/wiki/Endogeneity_(econometrics)">endogeneity</a>. Pricing changes do not happen in a vacuum. They are typically triggered by competitive pressure, regulatory shifts, or portfolio performance — all of which independently affect customer behavior. The observed correlation between a price change and an outcome change is contaminated by these shared drivers. You are not measuring the causal effect of price on behavior; you are measuring a tangled mess of simultaneous forces.</p><p>This is a textbook setting for <a href="https://en.wikipedia.org/wiki/Instrumental_variable">instrumental variable approaches</a>: finding an exogenous shock that shifts price but affects the outcome only through price. Regulatory changes — such as a cap on interchange fees or a new disclosure requirement — can sometimes serve as natural instruments. Threshold-based policies create opportunities for regression discontinuity designs, where customers just above and just below a policy cutoff provide a quasi-experimental comparison.</p><p>The practical consequence of ignoring endogeneity in pricing is that institutions systematically misestimate their own price sensitivity. They may believe demand is inelastic when it is actually elastic (because the price increase coincided with a favorable macro environment), or vice versa. These biased estimates flow directly into pricing strategy, revenue forecasts, and competitive positioning. The institution thinks it knows its customers’ willingness to pay. It does not. It knows a biased correlation that happens to look like an answer.</p><h3>The Compounding Error and the Agentic Imperative</h3><p><strong>These five problems are not isolated</strong>. They are deeply interconnected. A biased underwriting model produces a biased training dataset for credit line assignment. Biased credit line models generate distorted customer lifetime value estimates that feed into marketing attribution. Flawed attribution drives misallocated spend, which changes the composition of the customer portfolio, which distorts collections outcomes and pricing analyses downstream. Causality is not a per-model issue. It is a systemic integrity issue. The errors compound through the decisioning chain, and each link amplifies the bias of the one before it.</p><p>This compounding effect is exactly why the rise of agentic AI makes causal literacy an existential priority rather than a technical nicety. When a human analyst runs a biased model, the damage is bounded by judgment, review cycles, and organizational inertia. When an autonomous agent uses that same model to make thousands of decisions per minute — approving loans, setting limits, triggering collections actions, adjusting prices — the causal blind spot scales with the system.</p><p>The good news is that the gap between theory and practice has never been smaller. The causal inference methods I have described are not academic curiosities. They are implemented in production-grade libraries, they can be integrated into modern ML pipelines and feature stores, and they can be invoked as tools within agentic architectures. The “<a href="https://medium.com/@statnikov/the-causal-chasm-why-agents-not-just-models-may-climb-the-ladder-16c7025009f9">Reasoning Router</a>” concept, where an LLM orchestrator knows when to reach for a causal estimator instead of a predictive model is the logical next step in the evolution of enterprise AI systems.</p><p>The harder gap is organizational. Causal methods require a different data collection strategy: holdout groups, randomization budgets, instrumentation for natural experiments. They require a different evaluation framework — you cannot grade a causal model on AUC alone. And they require a different conversation between data science teams and business stakeholders, one where the question shifts from “how accurate is the model?” to “what is the model actually estimating, and is that the right quantity for this decision?”</p><p>The prediction machines we have built are extraordinary. But prediction is a tool, not a destination. <strong>The decisions that drive enterprise value in financial services are interventions, and interventions require causal reasoning</strong>. We have the theory. We now have the tools. The remaining barrier is the willingness to reframe the question.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=33d9f6540ce3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond Static Tests: Why Agentic AI Requires Dynamic Evaluation]]></title>
            <link>https://medium.com/@statnikov/beyond-static-tests-why-agentic-ai-requires-dynamic-evaluation-bcfaed828bea?source=rss-31a9961fd6bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/bcfaed828bea</guid>
            <category><![CDATA[llmops]]></category>
            <category><![CDATA[llm-evaluation]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Alexander Statnikov]]></dc:creator>
            <pubDate>Sun, 22 Feb 2026 01:13:42 GMT</pubDate>
            <atom:updated>2026-02-22T01:13:42.817Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Wf0teedVzdQoa2HOEiIx-w.png" /><figcaption>Images credits to Google Nano Banana Pro</figcaption></figure><p>We have spent decades optimizing how algorithms learn. But as we enter the era of autonomous agents, the more pressing challenge is accurately measuring what these systems actually “know.”</p><p>In Machine Learning, success hinges on generalization: the ability of a system to learn from historical data and apply those lessons accurately to unseen scenarios. A model that cannot generalize is not learning; it is simply functioning as an expensive database.</p><p>When deploying AI in high-stakes practical environments, generalization often fails in two distinct but related ways: overfitting and distribution shift. To understand why autonomous agents break traditional evaluation frameworks, we must first examine the mechanics of these twin failures.</p><h3>The Two Faces of Overfitting</h3><p>At their core, overfitting and distribution shift represent a divergence between the environment an AI expects and the one it encounters.</p><p>Overfitting occurs when a model fixates on historical data rather than underlying patterns. This typically happens in two distinct ways. The first is <strong>training data memorization</strong>. Instead of learning the principles of credit risk, a neural network might memorize quirks in its training set — for example, noticing that loan applicants from a specific zip code with seven-letter last names never defaulted in a given year. It learns noise, not underlying correlations. Common mitigations for this include regularization, early stopping, cross-validation, and feeding the model a significantly larger, more diverse dataset.</p><p>The second type of overfitting is more insidious: <strong>adaptive overfitting</strong> (sometimes referred to as validation overfitting).</p><p>Algorithms rely on “hyperparameters” — the structural settings of the model. To find the optimal configuration, engineers test the algorithm iteratively against a validation dataset. Imagine building an algorithmic trading model. You test your model against Q3 market data, and it loses money. You adjust the risk thresholds, momentum dials, and volatility caps, re-testing hundreds of times against that exact same historical dataset. Eventually, you find a configuration that yields a brilliant return.</p><p>You haven’t built a smarter trading model. You have simply rolled the dice enough times to find a configuration that, by chance, perfectly navigated the unrepeatable market noise of Q3. You overfit by exploiting the test itself.</p><p><em>(Technically, this is a failure of statistical evaluation — treating point estimates as absolute truth without accounting for confidence intervals across multiple testing cycles. However, the practical outcome is indistinguishable from overfitting: the model learns the noise of the test set rather than the underlying signal.)</em></p><h3>The Threat of Distribution Shift</h3><p>Distribution shift is a conceptually adjacent problem: the model performs exceptionally well on its target distribution, but the environment mutates.</p><p>Consider an AI underwriting system trained over a decade of low interest rates. It learns robust, generalized patterns about consumer repayment behaviors. Then, inflation spikes and interest rates double. The underlying economic reality shifts (concept drift). In the agentic world, shift isn’t just macroeconomic; it can be as granular as an underlying API changing its schema or a downstream tool altering its output format. The patterns the model learned are mathematically sound but no longer applicable to the new reality.</p><p>As we transition from prediction models to autonomous agents, the methods we use to monitor these divergences must fundamentally change.</p><h3>The Evolution of Evaluations</h3><p>Historically, we evaluated AI using ground-truth benchmarks and task-specific metrics (e.g., Precision, Recall, AUC, RMSE). When Large Language Models (LLMs) emerged, rigid math could no longer grade open-ended text. We shifted to LLM-as-a-judge frameworks to score outputs on coherence and relevance. However, these judges are famously stochastic, biased toward verbosity, and easily exploitable unless heavily calibrated with strict rubrics and pairwise rankings.</p><p>Today, we are moving from static models to systems with tools. An agent executes multi-step workflows: it ingests a query, writes a SQL script, retrieves proprietary data, runs a tool, and summarizes the findings. Evaluating an agent no longer means just checking a final answer; it means evaluating a <strong>trajectory</strong> — the sequence of observable actions, API calls, and intermediate artifacts.</p><p>As agents become more autonomous, it is imperative to have an independent oversight layer to audit these execution paths.</p><h3>Agentic Evals: The New Product Analytics</h3><p>This brings us to a critical technological convergence: Agentic Evaluations are becoming the evolution of Product Analytics.</p><p>A decade ago, product managers used platforms like Mixpanel or Amplitude to track the “user funnel.” They mapped the human journey to identify friction: <em>Did the user click the button? Did they drop off before checkout?</em> In the agentic era, the “user” navigating the software is the AI itself. Modern AI tools must track the “reasoning funnel.” <em>Did the agent correctly interpret the intent? Did it select the right tool? Did the database query return the correct context? Did the agent synthesize that data correctly, or did it hallucinate at the final step?</em> We are measuring how software navigates software. This requires an entirely new discipline of agentic analytics to understand exactly where autonomous workflows break down.</p><h3>The Danger of Static Golden Datasets</h3><p>To manage this complexity, we often rely on production observability. Modern instrumentation platforms monitor live, multi-step traces, allowing us to detect anomalies like tool error rates or latency spikes when real-world data drifts.</p><p>However, observability is not evaluation. Observability captures the trace; evaluation scores that trace against a rubric. And how we score those traces is exacerbating our overfitting problem.</p><p>To ensure safety, enterprise teams build “Golden Datasets” — fixed sets of high-quality test cases. Golden sets are absolutely essential for baseline regression testing, but relying on them exclusively triggers adaptive overfitting. If an engineering team runs an agent against the exact same static 500 test questions a thousand times while tweaking the system prompt, the team is actively specializing the agent to that exact wording.</p><p>The agent doesn’t literally “memorize” the data; rather, the iterative optimization loop strips away the model’s robustness to paraphrasing and edge cases. This creates an illusion of competence. The agent looks flawless on the evaluation dashboard, but the moment it hits production and encounters a query phrased slightly differently, the logic crumbles.</p><h3>The Imperative for Dynamic Evaluation</h3><p>We can no longer rely purely on static Golden Datasets. Evaluation must become a programmatic simulation. A robust defense requires golden sets combined with held-out sets, rotating suites, and continuous online monitoring.</p><p>The frontier of AI engineering requires systems that use generative models to constantly spin up novel, synthetically generated test cases. To avoid “synthetic drift” — where the generated tests diverge from reality — these dynamic suites must be heavily grounded, seeded from de-identified production logs and constrained by strict templates.</p><p>In any iterative optimization loop — whether automated or driven by a human engineer chasing a dashboard metric — systems naturally drift toward the metric they are measured against. Therefore, evaluations must operate like a cybersecurity red team: continuously shifting, adapting, and challenging the agent in novel ways.</p><p>If we rely purely on static tests to grade dynamic agents, we aren’t measuring capability; we are measuring narrow compliance. To build systems that can be trusted with critical enterprise workflows, our evaluations must evolve just as fast as the agents they govern.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bcfaed828bea" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Causal Chasm: Why Agents, Not Just Models, May Climb the Ladder]]></title>
            <link>https://medium.com/@statnikov/the-causal-chasm-why-agents-not-just-models-may-climb-the-ladder-16c7025009f9?source=rss-31a9961fd6bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/16c7025009f9</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[causal-inference]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[reasoning]]></category>
            <dc:creator><![CDATA[Alexander Statnikov]]></dc:creator>
            <pubDate>Sat, 14 Feb 2026 19:49:05 GMT</pubDate>
            <atom:updated>2026-02-14T19:49:05.498Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WfaLTdhqgZcAZx47s1WOrw.png" /><figcaption>Images credits to Google Nano Banana Pro</figcaption></figure><p>We have spent the last three decades perfecting the art of prediction. In the world of machine learning, we built engines that could look at <em>X</em> and predict <em>Y</em> with superb accuracy. But as any first-year statistics student knows, prediction is not explanation, and correlation is not causation.</p><p>For years, the “foundational” view of AI research has been grounded in Judea Pearl’s Ladder of Causation. As outlined in <a href="https://www.amazon.com/Book-Why-Science-Cause-Effect/dp/1541698967"><em>The Book of Why</em></a>, intelligence operates on three distinct rungs:</p><ol><li><strong>Association (Seeing):</strong> Identifying patterns. “If I see a customer buying diapers, they are likely to buy beer.” This is where 99% of modern Machine Learning lives.</li><li><strong>Intervention (Doing):</strong> Predicting the effect of an action. “If I <em>double</em> the price of the beer, will sales drop?” This requires understanding the mechanics of the system, not just the observation of it.</li><li><strong>Counterfactuals (Imagining):</strong> Reasoning about alternative realities. “What <em>would</em> have happened if we hadn’t doubled the price?”</li></ol><p>For a long time, the field of <strong>Causal AI</strong> developed slowly but firmly alongside the predictive boom. Researchers focused on theoretically sound algorithms that could discover causal structures from observational data — allowing us to make inferences about the future without running costly, sometimes impossible, experiments. These methods were rigorous but fragile, often highly sensitive to the nuances of the data they were fed.</p><p>Then, the Large Language Models (LLMs) entered the scene, and the conversation shifted overnight.</p><h3>The Illusion of Reasoning</h3><p>LLMs are, by definition, predictive engines. They are trained to predict the next token, a task firmly rooted in Rung 1 (Association). Yet, because they have ingested a near-infinite corpus of human knowledge — which itself contains causal explanations — they often <em>appear</em> to reason.</p><p>This has created a dangerous ambiguity. Are they simulating reasoning, or actually doing it?</p><p>A recent survey titled <a href="https://arxiv.org/abs/2602.06176"><em>Large Language Model Reasoning Failures</em></a> (Song et al., Feb 2026) throws cold water on the idea that scale alone solves this problem. The authors categorize reasoning failures into fundamental, application-specific, and robustness issues, noting that even the most advanced models stumble on simple logical hurdles when the phrasing changes slightly.</p><p>Skeptics like Gary Marcus <a href="https://garymarcus.substack.com/p/breaking-llm-reasoning-continues">have amplified these findings</a>, arguing that what looks like reasoning is often just sophisticated approximate retrieval. The criticism is that LLMs lack an internal “world model” — they don’t understand the physics of the data, only the statistics of the language describing it. They are shallow circuits trying to solve deep logic problems; without a fundamental architectural shift, they cannot reliably climb to Rung 2 or 3. They can mimic the <em>output</em> of a causal argument, but they cannot derive the <em>cause</em> from the raw signal.</p><h3>The Pragmatic Pivot: From Models to Systems</h3><p>If we stop our analysis at the model layer, we miss the actual revolution currently unfolding in enterprise AI. Most academic skeptics are effectively analyzing a “brain in a jar” — measuring the raw synaptic firing of a standalone model in a sterile environment. Meanwhile, the “builders” are giving that brain hands, eyes, and access to a library of specialized tools.</p><p>We are witnessing a fundamental shift from <strong>LLM-centric</strong> thinking to <strong>Agentic Systems</strong>.</p><p>In an agentic architecture, the LLM is no longer the sole source of intelligence; it is the orchestrator. Consider the sophisticated workflows now becoming commonplace in high-stakes environments like finance or engineering. Today, an agentic system can ingest a complex prompt, retrieve institutional data from proprietary feeds, invoke a deterministic simulation engine to run thousands of sensitivity analyses, and then synthesize the results into a multi-tabbed strategic model — all in a matter of minutes.</p><p>Does the LLM “understand” the nuances of debt covenants or the mathematical proof of an amortization schedule in a philosophical sense? Almost certainly not. But the System — the combination of the LLM’s linguistic interface and the deterministic tools it commands — effectively does.</p><p>This is the concept of the <strong>Reasoning Router</strong>. By offloading the “heavy lifting” of logic to deterministic engines (like Python scripts, SQL queries, or symbolic math solvers) and using the LLM merely to translate intent and verify outputs, we bypass the model’s inherent tendency toward hallucination. The LLM doesn’t need to <em>be</em> a perfect calculator if it is a perfect <em>user</em> of a calculator.</p><p>However, we must be cautious: this does not mean we have solved the “Causal Chasm.” We are currently advancing towards <strong>Functional Reasoning</strong>, not <strong>Theoretical Causality</strong>. The system can follow the <em>procedure</em> of a causal argument, and it can even be programmed to call a causal discovery algorithm to analyze a dataset. But it is still operating on a “procedural” level — it is simulating the steps of a reasoner without necessarily possessing the “world model” required to guarantee truth in a novel, unseen scenario.</p><p>This construct challenges the scientists who narrowly evaluate AI based on token-prediction benchmarks. If a system can query a database, detect a structural break in the data, run a causal inference test, and format a valid business recommendation, it becomes functionally indistinguishable from a reasoning entity for many knowledge work applications. We may not have reached the top of the Ladder of Causation yet, but we have built a very effective elevator that stops at several of the higher floors.</p><h3>The Ensemble Future: Innovating Beyond the LLM</h3><p>As we look toward the horizon, it is becoming increasingly clear that the path to true intelligence likely won’t be found by simply making existing models “larger.” Instead, we are entering a dual-track era of innovation: the development of entirely new model families and the rise of the <strong>Intelligent Ensemble</strong>.</p><h4>1. New Horizons: Specialized Model Families</h4><p>While text-based models dominate the current landscape, several alternative approaches are being pursued that prioritize “understanding” over “prediction.” These represent early-stage work, and while their long-term impact on the enterprise is not yet fully known, they offer a glimpse into a more grounded AI:</p><ul><li><strong>Spatial and Physical Intelligence:</strong> A new class of “World Models” is emerging that moves beyond language to understand the physics of 3D space. These systems are designed to simulate reality, allowing an AI to “imagine” the consequences of an action in a physical or geometric context before executing it.</li><li><strong>Neuro-Symbolic Integration:</strong> This approach attempts to marry the pattern-recognition strengths of neural networks with the rigorous, rule-based logic of symbolic AI. By embedding “hard” logical constraints into “soft” learning models, researchers hope to create systems that are not just more accurate, but more explainable and auditable in high-stakes environments.</li><li><strong>Joint Embedding Predictive Architectures (JEPA):</strong> Unlike models that predict the next word or pixel, these architectures focus on predicting high-level, abstract representations of the world. By ignoring the “noise” and focusing on the underlying structure, they aim to learn how different components of a system interact — a critical prerequisite for intervention-based reasoning.</li></ul><h4>2. The Agentic Ensemble: Practical Reasoning through Collaboration</h4><p>The second, and perhaps more immediate, breakthrough is the shift from monolithic models to <strong>Agentic Ensembles</strong>.</p><p>The most powerful systems of 2026+ likely won’t be a single, all-knowing brain. Instead, they will be collaborative ecosystems. Imagine an agentic coordinator that delegates tasks to a “cabinet” of specialists. One agent might be an LLM optimized for intent; another might be a specialized world model for physical simulation; a third might be a classical causal discovery algorithm running on live telemetry.</p><p>This is where “Practical Causality” wins. While it remains mathematically difficult to prove the formal causal properties of a massive, multi-agent system, the utility of such an ensemble is really promising. By combining the linguistic fluency of an LLM with the grounded logic of specialized models, we can build a system that can simulate “what-if” scenarios with a level of accuracy that a standalone model could never achieve.</p><h3>Conclusion: The Mission Remains the Same</h3><p>We are in the midst of a transition from AI as a “chatbot” to AI as Infrastructure.</p><p>The journey toward AGI isn’t just about climbing the Ladder of Causation in a textbook; it’s about building the elevators that allow our organizations to operate at the top floors every day. We must remain cautiously optimistic — acknowledging the “early work” status of these new families while building the agentic systems that can leverage their strengths.</p><p>The tools have changed, and the “physics” of our models has evolved, but the fundamental mission remains: to design systems that don’t just predict the future, but help us understand — and ultimately shape — the causes behind it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=16c7025009f9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Risk in 2026: What Changes When AI Enters the Loop]]></title>
            <link>https://medium.com/@statnikov/risk-in-2026-what-changes-when-ai-enters-the-loop-6d6ff464610c?source=rss-31a9961fd6bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/6d6ff464610c</guid>
            <category><![CDATA[risk-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[financial-services]]></category>
            <dc:creator><![CDATA[Alexander Statnikov]]></dc:creator>
            <pubDate>Sun, 08 Feb 2026 02:22:23 GMT</pubDate>
            <atom:updated>2026-02-08T02:22:23.481Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xjmAj9yQBHlzee0mtaajuQ.png" /><figcaption>Image credits to Google Nano Banana Pro.</figcaption></figure><p>There is a lot of noise right now about what AI will do to financial services. But if you cut through the hype and look at the daily reality of risk management, the shift is already happening. It isn’t just about faster computers or better chatbots. It is about a fundamental change in the “physics” of how we detect fraud, assess credit, and manage compliance.</p><p>If you are working in risk management today, you are likely feeling the friction between the old tools and the new threats. We are moving from a world of static rules to one of fluid reasoning. Here is what that actually looks like when you peel back the layers.</p><h3>The Architectural Shift: From Fragile Point-Solutions to Foundational Reasoning</h3><p>For the last decade, machine learning in risk has been a massive assembly line of discriminative models. These are task-specific, heterogeneous point-solutions: one model for account takeover, one for credit limit increases, one for fraud at point-of-sale. While effective, this architecture is inherently fragile. Every time the underlying data distribution shifts or a new product feature is launched, these hundreds of specific models require manual retraining and monitoring.</p><p>In 2026, we are not discarding these existing models, but we are fundamentally re-orienting them. The future belongs to a more consolidated architecture centered around Foundational Models for Tabular Data. Instead of dozens of small models, we are moving toward a few large-scale models that understand the “universal grammar” of financial transactions and user behavior across the entire enterprise. These foundational models provide a robust, shared representation of risk that individual decisioning layers can tap into, significantly reducing the maintenance burden and increasing the speed of deployment.</p><p>However, data representation is only half of the puzzle. The real breakthrough is the integration of Reasoning Engines. Foundational and Large Language Models allow us to finally solve the “cold start” problem that has plagued risk departments for years. Because these models have a pre-existing logical foundation, they can ingest unstructured data — like messy logs, web telemetry, or application notes — and reason through logical inconsistencies with almost no specific training data. They do not need to see ten thousand examples of a new attack to stop it. They simply need to recognize that the intent is logically flawed.</p><p>The architecture of 2026 will also likely incorporate broader families of models, specifically World Models. A World Model maintains an internal simulation of the entire environment. In a risk context, this means the system doesn’t just look at a transaction in isolation. It reasons about the counterfactuals: “If I approve this loan under these specific economic shifts, how does it ripple through the portfolio’s liquidity?” This combined stack allows us to move from simple pattern matching to a sophisticated understanding of intent, causality, and long-term consequences.</p><h3>The Two Moats: Data and Active Enterprise Context</h3><p>With every vendor pitching an “AI Solution,” there is a dangerous temptation to outsource intelligence. It is easy to plug in a third-party model and feel like you have modernized your stack.</p><p>But this creates a strategic trap. If you rely entirely on a vendor’s brain, you are renting your competitive advantage.</p><p>The true moat in 2026 consists of two things that you cannot buy:</p><p><strong>1. Proprietary Data:</strong> This is the fuel. Foundation large language models are commodities; they are trained on the public internet. Your specific customer history, transaction logs, and interaction data are the only way to fine-tune a generic model into a specialist for your business.</p><p><strong>2. Active Enterprise Context:</strong> This is the piece most organizations miss. Context is the digital reflection of your business model. It includes your risk policies, your specific operational playbooks, and the tribal knowledge of “how we do things here.”</p><p>If you outsource your decisioning to a vendor, you are giving them this context. You are training their model, not yours. To win, you must capture this context and keep it in-house. When you combine your private data with your active enterprise context, you build a system that acts with your specific judgment, automated at scale.</p><h3>The Paradox of the Data Silo</h3><p>The prevailing wisdom has long been that “data hoarding” is the ultimate competitive advantage. The assumption was simple: the institution with the largest data lake wins. But we are discovering a paradox. As these datasets grow in volume, they often decrease in utility due to internal fragmentation.</p><p>Large organizations are frequently paralyzed by the structural reality of their own infrastructure. Valuable insights are trapped in disconnected ecosystems: transactional data in one datamart, identity metadata in another, and regulatory reporting in a legacy system. When data is siloed this way, it loses its “connective tissue.” Without a unified view, even the most advanced AI cannot find the subtle correlations that signify high-alpha credit opportunities or sophisticated fraud rings.</p><p>This creates a significant strategic hurdle. Having petabytes of data is a liability if that data cannot be unified and fed into the foundational models described above. The industry is reaching a tipping point where the primary challenge is no longer the acquisition of data, but the architectural courage required to break these silos. The organizations that fail to address this internal friction will find themselves “data-rich but insight-poor,” possessing all the ingredients for intelligence but unable to bake the cake.</p><h3>The Death of the Lagging Indicator</h3><p>One of the biggest shifts I see is the move away from traditional credit scoring as the <em>primary</em> source of truth.</p><p>We have to be realistic here. FICO, or a related commonly accepted score like D&amp;B, isn’t disappearing overnight. These metrics have deep regulatory acceptance and serve as a vital common language for the industry. When you are selling a portfolio or reporting to a regulator, that score is the shorthand everyone understands. It is the currency of communication.</p><p>But let’s not confuse communication value with predictive value.</p><p>A credit score is fundamentally a lagging indicator. It is a summary of the past. It is simplistic in the modeling sense and often redundant. For example, you can frequently re-derive a score just by looking at trade lines. It tells you what a borrower <em>did</em>, not what they <em>can do</em>.</p><p>The future of underwriting is Cash Flow Capacity.</p><p>By using AI to analyze real-time inflows and outflows, we can predict a borrower’s actual capacity to pay in the moment. This allows for “positive selection,” identifying creditworthy customers who might have a thin file or a temporary dip in their score, but whose cash flow shows strong momentum. It changes credit risk from a binary “Yes/No” decision based on history to a dynamic assessment of future potential.</p><h3>Agentic Commerce and the New Defense</h3><p>We also have to prepare for a change in who the “customer” actually is. We are entering the era of Agentic Commerce.</p><p>Soon, consumers will use digital financial assistants to manage their money. These agents will automatically move deposits to high-yield accounts, negotiate loan rates, and optimize spending across twenty different institutions. This drives incredible efficiency for the consumer, but it creates a nightmare for traditional fraud defense.</p><p>Our current defenses rely on human friction: biometrics, device fingerprinting, and behavioral navigation patterns. But an AI agent doesn’t have a fingerprint. It operates at machine speed.</p><p>In this world, the vectors of attack shift from simple theft to Abuse and Integrity.</p><p>We have to worry about deepfakes, content safety, and the “hijacking” of the agents themselves. If a user’s financial assistant is compromised, it could be manipulated to perform actions that look legitimate but are actually malicious — like moving funds to a “safe” account that is actually controlled by an attacker.</p><p>To defend against this, leading organizations are moving toward Synthetic Stress Testing and automated “Red Teaming.” Companies are building “Attacker Agents” that tirelessly probe their own systems, looking for zero-day vulnerabilities in code and logic before a real bad actor finds them.</p><p>Our defense, therefore, must shift from identifying “bad users” to ensuring Agentic Integrity. We need real-time systems that verify not just the transaction, but the provenance of the instruction itself. Is this the user’s true intent, or is it a hallucinated command from a compromised bot? This requires a new layer of safety that sits above the transaction, monitoring the health and logic of the agent ecosystem itself.</p><h3>From Systems of Record to Systems of Intelligence</h3><p>For decades, the Governance, Risk, and Compliance (GRC) platform has been the center of the risk universe. But we have to be honest about its role: it was designed as a System of Record. It was a digital ledger where we documented our risks and mapped them to controls so that, during an audit, we could prove we had a plan. These systems were inherently static. They held the rules, but they didn’t have the eyes or the hands to enforce them in real time. In 2026, we are moving toward Systems of Intelligence. In this new stack, the risk platform isn’t just a place to store data; it is an active participant. It doesn’t just record that a human checked a box; it verifies that the underlying data matches the policy.</p><p>Historically, the gap between a policy and its enforcement was filled by human labor. Risk professionals acted as the “detective layer,” manually sampling data to verify that controls were actually working. If a policy stated that every high-risk transaction required secondary approval, a human auditor might check a random sample of fifty cases at the end of the month to ensure the rule was followed. A System of Intelligence eliminates the need for sampling by turning the policy into executable code. By using agentic workflows, the system acts as a continuous observer sitting directly in the transaction flow. It can audit 100% of activity against the organization’s active enterprise context in milliseconds. We are moving from a world of forensic accounting to one of real-time verification.</p><p>This shift fundamentally changes the cadence of risk management from periodic to continuous. In a System of Record, you might discover a control failure three months after it happened — often during a post-mortem or a painful regulatory exam. A System of Intelligence operates with Continuous Control Testing. If a configuration change in a cloud environment suddenly exposes sensitive data, or if a new marketing campaign inadvertently violates a consumer protection regulation, the system identifies the drift immediately. It can flag the anomaly, trigger a remediation workflow, or even halt the activity before it creates a liability.</p><p>This also transforms the relationship with regulators. In the old model, compliance meant spending weeks gathering evidence and proving that controls worked in the past. It was a high-friction exercise in defensive documentation. With a System of Intelligence, the evidence is the byproduct of the operation itself. Because the system is verifying every action against the policy as it happens, the audit trail is live and immutable. We are moving toward a future where submitting an audit report is replaced by providing regulators with a real-time, transparent view into the health and logic of the risk engine. The goal isn’t just to save time; it is to build an intelligent layer that ensures the organization’s risk appetite is enforced by the technology itself, not just documented in a filing cabinet.</p><h3>The Architect’s Mindset</h3><p>All of this changes the role of the risk professional. We are moving from a “Reviewer” mindset — checking the math after the fact — to an “Architect” mindset.</p><p>The Architect’s job is no longer to manually adjudicate the exceptions. It is to design the boundaries. You are auditing the “chain of thought” of the model, ensuring it aligns with your enterprise context. You are building the guardrails that allow the AI to reason safely without hallucinating a policy violation.</p><p>This requires a new level of fluency. You don’t need to be a machine learning engineer, but you do need to understand the “physics” of the models you are governing. You need to know when a model is confident and when it is guessing. You need to understand the data lineage well enough to know if your “System of Intelligence” is actually intelligent or if it is just amplifying the biases of your legacy data.</p><p>Culturally, this demands a shift from “Top-Down” mandate to “Bottom-Up” innovation. The best ideas in this new era won’t come from a boardroom strategy deck. They will come from the analysts and engineers who see the raw data every day. They are the ones who will spot a weird pattern in the logs and realize, “An agent could handle this.” The Architect’s role is to empower those builders, giving them the tools and the permission to turn that insight into a defense.</p><p>We are entering a period where the “Risk Officer” and the “Product Builder” are becoming the same person. You cannot manage risk from a distance anymore. You have to be in the engine room, designing the safety systems directly into the core of the product.</p><p>The risk landscape of 2026 is undefined, but it is not unmanageable. The tools have changed, but the mission remains the same: to protect value while enabling growth. The difference is that now, we have the engines to do it with precision we could only dream of a decade ago.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6d6ff464610c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Pendulum of Power: A Decade of AI Org Evolution]]></title>
            <link>https://medium.com/@statnikov/the-pendulum-of-power-a-decade-of-ai-org-evolution-cdd672818f45?source=rss-31a9961fd6bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/cdd672818f45</guid>
            <category><![CDATA[organization]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-leadership]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Alexander Statnikov]]></dc:creator>
            <pubDate>Sat, 31 Jan 2026 01:11:45 GMT</pubDate>
            <atom:updated>2026-01-31T01:11:45.378Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LVMY3R8QQJccipyPWiAYkw.png" /></figure><p>The question of where Data Science, Machine Learning and AI should live is a topic that remains at the top of the agenda for almost every organizational leader I meet today. It is a logical obsession because these teams represent some of the most specialized talent in any company. They are instrumental for growth and operational efficiency, and companies want to align them in a way that is meaningful for the business while keeping the talent engaged and retained. Over the last 17 years, I have had hundreds of conversations with leaders across multiple industries about this exact puzzle. As we move through 2026, I want to capture the trends I am seeing and show how the organizational pendulum is swinging once again.</p><h3>The Unicorn Era (2010 to 2014)</h3><p>In the early 2010s, we experienced the first major boom of Data Science and Machine Learning organizations. While ML is technically a subset of AI, we did not typically use the term “AI” to describe these teams back then. There was a fierce, almost frantic competition for talent that often came well before companies had the infrastructure ready to support it. Outside of the major tech giants and a few financial services leaders, these organizations were quite small. Data professionals were often scattered across the business, sometimes sitting in analytics and sometimes embedded in tech, working on what were essentially experimental projects in isolation.</p><h3>The Great Centralization (2015 to 2017)</h3><p>As the field matured around 2015, we saw a massive wave of centralization. Companies realized that they needed to share knowledge and business context more effectively. Central organizations began to pop up in almost every large firm. In very large corporations that were split into multiple distinct businesses, it was common to see multiple central organizations operating in parallel. This era was defined by a desire for standards and a unified data strategy.</p><h3>Hub-and-Spoke and the Great Divorce (2018 to 2021)</h3><p>By 2018, the friction of pure centralization led to the emergence of the hub and spoke model. Business departments wanted more control over their own roadmap and felt that central teams lacked the necessary context to be effective. This period also marked a formal divorce between disciplines. Data Science became increasingly focused on product analytics and business insights, while Machine Learning evolved into a more engineering oriented discipline.</p><h3>The GenAI Democracy (2022 to 2024)</h3><p>The generative AI revolution changed math again. Suddenly, software engineers were jumping on the “AI bandwagon” and building applications without necessarily needing a dedicated Machine Learning engineer for every task. The ML specialists quickly found a new and deeper purpose in fine tuning or training large language models and managing an increasingly complex infrastructure. Even so, the teams remained loosely decentralized, with AI and ML sitting on the engineering side while Data Science stayed aligned with product and analytics.</p><h3>The 2026 Reality: The Rise of Foundational Hubs</h3><p>Now that we are firmly in the AI age, many companies are exploring how foundational models can reshape their operations. This is creating a fascinating shift in ownership. In the past, you might have had five different teams building five different time series models for five different lines of business, each using different signals. Today, the emphasis has shifted toward building or fine tuning a single foundational model that captures all those signals in one place. This requires a level of depth and collaboration that is difficult to achieve in a fragmented environment.</p><p>Because of this technological shift, we are seeing a move back toward centralization. However, these 2026 central teams are much leaner than the massive organizations of a decade ago because the technology itself has become so much more powerful. This shift primarily affects the AI and ML engineers. Data Science teams, by contrast, remain largely decentralized so they can continue providing product teams with the rapid insights they need for daily decision making.</p><h3>The Unit Economics of AI</h3><p>While the technical reasons for centralization are compelling, the financial reality often seals the deal. In a decentralized or embedded model, companies often face high compute costs because multiple teams are essentially paying a “redundancy tax” on GPUs/infra spend for similar tasks. A centralized foundational hub allows for optimized inference pipelines that drastically lower these costs. Furthermore, while a decentralized approach might seem to offer a faster time to market for shallow AI wrappers, it usually stalls when it comes to deeper innovation. A foundational hub might take longer to stand up initially, but it makes the rollout of secondary features almost instantaneous. Finally, this shift moves machine learning engineers away from repetitive plumbing and toward high value model design, tuning, and training, all while transforming fragmented data silos into a unified “360 degree view” of the customer.</p><h3>The Trust Factor</h3><p>While technology and business context justify these shifts, we cannot ignore the human factor. The variability of org structures is often driven by trust and the personal preferences of key stakeholders. Many business leaders simply prefer to have full control of their resources. They want to be able to “unblock themselves” without depending on a central team that might have different priorities. A business leader might be at peace with a central organization when things are going well, but the moment there is a disagreement, they often prefer to be in full control of their own destiny.</p><p>I have often heard these massive organizational decisions boiled down to very simple sentiments. A CEO might say, “I trust this person to be leader of central org,” or they might simply state, “I do not see value of central team”. In the eyes of the business, there is usually very little debate once these personal conclusions are reached. The friction between a business unit and a central organization can become so significant that it eventually outweighs any technical or career development benefits. Relationships are, and will always be, a critical factor in how these charts are drawn.</p><h3>Conclusion</h3><p>Looking ahead, the current trend shows AI and ML resources centralizing once more and aligning closely with engineering, while Data Science stays decentralized or aligned with product. I expect this to be the dominant model until we successfully deploy the next generation of robust, foundational models. At that point, AI and ML expertise might become so ubiquitous that the pendulum swings back toward decentralization yet again. The cycle of the organizational pendulum never truly ends: it just finds a new center of gravity based on the tools we have at our disposal.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cdd672818f45" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Agentic Moat: Why Enterprise Context is the Foundation of Modern AI]]></title>
            <link>https://medium.com/@statnikov/the-agentic-moat-why-enterprise-context-is-the-foundation-of-modern-ai-b78e41fa36b6?source=rss-31a9961fd6bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/b78e41fa36b6</guid>
            <category><![CDATA[ai-leadership]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-strategy]]></category>
            <category><![CDATA[enterprise-architecture]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Alexander Statnikov]]></dc:creator>
            <pubDate>Sat, 24 Jan 2026 20:39:02 GMT</pubDate>
            <atom:updated>2026-01-24T20:39:02.572Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XX8kqhQgPD_yLZGg5vGC8g.png" /><figcaption>Credits to Google Nano Banana Pro.</figcaption></figure><p>Everyone buys into the dream of a unified view of the customer that unlocks the signals essential for growth. It sounds intuitive and fundamental to any digital business. Yet, over 25 years of turning AI into revenue-driving products, I have observed a recurring reality: Data silos are not just persisting; they are proliferating at an accelerated rate.</p><p>As we race into the era of Agentic AI, the pressure to deliver immediate value is creating a “fragmentation tax.” We are frequently trading a sustainable, long-term AI advantage for short-term pilot wins. If we do not evolve our architecture now, we are not just building AI; we are building a new generation of technical debt that will be significantly harder to pay down later.</p><h3>The Pragmatic Trap of the Lean Financial Stack</h3><p>Consider a lean, modern financial services organization. To stay competitive, they have been admirably pragmatic. Their technology stack is often a patchwork of essential, specialized systems. For instance, they might use a banking core for transactions, a separate analytics platform for day-to-day reporting, and specialized point solutions for loan origination or compliance screening.</p><p>In this environment, there often is no central brain. Data exists in the spaces between these systems. Analytics teams survive by building brittle pipelines just to keep reports accurate. On the surface, the business looks functional and perhaps even optimized for its current scale.</p><p>Then, the AI wave arrives.</p><p>Today, leadership is pitched daily by innovative startups offering agentic solutions that promise to automate everything from customer success to marketing compliance. Because these startups are hungry for growth, they often offer low-fee pilots. The ROI looks undeniable, and the user experience is often superior to anything built in-house. This is where the catch begins to take root.</p><h3>The Hidden Cost of Agentic Silos</h3><p>The challenge is that every new third-party agent introduced often becomes a new data island. These solutions are almost always cloud-based and exist within the vendor’s own infrastructure. Over time, they accumulate your most precious asset, your customer interaction data, in a vacuum.</p><p>In the moment, it seems fine. There is no obvious reason why a customer success agent would need to speak to a marketing compliance reviewer. However, this lack of connectivity creates systemic risks that I have seen across the industry:</p><ul><li><strong>The Insights Gap:</strong> Discovering deep, cross-functional insights becomes virtually impossible when data is trapped in separate clouds.</li><li><strong>The Integration Tax:</strong> Engineering teams can quickly become “integration janitors,” maintaining dozens of connections that eventually outweigh the business benefit of AI tools.</li><li><strong>The Security Surface Area:</strong> Every third-party cloud instance is a new potential point of failure. A data breach at even a minor vendor can have a disproportionate impact on your brand and regulatory standing. Multiplying this risk by a dozen different agents creates a surface area that is incredibly difficult to monitor, let alone secure.</li></ul><p><strong>Most critically, there is the loss of historical memory.</strong> When you decide to switch vendors, you do not just lose a tool; you lose the context and the learning that the agent accumulated. Without a central foundation, your organization lacks a durable context, forcing every new agent to start from zero rather than building on the collective intelligence of the enterprise.</p><h3>From Static Records to Active Intelligence</h3><p>In the legacy world, we relied on the <strong>System of Record</strong>. This was primarily the banking core or other legacy transactional systems. These were repositories of <strong>static data</strong>: information sitting in a database waiting for a manual query.</p><p>To win in the agentic era, we must transition to a <strong>System of Intelligence</strong>. This shift is about moving from static data to <strong>active context</strong>. The goal is no longer just to store data, but to create an enterprise-wide context that agents can tap into in real-time.</p><p>Instead of sending your data to an agent, you invite the agent to work on your data platform. This allows the agent to act not just on the task at hand, but with the full weight of the organization’s historical memory and cross-functional signals behind it. This creates a feedback loop where every agentic action enriches the central context for every other agent.</p><h3>Building the Moat: Unified Access over Centralized Preaching</h3><p>This is not a call for old-school, rigid data centralization. In an agile environment, that approach is far too slow. Instead, the modern moat is built on unified access and architectural modularity.</p><ul><li><strong>Federated Context:</strong> Whether you utilize a data lake or a modern mesh architecture, the goal is to ensure data remains under your sovereignty. This ensures it can be used by downstream agents without constantly moving it into third-party environments. You are not forcing everyone into one massive database, but you are providing a unified access layer for your agentic workforce.</li><li><strong>Native Execution:</strong> Modern platforms now allow third-party applications and agents to run natively. This means the code comes to the data, not the other way around. This reduces the risk of data breaches and ensures you maintain a unified 360-degree view of your operations.</li><li><strong>Modular Agility:</strong> A solid data platform makes your business architecture modular. You can “hire” an AI agent for a specific task today, and if a better one emerges next year, you can “fire” the old one without losing your business intelligence. Your memory stays with you, not the vendor.</li></ul><h3>The Speed of the Feedback Loop</h3><p>Building these foundations is not just about data plumbing. It is about establishing excellence in risk and revenue. In my experience, the most powerful AI is the one that has the most context.</p><p>By building an enterprise-wide context and requiring your agents to play by your rules, you are not just mitigating risk. You are building a system that can learn, adapt, and out-iterate the competition. In the agentic world, the ultimate moat is the speed of your feedback loop, and that loop is powered by your data.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b78e41fa36b6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Most Business Data Isn’t Flat: Why Relational Learning Still Matters in the LLM Era]]></title>
            <link>https://medium.com/@statnikov/most-business-data-isnt-flat-why-relational-learning-still-matters-in-the-llm-era-2af11eb948d6?source=rss-31a9961fd6bc------2</link>
            <guid isPermaLink="false">https://medium.com/p/2af11eb948d6</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[relational-databases]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Alexander Statnikov]]></dc:creator>
            <pubDate>Sun, 18 Jan 2026 19:37:28 GMT</pubDate>
            <atom:updated>2026-01-18T19:37:28.681Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H8JIxzebbCxVvcnFvPbHMA.png" /><figcaption>Credits to Google Nano Banana Pro.</figcaption></figure><h3>Most business data isn’t flat</h3><p>I have long been fascinated by relational learning for a simple reason: most real-world data is relational. Customers have accounts, accounts have transactions, transactions have merchants, merchants have categories, and everything evolves over time. Yet a lot of machine learning practice still assumes the world is a single, perfectly “flattened” table.</p><p>That mismatch creates a hidden tax in enterprise ML. The value is in the relationships, but the models often see only the spreadsheet.</p><h3>The hidden tax: feature engineering</h3><p>I felt this gap most clearly when I moved from academia into financial services. The company I joined was ahead of the market, with hundreds of models in production. But those models were supported by large teams who spent a huge portion of their time manually engineering features from underlying relational data.</p><p>If you have ever built models on multi-table data, you know the drill: join tables (carefully); aggregate behavior (over windows); create ratios, deltas, trends; and repeat, then repeat again.</p><p>This work was valuable because new features often added signal and improved performance. But it was also grossly inefficient, hard to reproduce, and prone to overfitting. In practice, feature engineering became a craft activity that scaled with headcount.</p><h3>A small experiment that changed my mind</h3><p>I could not accept that status quo, so we tried something different. With a small team, we built an intelligent system to automate feature engineering and evaluation directly from relational data.</p><p>The idea was simple: treat feature discovery as a search problem.</p><p>Instead of asking humans to dream up transformations, the system explored candidate features, scored them, learned which transformations were productive, and iterated. The early framework we used resembled reinforcement learning: propose, evaluate, update, repeat.</p><p>The initial results were immediate and encouraging. In our first iteration, we could match or even beat the performance of FICO, a third-party score built on tradeline data, using only the raw relational inputs underlying that model. That gave us an important validation signal: the automation was not just producing noise, it was discovering real structure.</p><p>Over the next year, we applied the tool across multiple risk models. Each time, it surfaced features that improved performance and were still intuitive to domain experts. Looking back, we were only scratching the surface.</p><h3>Why this still matters in the LLM era</h3><p>Fast forward to today’s AI moment. Large language models have transformed workflows built around text. But the core enterprise ML problem remains: most important business decisions still depend on structured, relational data.</p><p>And LLMs do not naturally “see” structured data the way they see language. When we force them to work with tables, we often flatten tables into token sequences and lose the meaning encoded in schemas, relationships, types, and semantics.</p><p>The common workaround is indirect: ask the model to generate code, run it, then hope the result is correct. That can work for simple tasks, but it breaks easily in real enterprise environments, where ambiguous columns, messy joins, and brittle systems are the norm.</p><p>As a result, high-stakes sectors like financial services often remain anchored to traditional ML stacks for structured problems. In other words, we have powerful models for language, but we are still paying the same “relational tax” when the data lives in databases.</p><h3>What I think comes next</h3><p>There are real exceptions and real progress, but they are uneven: First, sequential and time series modeling is making big strides because transformer-style architectures can be applied more natively. Second, world models are advancing rapidly, but most breakthroughs have been concentrated in images and video, not enterprise relational systems. Third, agents and tooling can automate parts of the ML workflow, but they often struggle at the most critical step: reliably constructing datasets and features from multi-table enterprise data. Fourth, tabular (and relational) foundation models are the most direct attempt to attack the core issue: learning the structure of rows, columns, and schemas directly, not as flattened text. The structured datasets require models built natively for structured data, so they can understand schemas and relationships from the ground up.</p><p>To me, this is the most promising path to “leapfrogging” the current enterprise approach, where we maintain sprawling portfolios of task-specific models and pour time into manual feature engineering. If we get this right, we move from handcrafted features to learned representations that generalize across use cases, with less human toil and more consistent performance.</p><h3>The bottom line</h3><p>Enterprises are not short on data. They are short on usable representations of relational reality.</p><p>Relational learning is still the missing bridge between how data is stored and how models learn. We can keep flattening the world and paying the feature-engineering tax, or we can build systems that treat relational structure as the starting point.</p><p>That is why I think the next major wave of enterprise AI will be models and platforms that finally learn from the data businesses actually run on: tables, relationships, and time.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2af11eb948d6" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>