<?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 Umamaheswara rao Pragada on Medium]]></title>
        <description><![CDATA[Stories by Umamaheswara rao Pragada on Medium]]></description>
        <link>https://medium.com/@uma.pragada?source=rss-6c9988b9f264------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*WahSC5NJvIatLrc1za9naw.jpeg</url>
            <title>Stories by Umamaheswara rao Pragada on Medium</title>
            <link>https://medium.com/@uma.pragada?source=rss-6c9988b9f264------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 06:15:52 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@uma.pragada/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[Garbage In, Gospel Out: Why AI Governance is the New Data Liability]]></title>
            <link>https://medium.com/@uma.pragada/garbage-in-gospel-out-why-ai-governance-is-the-new-data-liability-b6979ee92a21?source=rss-6c9988b9f264------2</link>
            <guid isPermaLink="false">https://medium.com/p/b6979ee92a21</guid>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-quality]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Umamaheswara rao Pragada]]></dc:creator>
            <pubDate>Fri, 14 Aug 2026 19:30:59 GMT</pubDate>
            <atom:updated>2026-08-14T19:30:59.001Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>The oldest rule in computing just became the most dangerous one — and AI governance is the only thing standing between your enterprise and confident nonsense.</strong></p><p>For fifty years, <strong>“Garbage In, Garbage Out” (GIGO)</strong> was a comforting law. It was comforting because the garbage announced itself: a broken formula returned a #REF! error, or a corrupted feed produced a report where the numbers clearly did not foot. Bad input produced <em>visibly</em> bad output, and someone caught it before the board pack went out.</p><p><strong>Generative AI has broken that contract.</strong></p><p>When you feed a Large Language Model (LLM) incomplete, stale, or contradictory data, it does not return an error. Instead, it returns four elegant, confidently formatted paragraphs in a confident executive register with a recommendation at the end. It looks exactly like perfect data output. We have moved from “Garbage Out” — the version you could catch — to <strong>“Garbage Out that reads like Gospel”</strong> — the version that gets signed.</p><p><strong>Most Organisations are Governing Only One-Third of Their “Garbage”</strong></p><p>When a CDO says they have a “data quality programme,” they usually mean the structured data in the warehouse. That is only one of three “garbage streams” flowing into an AI system:</p><ol><li><strong>Enterprise Data:</strong> The governed structured estate, plus the “dark data” — decades of PDFs, outdated policies, and meeting notes sitting in SharePoint folders that are now used as RAG (Retrieval-Augmented Generation) indexes.</li><li><strong>The Pretraining Corpus:</strong> The internet’s errors, biases, and outdated facts that come pre-loaded in the model which you did not curate and cannot audit.</li><li><strong>Context &amp; Feedback:</strong> Hurriedly written system prompts and user feedback that often rates the <em>fluency</em> of an answer rather than its accuracy — which you then use to tune the system.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aXAkA38kio7gDSa04WXPhQ.png" /><figcaption><em>Most AI governance programmes cover the top band of the first stream only</em></figcaption></figure><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —</p><p><strong>When Abstractions Fail: Real-World Consequences</strong></p><p>This is not philosophical; it is a liability. Failures move budgets in a way that abstractions cannot.</p><ul><li><strong>Air Canada (2024):</strong> A tribunal held the airline liable for a discount its chatbot invented, rejecting the defence that the AI was a separate legal entity. <strong>Your model’s output is your statement.</strong></li><li><strong>Mata v. Avianca:</strong> Lawyers were sanctioned for submitting fictional, ChatGPT-generated legal citations that the model confidently claimed were real when challenged.</li><li><strong>Zillow Offers:</strong> An algorithmic valuation model failed to account for shifting market conditions, leading to hundreds of millions in writedowns.</li><li><strong>The “Hiding” RAG Index:</strong> In many enterprises, multiple versions of a policy exist in a vector store. The model synthesises a policy that has never existed by picking bits from each, citing a real filename while doing so.</li></ul><p><strong>Can Governance Stop Hallucinations?</strong></p><p>Hallucination has two distinct sources, and they require different answers:</p><ul><li><strong>Architectural:</strong> The model predicts the next likely token and has no internal “I don’t know” signal. Governance can <strong>shrink the surface area</strong> here, but not remove it.</li><li><strong>Contextual:</strong> This is a <strong>data governance defect</strong>. It happens when a model is given stale or ambiguous context. This is the majority of enterprise hallucination, and it <strong>is addressable.</strong></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JmSeLg6P-PJvpliXX5xrbg.png" /><figcaption><em>Governance cannot eliminate the architectural column. It can eliminate the contextual one</em></figcaption></figure><p><strong>The New Dimensions of Data Quality</strong></p><p>The classic “six dimensions” (Accuracy, Completeness, etc.) are no longer enough. To manage AI, you must measure seven new dimensions:</p><ul><li><strong>Groundedness:</strong> Is every claim traceable to a retrieved source?</li><li><strong>Provenance:</strong> Do we know where this content originated and who owns it?</li><li><strong>Representativeness:</strong> Does the data reflect the population the model will serve?</li><li><strong>Semantic Clarity:</strong> Does “active customer” mean the same thing everywhere?</li><li><strong>Freshness at Inference:</strong> Was the data true at the <em>moment</em> the model used it?</li><li><strong>Consent:</strong> Is the data lawfully permitted for this specific use?</li><li><strong>Safety &amp; Toxicity:</strong> What must never reach an output?</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*28RXdxQFyLJ7NRgZlC6ecw.png" /><figcaption><em>The classic set still holds. It no longer covers how AI consumes data</em></figcaption></figure><p><strong>The Escalation of Autonomy</strong></p><p>The risk compounds as we move toward <strong>Autonomous Agents</strong>. With a copilot, bad data produces a bad <em>suggestion</em>. With an agent that has tool access, bad data produces a <strong>bad action, already taken</strong> — an order placed or a payment released with no human in the loop.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QDPF_JC5_pC_O3ebyuhjCw.png" /><figcaption><em>Every increment of autonomy must be matched by an increment of data quality</em></figcaption></figure><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —</p><p><strong>Where to Start on Monday</strong></p><ol><li><strong>Tier use cases by consequence:</strong> Don’t govern a summariser the same way you govern a credit-limit agent.</li><li><strong>Declare your RAG corpus a data product:</strong> Assign owners, freshness SLAs, and retirement dates to SharePoint folders.</li><li><strong>Enforce Provenance:</strong> No citation, no answer.</li><li><strong>Label AI-generated artifacts:</strong> Prevent the “contamination loop” by tagging synthetic data at the point of creation so it doesn’t rot your future training sets.</li><li><strong>Wire AI incidents into the data quality backlog:</strong> Fix the cause (the duplicate record), not just the symptom (the hallucination).</li></ol><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — -</p><p><strong>The Bottom Line:</strong> AI governance is not a brake; it is the steering and the seatbelt. In the AI era, data quality is your model’s ground truth. Garbage in still means garbage out — it is just that the garbage is now beautifully written.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F-xbHi4ERMI6fIbD_99hBg.png" /><figcaption><strong>“AI governance is not the brake. It is the steering, the instrumentation, and the seatbelt — the set of controls that lets an organisation go fast without discovering, at scale and in public, that it was confidently wrong the whole time.”</strong></figcaption></figure><p><strong>What’s the most expensive “confidently wrong” AI output you’ve seen? Share your stories in the comments.</strong></p><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — -</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b6979ee92a21" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Where Data Governance Ends and AI Governance Begins]]></title>
            <link>https://medium.com/@uma.pragada/where-data-governance-ends-and-ai-governance-begins-5f5ed1d5064c?source=rss-6c9988b9f264------2</link>
            <guid isPermaLink="false">https://medium.com/p/5f5ed1d5064c</guid>
            <category><![CDATA[model-risk]]></category>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[banking-technology]]></category>
            <category><![CDATA[responsible-ai]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[Umamaheswara rao Pragada]]></dc:creator>
            <pubDate>Thu, 18 Jun 2026 11:22:53 GMT</pubDate>
            <atom:updated>2026-06-18T11:22:53.880Z</atom:updated>
            <content:encoded><![CDATA[<p>In banking, it’s not a line — it’s a handoff zone. And the handoff zone is where incidents happen.</p><p><strong><em>By Umamaheswara Rao Pragada — Principal Enterprise Architect, Data &amp; AI</em></strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DwRZorMZiaZNh_ltwrfLPw.png" /><figcaption>Data governance is the fuel, AI governance is the vehicle — and the contested middle is the pump, where ownership quietly goes missing.</figcaption></figure><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — -</p><p>Here’s a scenario that plays out in banks everywhere.</p><p>A team builds a RAG-based assistant that answers policy questions for the compliance department. The data team has done its job: the source documents are classified, access-controlled, and lineage-tracked. The AI team has done its job: the model is validated, the prompts are tested, hallucination checks are in place.</p><blockquote>Six months later, an internal review asks a simple question. The vector database holds embeddings of confidential policy documents — who classified them? What’s their retention period? Who approved the cloud region they sit in?</blockquote><p>Silence.</p><p>The data team assumed embeddings were a “model artifact” — the AI team’s problem. The AI team assumed embeddings were “just data in a different format” — already covered by data governance.</p><blockquote>Both teams did their jobs perfectly. Nobody owned the gap between the jobs.</blockquote><p>This is the real boundary problem in financial services AI today. Not “do we need AI governance?” — every regulator has answered that for us. The harder question is: <strong>where does data governance end, where does AI governance begin, and who owns the messy middle?</strong></p><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —</p><h3>Why banks get this confused</h3><p>Two reasons, and both are understandable.</p><p><strong>First, the disciplines grew up separately.</strong> Data governance in banking matured over the last 15 years — driven by BCBS 239, by the rise of the Chief Data Officer, and by frameworks like DAMA-DMBOK. It thinks in terms of <em>assets</em>: tables, files, reports. Quality, lineage, ownership, classification.</p><p>AI governance is arriving now — driven by the EU AI Act, ISO 42001, NIST’s AI Risk Management Framework, and growing supervisory attention from central banks. It thinks in terms of <em>behavior</em>: decisions, outcomes, fairness, accountability for what a system <em>does</em>.</p><p><strong>Second, AI systems blur the asset/behavior distinction.</strong> A traditional report is governed once and behaves predictably forever. An AI system’s behavior can change because someone edited a prompt, because the underlying data drifted, or because a vendor silently updated a model. The “asset” and the “behavior” are no longer the same thing — and each needs its own discipline.</p><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — -</p><h3>What data governance owns (and should keep owning)</h3><p>Data governance governs <strong>the asset, wherever it goes</strong> — including into an AI system:</p><ul><li><strong>Data quality</strong> — completeness, accuracy, timeliness rules and scorecards</li><li><strong>Metadata and the business glossary</strong> — what does this field actually mean?</li><li><strong>Lineage</strong> — where did this data come from? (BCBS 239 made this non-negotiable for risk data)</li><li><strong>Classification and privacy</strong> — public, internal, confidential; PII handling</li><li><strong>Ownership and stewardship</strong> — a named human accountable for each domain</li><li><strong>Retention and residency</strong> — how long, and in which jurisdiction</li></ul><p>Here’s the part many AI teams miss: <strong>none of this stops applying just because the data enters an AI pipeline.</strong> A confidential document doesn’t become less confidential because it was chunked and embedded. Data governance doesn’t “end” — its <em>scope of artifacts</em> ends, and that’s a different thing.</p><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — -</p><h3>What AI governance adds (that data governance never asked)</h3><p>AI governance governs <strong>behavior, decisions, and impact</strong>. It asks questions that have no equivalent in DMBOK:</p><ul><li><strong>Use-case risk tiering</strong> — a chatbot suggesting lunch and a model informing credit decisions are not the same risk. The EU AI Act’s risk-based approach makes this explicit, and credit scoring sits firmly in the high-risk tier.</li><li><strong>Model validation</strong> — does it actually work, on data it hasn’t seen, under conditions that will occur?</li><li><strong>Fairness and bias</strong> — does it treat customer segments equitably? Can we prove it?</li><li><strong>Explainability</strong> — can we tell a customer (or a regulator) <em>why</em> the system decided what it decided?</li><li><strong>Human oversight</strong> — where must a human review, approve, or be able to override?</li><li><strong>Monitoring and drift</strong> — is the system still behaving as validated, six months later?</li><li><strong>Prompt and agent change control</strong> — in generative AI systems, editing a prompt changes behavior with no code deploy. That edit needs the same discipline as a code release.</li></ul><p>The good news for banks: <strong>this is not a foreign discipline.</strong> Banks have run Model Risk Management for two decades — the Federal Reserve’s SR 11–7 guidance has required independent validation, documentation, and ongoing monitoring of models since 2011. AI governance is largely MRM extended to a new class of models, plus new dimensions (fairness, generative behavior, prompts) that classical MRM never contemplated. You’re not building a new muscle. You’re stretching one you already have.</p><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — -</p><h3>The contested middle — where incidents are born</h3><p>Between the two domains sits a set of artifacts that are <em>simultaneously data and model material</em>. This is the zone from the opening story, and in my experience it’s where governance fails in practice:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-qVdu8uTWPeKbQCzJt7ATQ.png" /><figcaption>The contested middle: seven artifact types that each team plausibly believes the other side already owns — which is exactly how they end up owned by no one</figcaption></figure><p>The pattern in every row is the same: <strong>each side has a plausible reason to believe the other side owns it.</strong> And artifacts that two teams plausibly don’t own are artifacts nobody owns.</p><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —</p><h3>A frame that works: the fuel, the vehicle, and the pump</h3><p>Here’s the mental model I’d offer:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GDBa5RqoLKzRDZBH5XrHCQ.png" /><figcaption>One compliance-assistant use case traced end to end: clean fuel (data governance), the contested pump (the handoff), and the moving vehicle (AI governance). Each box names a real banking artifact — and who should own it.</figcaption></figure><ul><li><strong>Data governance governs the fuel</strong> — its quality, its source, its handling, wherever it travels.</li><li><strong>AI governance governs the vehicle and the journey</strong> — what the system does, how it’s driven, who’s accountable when it takes a wrong turn.</li><li><strong>The handoff happens at the pump</strong> — and the pump needs a named owner.</li></ul><p>The “pump” rule in practice: <strong>every artifact in the contested middle gets an explicit, named owner — decided deliberately, not assumed.</strong> It genuinely doesn’t matter much whether embeddings end up owned by the data office or the AI function. What matters is that the decision is written down, so that “I thought you had it” becomes impossible.</p><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —</p><h3>One operating model, not two councils</h3><p>The instinct in many organizations is to stand up a brand-new AI governance committee, separate from the data governance council. I’d argue that’s exactly wrong — for banks especially. You end up with two forums claiming the same artifacts, double approvals for every AI use case, and the contested middle falling between two chairs instead of one.</p><p>The better pattern: <strong>extend the operating model you have.</strong></p><ol><li><strong>Same council, expanded seats</strong> — add model risk, compliance, and AI engineering to the existing data governance forum.</li><li><strong>Same intake, new checkpoints</strong> — when a use case involves AI, the existing data approval flow gains AI-specific gates: risk tier, validation evidence, human-oversight design.</li><li><strong>Same stewards, new artifact types</strong> — extend data stewards’ scope to cover the contested middle, or appoint “AI artifact stewards” within the same structure. Either works; ambiguity doesn’t.</li><li><strong>Same lineage backbone, extended reach</strong> — your metadata and lineage tooling should trace not just table-to-report, but document-to-embedding-to-prompt-to-output. If your catalog can’t see your vector database, that’s a finding.</li></ol><p>This is also what regulators implicitly expect. BCBS 239 lineage, SR 11–7 validation, and EU AI Act documentation requirements all converge on the same demand: <strong>an unbroken chain of accountability from source data to AI-driven decision.</strong> One operating model can produce that chain. Two rarely do.</p><h3>The people question: data roles vs. AI roles</h3><p>One operating model still needs the right seats at the table — and this is where the role question comes in. Are AI governance roles just data governance roles with a new label? Mostly no. Each classic data role has an AI counterpart, but the counterpart asks a genuinely different question.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZXBkQayoN9RvpSNb1MRSaw.png" /><figcaption>Five classic data roles extend cleanly into AI counterparts; two — the independent validator and the responsible-AI officer — have no ancestor in DMBOK and must be created new.</figcaption></figure><p>Most AI roles are extensions of roles banks already have — but the independent validator and the responsible AI officer are genuinely new.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lbO1VDi4ChN_MItKaemCGg.png" /><figcaption>Data roles vs. their AI counterparts. The amber rows mark the two roles that cannot simply be relabelled from an existing data function — they are genuinely new.</figcaption></figure><p>Three takeaways from the mapping:</p><ol><li><strong>Most AI roles are extensions, not inventions.</strong> Owner, steward, custodian, council — the pattern transfers. What changes is the unit of governance: from data <em>domain</em> to AI <em>use case</em>, and from asset quality to system behavior.</li><li><strong>Two roles cannot be extensions: the independent validator and the responsible AI officer.</strong> The first because independence is the whole point (three-lines-of-defense logic banks already live by), the second because the question it asks — fairness and impact — has no ancestor in DMBOK.</li><li><strong>The AI Artifact Steward is the cheapest fix on this list.</strong> One paragraph in an existing steward’s mandate — “you also own the embeddings, the RAG corpus, and the prompt library for your domain” — closes the exact gap where most incidents are born.</li></ol><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — -</p><h3>The bottom line</h3><p>The question &quot;where does data governance end and AI governance begin?&quot; has a slightly uncomfortable answer: <strong>data governance never ends — it follows the data into the AI system. AI governance begins the moment data starts driving behavior and decisions.</strong> Between them sits a contested middle that belongs to whoever you explicitly assign it to — and to nobody, if you don&#39;t.</p><blockquote><strong>Data governance governs the fuel. AI governance governs the vehicle and the journey. The handoff happens at the pump — and the pump needs an owner.</strong></blockquote><p>Banks don&#39;t need a second governance empire. They need one extended operating model, a named owner for every artifact in the middle, and the honesty to treat every &quot;I thought you had it&quot; as the audit finding it really is.</p><p>— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — — -</p><p><em>This post is part of my series on data and AI governance that actually works in practice. Coming next: how to extend a data governance operating model for AI, step by step. Tell me in the comments — in your organization, who owns the embeddings?</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5f5ed1d5064c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Critical Data Elements: How to Govern What Matters Without Boiling the Ocean]]></title>
            <link>https://medium.com/@uma.pragada/critical-data-elements-how-to-govern-what-matters-without-boiling-the-ocean-c6e08e09eadc?source=rss-6c9988b9f264------2</link>
            <guid isPermaLink="false">https://medium.com/p/c6e08e09eadc</guid>
            <category><![CDATA[data-management]]></category>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[data-quality]]></category>
            <category><![CDATA[critical-data-element]]></category>
            <category><![CDATA[banking]]></category>
            <dc:creator><![CDATA[Umamaheswara rao Pragada]]></dc:creator>
            <pubDate>Wed, 17 Jun 2026 09:10:43 GMT</pubDate>
            <atom:updated>2026-06-17T09:28:46.342Z</atom:updated>
            <content:encoded><![CDATA[<p>You can’t govern everything. A practical, risk-based method for choosing the thirty fields that actually matter — and the courage to leave the rest.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wS9cI0wDHVUSg9enxhIsHw.png" /><figcaption>Part 3 of 3 — Data Governance That Actually Works: how to govern what matters without boiling the ocean.</figcaption></figure><p><em>This is the final part of a three-part series on practical data governance. Part 1 made the case for governance as a risk capability. Part 2 examined the operating model and the stewardship layer where most programs fail. Part 3 tackles the question every program faces on day one: where to begin, when you cannot govern everything at once. Each post stands on its own, but together they build a mental model practitioners can apply on Monday morning.</em></p><p>There is a failure mode in data governance so common it deserves its own name. A program launches with energy and ambition. The team decides, reasonably enough, that good governance means governing the data — all of it. They begin cataloging every field, defining every term, writing quality rules for everything in sight. Six months later, they have a sprawling, half-finished catalog, a team that is exhausted, and a sponsor who is asking, not unreasonably, what the program has actually delivered.</p><p>The answer is: a little bit of everything, which is to say, nothing anyone can point to.</p><p>This is boiling the ocean, and it is fatal because it confuses <em>coverage</em> with <em>value</em>. Not all data is equally important. A field that drives a regulatory submission and a field that records an internal comment are both data, but governing them with equal intensity is a profound misallocation of scarce attention. The most important discipline in data governance is not the discipline of governing. It is the discipline of choosing what to govern — and, just as importantly, what to leave alone for now.</p><h3>The idea of a Critical Data Element</h3><p>A Critical Data Element, or CDE, is exactly what the name suggests: a data field whose accuracy, completeness, and timeliness genuinely matter to the organization. Not every field is critical. In most organizations, a surprisingly small fraction of fields carries the overwhelming majority of the risk and value. The art is identifying them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tcDR-OdryyLJSr45isoS0g.png" /></figure><p>Think of it as a funnel. At the top sits all enterprise data — thousands of fields across every system, far more than any program could govern well. Narrow to the data that is actually used in reports, decisions, and processes, and the number shrinks. Narrow again to the data that carries genuine risk — regulatory, financial, fraud-related, or reputational — and you are left with a manageable set. At the bottom of the funnel sit the Critical Data Elements: typically thirty to fifty fields that a program should govern first, govern thoroughly, and govern before anything else.</p><p>Thirty to fifty may sound modest against an estate of thousands. That modesty is the point. Thirty elements governed thoroughly — with agreed definitions, clear ownership, traceable lineage, and active quality monitoring — deliver more risk reduction than five hundred elements governed superficially. Depth on what matters beats breadth across what doesn’t.</p><h3>How to choose them</h3><p>Identifying CDEs is not a matter of intuition or seniority of opinion. It is a scoring exercise, and running it as one keeps the decision defensible and free of politics. I find five criteria sufficient, and in a banking context they map naturally onto the risks that keep executives awake.</p><p>The first is <strong>regulatory impact.</strong> Does this field feed a regulatory return, a financial-crime monitoring process, or fall within a compliance scope such as card data security? If a regulator could ask about it, it is a strong candidate.</p><p>The second is <strong>financial impact.</strong> Does the field flow into settlement, into fraud-loss calculations, into revenue recognition? Money attracts scrutiny, and fields that touch money inherit that scrutiny.</p><p>The third is <strong>operational impact.</strong> Does an error in this field cause incidents, customer complaints, or reconciliation breaks? Fields that generate operational pain when they go wrong are fields worth getting right.</p><p>The fourth is <strong>reputational impact.</strong> Would a public failure involving this data damage trust — with customers, partners, or the market? Some data carries weight out of proportion to its volume.</p><p>The fifth is <strong>downstream reuse.</strong> How many reports, dashboards, models, and processes depend on this single field? A field consumed by a dozen downstream uses is a single point of failure worth governing, because one error propagates everywhere.</p><p>Score each candidate field against these five, weight them according to your organization’s risk appetite, and a ranked list emerges. The fields at the top are your CDEs. The exercise is not perfectly precise, and it does not need to be. Its purpose is to replace “govern everything” and “govern whatever the loudest voice wants” with a transparent, risk-based prioritization that the council can stand behind.</p><p>Here is how a handful of fields familiar to anyone in payments or banking might score.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/659/1*by6GTJiGca6xaVLs6uQlGg.png" /></figure><p>The discipline this table represents is the discipline of saying no. The marketing-preference flag is not unimportant, but it is not where scarce governance capacity should go first. Saying so explicitly — and recording why — is what separates a program with focus from one that drifts.</p><h3>What “governing” a CDE actually means</h3><p>Once a field is designated a CDE, governing it is not a vague aspiration. It is a concrete and finite set of actions, and a useful way to know whether a CDE is genuinely governed is to check it against four questions.</p><p>Does it have an agreed definition? Someone — a steward who knows the domain — has written down what the field means, and the relevant stakeholders have accepted that definition. This sounds trivial until you discover how often a single field means three different things to three different teams.</p><p>Does it have a clear owner? A named steward is accountable for the field’s definition and quality, with the authority and the time to exercise that accountability.</p><p>Can its lineage be traced? You can show where the field originates, how it is transformed, and where it ends up. When a regulator asks how a number was derived, the answer is a diagram, not a three-day investigation.</p><p>Is its quality monitored? There are active rules checking that the field is accurate, complete, timely, and valid — and someone is alerted when those rules are breached, before the bad data reaches a report or a model.</p><p>A CDE that satisfies all four is governed. A CDE that satisfies only some is a work in progress, and naming the gap honestly is more useful than pretending the work is done.</p><h3>The strategic payoff of starting small</h3><p>There is a counterintuitive truth buried in the CDE approach. Starting small is not a compromise forced by limited resources. It is the strategically superior choice even when resources are ample.</p><p>A program that governs thirty CDEs thoroughly within ninety days has something to show: a working capability, a demonstrable reduction in risk on the fields that matter most, and — crucially — a proof point that converts skeptical executives into sponsors. A program that attempts to govern everything has, at ninety days, nothing finished and nothing to demonstrate. The first program earns the credibility and the funding to expand. The second program is quietly wound down.</p><p>Governance scales by earning the right to scale. The CDE approach is how a program earns that right: deliver undeniable value on a focused scope, then use that success to justify the next tranche of investment. Breadth is the reward for depth, not a substitute for it.</p><h3>Where this leaves us — and the series</h3><p>Critical Data Elements are the answer to the hardest practical question in governance: where to begin. By scoring fields against regulatory, financial, operational, reputational, and reuse impact, a program can identify the thirty to fifty fields that carry most of the risk, govern them thoroughly, and demonstrate value before attempting more. The discipline of governance turns out to be, in large part, the discipline of choosing — and of having the courage to leave the unimportant ungoverned.</p><p>Across this series, a single argument has been building. In Part 1, we saw that governance is the shift from reactive crisis to proactive control — a shift that, in banking especially, is the difference between an institution that is trusted and one that is merely tolerated. In Part 2, we saw that this shift requires a real operating model, one that lives or dies on whether the stewardship layer is given genuine time, framing, and authority. And in Part 3, we have seen that the model must start focused, governing what matters before reaching for what doesn’t.</p><p>The thread connecting all three is that data governance is not, at its core, a technical problem. It is an organizational discipline that technology serves. The frameworks are well documented and freely available. What is rare — and what separates programs that work from programs that produce documents — is the judgment to apply them to the reality of how organizations actually behave.That judgment cannot be certified or downloaded. It is built, one focused success at a time.</p><p><em>This concludes the three-part series on practical data governance. Part 1 covered proactive governance as risk mitigation; Part 2 covered the operating model and the stewardship layer; this part covered Critical Data Elements and prioritization.</em></p><p><em>If the series was useful, I’d welcome your own war stories — the place governance worked, and the place it didn’t, are often the same lesson seen from two sides.</em></p><p><strong>Tags:</strong> #DataGovernance #DataManagement #EnterpriseArchitecture #Banking #FinancialServices #DataQuality #DMBOK #DataStrategy</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c6e08e09eadc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Data Governance Operating Model: From Org Chart to Operating Capability]]></title>
            <link>https://medium.com/@uma.pragada/the-data-governance-operating-model-from-org-chart-to-operating-capability-475d0e6d3750?source=rss-6c9988b9f264------2</link>
            <guid isPermaLink="false">https://medium.com/p/475d0e6d3750</guid>
            <category><![CDATA[data-stewardship]]></category>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[operating-models]]></category>
            <category><![CDATA[finance-and-banking]]></category>
            <category><![CDATA[data-management]]></category>
            <dc:creator><![CDATA[Umamaheswara rao Pragada]]></dc:creator>
            <pubDate>Thu, 11 Jun 2026 06:46:24 GMT</pubDate>
            <atom:updated>2026-06-11T06:46:24.772Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Every program has the pyramid slide. Almost none survive contact with the stewardship layer. Here’s why — and what to do about it.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kItkYqSKWGIPUPSDoWvXsA.png" /><figcaption>Part 2 of 3 — Data Governance That Actually Works: the operating model and the layer where most programs quietly fail.</figcaption></figure><blockquote>This is the second of a three-part series on practical data governance. Part 1 made the case for why governance matters as a risk capability. Part 2 turns to structure: how to build an operating model that actually works, and the one layer where most programs quietly fail. Part 3 will tackle where to begin. Each post stands on its own, but together they build a mental model practitioners can apply on Monday morning.</blockquote><p>There is a particular slide that appears in almost every data governance program. It shows a tidy pyramid of committees — a steering group at the top, a council beneath it, an office below that, stewards at the base. Everyone nods. The slide is approved. And then, six months later, the program has produced policies that nobody follows and a glossary that nobody reads.</p><p>The slide was not wrong. It was just incomplete in the way that matters most: it described the <em>structure</em> without describing the <em>work</em>. An operating model is not an org chart. It is a division of labor in which each layer does something the others cannot, and the whole only functions when every layer is genuinely doing its part.</p><p>Let me walk through the model that works, and then spend the rest of this article on the layer where it most often breaks.</p><h3>The five layers</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0LvCv-uxUkBsbQKUXUQJdA.png" /><figcaption>Strategy, tactics, engine, hands, feet — each layer is a distinct division of labor. Remove one and governance becomes a document.</figcaption></figure><p>The model has five layers, and the cleanest way to remember them is by what each one owns: <strong>strategy, tactics, engine, hands, feet.</strong></p><p>The <strong>Steering Committee</strong> owns strategy. This is the executive layer — the most senior leaders in the organization, who release funding, set direction, and provide the political cover that governance needs to survive competing priorities. They do not run the program. They protect it. In a regulated institution, this group also carries the relationship with the regulator, which means governance is never far from the board’s line of sight.</p><p>The <strong>Data Governance Council</strong> owns tactics. This is where the cross-functional decisions get made — which data to prioritize, which policies to ratify, how to resolve a dispute between two business areas that define the same term differently. The council is deliberately diverse: senior leads from operations, risk, compliance, technology, fraud, and analytics, each representing a constituency that governance must serve. A council that approves everything unanimously is not doing its job; the friction is the point.</p><p>The <strong>Data Governance Office</strong> owns the engine. This is the only permanent, full-time team in the model. It curates the business glossary, runs the metadata and lineage tooling, configures the quality rules, drafts the policies, and trains the stewards. The office is rarely glamorous — it is daily triage of quality alerts and patient curation of definitions — but it is the layer that does these things every single day. Without it, everything else is occasional.</p><p>The <strong>Stewardship Teams</strong> own the hands. Stewards are the people who actually know the data — what a field means, why a value looks the way it does, which edge cases matter. They are drawn from the business, not from a central function, and they own the definitions and quality rules for their domain.</p><p>The <strong>Local Governance Committees</strong> own the feet. In any organization large enough to have divisions, enterprise standards must be translated into divisional execution. Local committees take the enterprise standard — “every critical data element must have an owner and a quality rule” — and decide, for their division, which elements those are and how the work gets done.</p><p>Authority flows down through these layers. Knowledge flows up. When both flows are healthy, governance is a living capability. When either is blocked, it calcifies into paperwork.</p><h3>The layer where it fails</h3><p>If I had to predict where a given governance program will struggle, I would not bet on the steering committee or the office. I would bet on the stewards. The stewardship layer is where the most programs quietly die, and the reasons are worth understanding because they are almost never technical.</p><p>The first failure is <strong>stewardship as a title, not a role.</strong> An executive announces that the head of a department is now the data steward for that domain. The title is conferred; no time is allocated; no expectations are set. The steward, already fully occupied, treats it as the honorific it functionally is. Nothing changes. The lesson here is uncomfortable but clear: a steward without protected time is not a steward, and the failure belongs to whoever assigned the role without the capacity to perform it.</p><p>The second failure is <strong>asking for stewardship in the abstract.</strong> A governance office sends out a request: please nominate data stewards. The request lands as one more corporate ask competing with real deadlines, and it is quietly deprioritized. The organizations that succeed do something different. They do not ask for stewards in the abstract; they ask people to take ownership of specific problems those people already have. If a particular reconciliation breaks every month and frustrates a specific team, that team does not need to be persuaded to own the underlying data — they need to be shown that stewardship is the path to making the pain stop. Stewardship framed as a solution to an existing problem is accepted. Stewardship framed as a new obligation is resisted.</p><p>The third failure is <strong>stewards without authority.</strong> A steward is told they own a definition but has no mechanism to enforce it, no forum to escalate to, no backing when a more powerful colleague overrides them. Ownership without authority is just accountability for things you cannot control — and capable people learn quickly to avoid such roles. This is why the council layer matters so much: it is the steward’s backstop, the place where a steward’s decision is defended rather than quietly reversed.</p><p>The pattern across all three failures is the same. The stewardship layer fails for organizational and human reasons, not technical ones. And because the failures are human, they cannot be fixed with better tooling. They are fixed with better role design, genuine time allocation, and real authority.</p><h3>What each layer actually decides</h3><p>To make this concrete, it helps to see the kinds of decisions that belong at each layer. Decisions placed at the wrong layer are another common failure — a council that drowns in operational minutiae, or an office forced to make strategic calls it has no authority to make.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/673/1*AEpoF81v4RqBp9b06t8cZQ.png" /></figure><p>The test for whether a decision sits at the right layer is simple: push it as far down as it can responsibly go. Strategic and cross-functional decisions rise to the council and steering committee; everything else belongs closer to the people who know the data. A program in which every decision escalates upward has not built governance — it has built a bottleneck.</p><h3>A federated balance, not a central authority</h3><p>There is a temptation, especially early, to centralize everything. The governance office wants to own all definitions, set all rules, control all access. It feels efficient. It is, in fact, the fastest route to irrelevance, because a central function can never know the data as well as the people who work with it daily, and a standard imposed without local ownership is a standard that will be quietly ignored.</p><p>The model that works is federated. The center sets the <em>what</em> — the standards, the methods, the non-negotiable principles. The edges decide the <em>which</em> and the <em>how</em> — which elements matter for their domain, how the work fits their reality. The council is the forum where these two forces meet and negotiate. Done well, federation gives an organization both consistency and ownership, which is the combination that makes governance stick.</p><h3>Where this leaves us</h3><p>A governance operating model is not the slide with the pyramid. It is a working division of labor in which the steering committee provides cover, the council makes decisions, the office runs the engine, the stewards supply the knowledge, and the local committees execute. The structure is necessary but not sufficient: the program lives or dies on whether the stewardship layer is real, and that depends on time, framing, and authority far more than on any tool.</p><p>But even a perfectly structured model faces an immediate and humbling question on day one. You cannot govern everything at once. So where do you begin?</p><p><em>This is Part 2 of a three-part series on practical data governance. Part 1 made the case for proactive governance as risk mitigation. Part 3 tackles prioritization through Critical Data Elements — how to start without boiling the ocean.</em></p><p><em>Have you seen the stewardship layer succeed or fail in your own organization? I’d be glad to compare notes.</em></p><p><strong>Tags:</strong> #DataGovernance #DataManagement #EnterpriseArchitecture #Banking #FinancialServices #DataQuality #DMBOK #DataStrategy</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=475d0e6d3750" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data Governance as Proactive Management & Risk Mitigation]]></title>
            <link>https://medium.com/@uma.pragada/data-governance-as-proactive-management-risk-mitigation-ff59e53e3337?source=rss-6c9988b9f264------2</link>
            <guid isPermaLink="false">https://medium.com/p/ff59e53e3337</guid>
            <category><![CDATA[banking]]></category>
            <category><![CDATA[data-management]]></category>
            <category><![CDATA[data-quality]]></category>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[financial-services]]></category>
            <dc:creator><![CDATA[Umamaheswara rao Pragada]]></dc:creator>
            <pubDate>Tue, 09 Jun 2026 21:00:49 GMT</pubDate>
            <atom:updated>2026-06-09T21:20:21.593Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A practitioner’s view on the shift from firefighting to control — and why banking, more than any industry, can’t afford to skip it.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6jVWa_COewLfJR80xkh_bg.png" /><figcaption>Part 1 of 3 — Data Governance That Actually Works: a practitioner’s series on banking data governance.</figcaption></figure><blockquote>This is the first of a three-part series on practical data governance, written from the implementation side rather than the framework side. Part 1 makes the case for why governance matters — as a risk capability, not paperwork. Part 2 examines the operating model, and the single layer where most programs quietly fail. Part 3 tackles the question every program faces on day one: where to begin, when you cannot govern everything at once. Each post stands on its own, but together they build a mental model practitioners can apply on Monday morning.</blockquote><p>Picture a Monday morning that plays out somewhere in financial services almost every week.</p><p>A regulator has asked a simple question: how was a particular figure in last quarter’s submission derived? It should take ten minutes. Instead, three teams spend three days. The figure was assembled from two source systems, transformed by a pipeline nobody fully remembers, and reconciled against a report whose definition of the underlying metric quietly differs from the one used upstream. Each team produces a slightly different number. None can prove which is correct. The regulator, meanwhile, is still waiting — and now wondering what else can’t be answered quickly.</p><p>Nobody in this story did anything obviously wrong. There was no breach, no fraud, no negligence. And yet the organization just paid a real cost — in senior time, in regulatory confidence, and in trust — for a problem that was entirely preventable.</p><p>This is what the absence of data governance looks like. Not a dramatic failure, but a steady tax paid in crises that didn’t have to happen.</p><p><strong>What data governance actually is</strong></p><p>Strip away the frameworks and the certifications, and data governance is a deceptively simple idea. It is the exercise of authority, control, and shared decision-making — planning, monitoring, and enforcement — over how an organization manages its data.</p><p>That definition is worth slowing down on, because each word carries weight. Authority means someone is empowered to set standards and make them stick. Control means access, quality, and classification are deliberately managed, not left to chance. Shared decision-making means the people who own the business context are part of the decisions, not subjects of them. And enforcement means the standards have consequences, not just aspirations.</p><p>But the most useful way to understand governance is not through its definition. It is through the shift it represents.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xuaCfEQf7b_im7nbGElREQ.png" /><figcaption><strong><em>The reactive-to-proactive shift: the cost doesn’t disappear, it moves — from crisis paid on the regulator’s schedule to discipline paid on your own.</em></strong></figcaption></figure><p><strong>The shift from reactive to proactive</strong></p><p>Without governance, organizations operate reactively. They fix broken reports after the report is broken. They investigate breaches after the data has already left. They scramble to answer regulators after the question has been asked. The work is real and often heroic — but it is firefighting, and firefighting is expensive precisely because it is unplanned.</p><p>With governance, the same organization operates proactively. Reports reconcile because the definitions were agreed before the report was built. Sensitive data is protected because it was classified the moment it entered the estate. Regulators get clear answers because lineage was captured as the data flowed, not reconstructed under pressure. The work is still real — but it is planned, predictable, and defensible.</p><p>The cost does not disappear. It moves. In a reactive organization, the cost is paid in crisis. In a proactive one, it is paid in discipline. The entire economic argument for governance rests on a single observation: discipline is far cheaper than crisis, and it is paid on your schedule rather than the regulator’s.</p><p><strong>Why this matters more in banking than almost anywhere</strong></p><p>Every industry benefits from governance, but financial services has a structural reason to take it more seriously than most. Banking runs on data the way other industries run on inventory or equipment. A loan decision, a fraud score, a regulatory return, a settlement instruction — these are not supported by data, they are data. When the data is wrong, the business is wrong, often before anyone notices.</p><p>Consider how the reactive-versus-proactive distinction plays out across the situations any banking data professional will recognize.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/647/1*NJHjNLUGatxlPqzU1hKRVA.png" /></figure><p>Read down the right-hand column and a pattern emerges. Proactive governance is not about producing more documents. It is about converting unpredictable, crisis-driven costs into predictable, manageable ones.</p><p><strong>The five risks governance is really mitigating</strong></p><p>When executives ask what governance is for, the honest answer is risk reduction. It helps to be specific about which risks. I find it useful to group them into five, because naming them turns an abstract program into a concrete value proposition.</p><p>The first is rework — the broken reports and reconciliation breaks that consume engineering and analyst time. The second is regulatory risk — the findings, queries, and potential penalties that come from being unable to demonstrate control. The third is reputational risk — the erosion of trust, internal and external, when data cannot be relied upon. The fourth is revenue risk — the fraud losses, missed opportunities, and failed product launches that flow from bad data. And the fifth is resource drain — the slow bleed of senior attention consumed by firefighting that should have been prevented.</p><p>Every governance investment can be justified by pointing to which of these five it reduces. If a proposed initiative doesn’t clearly reduce one of them, it probably isn’t governance — it’s bureaucracy wearing governance as a costume.</p><p><strong>A caution from the field</strong></p><p>It would be dishonest to present governance as a guaranteed win, so let me offer the most common way it goes wrong. Organizations frequently mistake documentation for governance. They produce a glossary, a policy library, a catalog — and then declare victory while the actual behavior of the organization remains entirely unchanged. The reports still don’t reconcile. The regulator still waits three days.</p><p>Governance is not the artifacts. It is the change in behavior the artifacts are supposed to produce. A glossary that no one consults is not governance; it is a document. The test of a governance program is never how many policies it has written. It is whether the Monday-morning scenario at the start of this article still happens.</p><p><strong>Where this leaves us</strong></p><p>Data governance, properly understood, is not a compliance exercise grudgingly endured. It is the discipline that determines whether an organization manages its most important asset deliberately or improvises with it under pressure. In banking, where the data is the business and the regulator is always one question away, that distinction is the difference between an institution that is trusted and one that is merely tolerated.</p><p>Understanding why governance matters is the easy part. The hard part is structuring it so that it actually works — so that authority reaches the people who hold the knowledge, and standards survive contact with operational reality. That structure is the subject of <em>Part 2: the operating model, and the single layer where most governance programs quietly fail</em>.</p><p><em>Part 2 publishes next Tuesday. If you’d like to read it when it lands, the easiest way is to follow me here on Medium.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ff59e53e3337" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Your Data Strategy Keeps Failing: The Hidden Tension Between Aspiration and Architecture]]></title>
            <link>https://medium.com/@uma.pragada/why-your-data-strategy-keeps-failing-the-hidden-tension-between-aspiration-and-architecture-cc9bb8f5a855?source=rss-6c9988b9f264------2</link>
            <guid isPermaLink="false">https://medium.com/p/cc9bb8f5a855</guid>
            <category><![CDATA[data-strategy]]></category>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[enterprise-architecture]]></category>
            <category><![CDATA[data-management]]></category>
            <category><![CDATA[dama-dmbok]]></category>
            <dc:creator><![CDATA[Umamaheswara rao Pragada]]></dc:creator>
            <pubDate>Fri, 13 Feb 2026 05:32:37 GMT</pubDate>
            <atom:updated>2026-02-13T05:32:37.613Z</atom:updated>
            <content:encoded><![CDATA[<p>Two frameworks from the DAMA Body of Knowledge reveal the uncomfortable truth about how organizations build — and should build — their data management capabilities</p><p>I’ve spent over two decades architecting enterprise data platforms across banking and financial services — from enterprise data warehouses serving millions of transactions to modern AI-driven analytics pipelines. Across every engagement, every transformation program, and every “data strategy roadmap” presentation, I’ve watched the same pattern repeat itself with almost theatrical predictability.</p><p>Organizations build the roof before they lay the foundation. And then they wonder why the house leaks.</p><p>Two frameworks from the DAMA Body of Knowledge — Aiken’s Pyramid and Sue Geuens’ Functional Area Dependencies model — capture this tension with remarkable clarity. Together, they tell us not just <em>what</em> data management looks like when done right, but <em>why</em> it so often goes wrong. Understanding both is essential for anyone serious about building data capabilities that actually endure.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yZsea3ktn3t0OHIY_CpWQQ.png" /></figure><p>• • •</p><h3>Aiken’s Pyramid: The Honest Architecture of Maturity</h3><p>Peter Aiken took the familiar DAMA Wheel — that circular diagram showing Data Governance at the center surrounded by ten knowledge areas — and reimagined it as a pyramid. This wasn’t merely an aesthetic choice. It introduced two powerful ideas simultaneously.</p><p>The first idea is structural dependency. Like any physical pyramid, the lower layers must support the upper ones. At the base sits Data Governance, flanked by Data Architecture, Data Quality, and Metadata Management. Above them: Data Storage &amp; Operations, Data Security, Data Modeling and Design. Higher still: Data Integration &amp; Interoperability. And only at the very top — resting on everything below — sit Data Warehousing, Business Intelligence, Reference &amp; Master Data, Documents &amp; Content, and the pinnacle: Advanced Practices including Analytics, Mining, and Big Data.</p><p>The message is architecturally intuitive: you cannot have reliable analytics without reliable integration, which requires reliable storage and security, which demands proper architecture, quality controls, and governance. Skip a layer, and the structure becomes unstable.</p><p>The second idea is more subversive — and more honest. Aiken acknowledged that organizations rarely build in this ideal sequence. In practice, these capabilities may be “put in place in an arbitrary order.” A bank might implement a data warehouse before establishing formal data governance. A retailer might deploy advanced analytics before standardizing metadata management. The pyramid describes the ideal structural relationship, not the actual construction sequence.</p><p>This dual nature — aspiration and realism — is what makes Aiken’s Pyramid so enduring. It holds up the mirror and says: <em>here’s what a well-architected data capability looks like, and here’s the messy reality of how you’ll probably build it.</em></p><p>• • •</p><h3>Geuens’ Dependencies: The Structural Truth</h3><p>Sue Geuens took a different approach entirely. Rather than describing maturity phases, she mapped the functional dependencies between DAMA knowledge areas — the actual structural relationships that exist whether organizations acknowledge them or not.</p><p>Her framework arranges the data management functions into a clear dependency stack. At the top sits Business Intelligence and Analytics — the capabilities everyone wants, the dashboards executives demand, the insights that justify the entire data investment. Directly below: Master Data Management and Data Warehousing — the solutions that BI directly depends on. You cannot produce trustworthy analytics without clean master data and well-structured warehouse solutions.</p><p>Below that: Systems and Applications — the feeding systems that populate master data repositories and data warehouses. These are the operational platforms, the core banking systems, the CRM tools, the ERP environments. Supporting the systems layer: Data Quality, Data Design, and Data Integration &amp; Interoperability — the practices that ensure systems produce and exchange reliable data. Without these, systems produce garbage, and garbage flows downstream to poison everything above.</p><p>At the very foundation: Data Governance — and within it, Metadata Management, Data Security, Data Architecture, and Reference Data Management. These are the bedrock capabilities on which every other function depends.</p><p>The elegance of Geuens’ model is its directness. There are no phases, no acknowledgment of arbitrary ordering. It simply states: <em>these are the dependencies. Violate them at your peril.</em></p><ul><li>• •</li></ul><h3>The Uncomfortable Synthesis: Data Debt Is Real</h3><p>Here’s what two decades of hands-on experience have taught me about where these two frameworks intersect — and what they collectively mean for practitioners.</p><blockquote>Aiken tells us how organizations behave. Geuens tells us why that behavior has consequences.</blockquote><p>In financial services, I’ve lived through this cycle repeatedly. A business unit demands a new analytical capability — say, real-time credit risk monitoring. The project gets funded, the data warehouse gets extended, the dashboards get built. Six months later, someone discovers that the underlying customer master data has a 12% duplication rate across three source systems. The risk scores are wrong. The dashboards are lying. The remediation project costs three times the original build.</p><p>This is what I call data debt — the compounding cost of building upper layers without properly establishing lower ones. Like technical debt in software engineering, data debt accrues interest. Every report, every model, every dashboard built on ungoverned, unvalidated, poorly-architected data becomes a liability that grows more expensive to fix over time.</p><p>The pattern is so consistent across organizations that it’s almost a law: <em>the cost of retrofitting foundational data management capabilities increases exponentially with the volume of dependent capabilities built on top.</em></p><ul><li>• •</li></ul><h3>Five Lessons from the Intersection</h3><p>Having navigated this terrain across multiple large-scale banking transformations, here’s what I believe practitioners need to internalize:</p><p>First, governance is not optional — it’s structural. Both frameworks place Data Governance at the foundation. This isn’t philosophical; it’s architectural. Without governance, you lack the decision-making framework to resolve the inevitable conflicts that arise when multiple systems, multiple business units, and multiple definitions of “customer” or “revenue” collide. Organizations that treat governance as a compliance checkbox rather than an operational capability are the ones that struggle most with data quality, integration, and ultimately analytics.</p><p>Second, metadata is the nervous system, not a nice-to-have. Geuens places Metadata Management within the governance foundation — not alongside data warehousing or analytics. This is profoundly correct. Without comprehensive metadata — business glossaries, data lineage, technical mappings — every other function operates blind. I’ve seen organizations spend millions on data quality tools without first cataloging what data they have, where it comes from, and what it means. It’s like performing surgery without an anatomy chart.</p><p>Third, the “arbitrary order” reality demands conscious debt management. Aiken’s acknowledgment that organizations build capabilities in non-ideal sequences isn’t a license for chaos — it’s a call for conscious trade-off management. If you must build a data warehouse before establishing formal governance (and sometimes business reality demands this), then at minimum: document what you’re deferring, estimate the remediation cost, and build a concrete plan to address the foundation. Treat it like financial debt — acceptable when managed, catastrophic when ignored.</p><p>Fourth, integration is the layer that reveals all sins. Notice that in both frameworks, Data Integration &amp; Interoperability sits in a critical middle position — above the foundational practices, below the analytical outputs. This is because integration is where poor architecture, inconsistent metadata, weak governance, and bad data quality all become visible. Integration doesn’t create these problems; it exposes them. Organizations that struggle with data integration are almost always dealing with upstream foundational failures.</p><p>Fifth, the dependency chain means transformation is bottom-up, even when sponsorship is top-down. Executive sponsors fund analytics and BI because that’s where visible business value lives. But sustainable delivery requires investing downward. The most successful data transformations I’ve been part of took a “boring first” approach: establish governance, define the business glossary, build the metadata catalog, standardize reference data — and then accelerate the visible capabilities on a solid foundation. The irony is that this approach delivers faster over a three-to-five year horizon, even though it appears slower in the first six months.</p><p>• • •</p><h3>A Practical Framework for Sequencing</h3><p>For practitioners wrestling with where to start, I propose a synthesis of both models — a phased approach that respects Geuens’ dependencies while acknowledging Aiken’s reality:</p><p>Phase 1 — Establish the Foundation (Governance + Architecture + Metadata): Stand up a Data Governance operating model — not a massive bureaucracy, but a lightweight decision-making framework. Define your enterprise data architecture principles. Begin your business glossary and metadata catalog with the most critical business terms and data assets.</p><p>Phase 2 — Strengthen the Core (Quality + Security + Design + Integration): Implement data quality monitoring on your most critical data flows. Establish security classifications and access controls. Standardize data modeling practices and build integration patterns. This is where you make your systems layer trustworthy.</p><p>Phase 3 — Build the Analytical Capabilities (Warehouse + Master Data + BI): With foundations in place, build your data warehouse, master data management, and BI capabilities. These will be faster to implement, more reliable, and cheaper to maintain because the underlying layers are solid.</p><p>Phase 4 — Advance and Innovate (Analytics + AI + Big Data): Now — and only now — invest in advanced analytics, machine learning, and big data capabilities. These are the most dependency-heavy functions in the entire stack, and they will only deliver value if every layer below them is trustworthy.</p><ul><li>• •</li></ul><h3>The Bottom Line</h3><p>Aiken’s Pyramid and Geuens’ Dependencies Framework are not competing models — they are complementary lenses on the same fundamental challenge. Aiken gives us the maturity perspective: organizations evolve toward better data management, often in imperfect sequences. Geuens gives us the architectural perspective: regardless of the order you build things, the dependency relationships are immutable.</p><p>The organizations that thrive are the ones that use Aiken’s honesty to plan realistically and Geuens’ clarity to build structurally. They acknowledge the mess while building toward the ideal. They manage data debt consciously rather than discovering it in crisis.</p><p>In an era where every organization aspires to be “data-driven” and “AI-powered,” these foundational frameworks from DAMA remind us of a timeless truth: <strong>the quality of your insights can never exceed the quality of your foundations.</strong></p><blockquote><em>Build from the bottom. The view from the top is worth the patience.</em></blockquote><p><em>Umamaheswara Rao Pragada is a Principal Enterprise Architect for Data &amp; AI with over 20 years of experience in banking technology, specializing in enterprise data platforms, data governance, and AI-driven solutions for financial services. He currently leads Enterprise Data Warehouse and AI initiatives at a major financial institution.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cc9bb8f5a855" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data Governance Roles Demystified: A Practitioner’s Guide]]></title>
            <link>https://medium.com/@uma.pragada/data-governance-roles-demystified-a-practitioners-guide-a4042351153c?source=rss-6c9988b9f264------2</link>
            <guid isPermaLink="false">https://medium.com/p/a4042351153c</guid>
            <dc:creator><![CDATA[Umamaheswara rao Pragada]]></dc:creator>
            <pubDate>Sat, 27 Dec 2025 12:09:33 GMT</pubDate>
            <atom:updated>2025-12-27T12:09:33.849Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why Most Data Governance Programs Struggle with Role Clarity</h3><p><em>Part 1 of a 2-Part Series on Building Effective Data Governance Teams</em></p><p>I’ve spent over two decades implementing enterprise data systems in banking and financial services. And if there’s one pattern I’ve seen repeatedly, it’s this: <strong>organizations invest heavily in data governance tools and frameworks, yet struggle because nobody truly understands who does what.</strong></p><p>Sound familiar?</p><p>You’re not alone. According to Gartner, through 2025, 80% of organizations seeking to scale digital business will fail because they don’t take a modern approach to data governance. In my experience, the root cause often isn’t the framework or technology — it’s role confusion.</p><p>In this article, I’ll share a practitioner’s perspective on the six essential data governance roles, and more importantly, reveal a key insight that transformed how I approach DG implementations.</p><h3>The Six Essential Data Governance Roles</h3><p>Before we dive in, let me be clear: these aren’t roles I invented. They’re well-established in frameworks like DMBOK (Data Management Body of Knowledge). But what I’ve learned through years of implementation is that <em>understanding</em> these roles and <em>operationalizing</em> them are two very different things.</p><p>Let’s break them down.</p><h3>1. Data Owner</h3><p><strong>Who they are:</strong> Senior business executives accountable for a specific data domain.</p><p><strong>What they do:</strong></p><ul><li>Approve data access requests and authorization levels</li><li>Define data quality requirements and acceptable thresholds</li><li>Authorize data sharing agreements</li><li>Resolve escalated data issues</li><li>Represent their domain in governance councils</li></ul><p><strong>The reality check:</strong> Data Owners often don’t know they’re Data Owners. In one implementation, I found that the Head of Retail Banking had been making data access decisions for years without realizing he was functioning as the Data Owner for customer data. Formalizing this role gave him — and his team — clarity and accountability.</p><h3>2. Data Steward</h3><p><strong>Who they are:</strong> The hands-on practitioners who manage data definitions, metadata, and quality rules daily.</p><p><strong>What they do:</strong></p><ul><li>Maintain business glossary and data dictionary entries</li><li>Define and document data quality rules</li><li>Manage data lineage documentation</li><li>Investigate and resolve data quality issues</li><li>Ensure metadata consistency across systems</li></ul><p><strong>The reality check:</strong> This is often the most misunderstood role. Data Stewards aren’t IT people, and they’re not business people — they’re the bridge between both. They need to understand what “Customer Balance” means to the business AND how it flows through your core banking system, data warehouse, and reporting layer.</p><h3>3. Data Custodian</h3><p><strong>Who they are:</strong> Technical professionals responsible for the safe storage, security, and infrastructure management of data.</p><p><strong>What they do:</strong></p><ul><li>Implement and maintain database infrastructure</li><li>Execute backup, recovery, and disaster recovery procedures</li><li>Implement security controls and access management</li><li>Monitor system performance and optimize storage</li><li>Ensure compliance with technical security standards</li></ul><p><strong>The reality check:</strong> Data Custodians are often confused with Data Owners. Here’s the key difference: <strong>Custodians protect and store the data; Owners decide who can use it and how.</strong> Your DBA implements role-based access controls — but the Head of Treasury decides who gets access to trading data.</p><h3>4. Data Producers</h3><p><strong>Who they are:</strong> The systems, applications, and organizational units that create, capture, or generate data.</p><p><strong>What they do:</strong></p><ul><li>Capture data accurately at the point of origin</li><li>Adhere to data quality standards at entry</li><li>Provide timely data feeds to downstream systems</li><li>Document data generation processes</li><li>Report data quality issues to stewards</li></ul><p><strong>In financial services context:</strong> Your core banking systems, card management platforms, treasury systems, and trading platforms are all Data Producers. In a typical bank, you might have 15+ source systems generating data — each with its own data formats, update frequencies, and quality characteristics.</p><h3>5. Data Consumers</h3><p><strong>Who they are:</strong> The users, applications, and systems that access and utilize data for analysis, reporting, and decision-making.</p><p><strong>What they do:</strong></p><ul><li>Use data in accordance with governance policies</li><li>Report data quality issues to stewards</li><li>Provide feedback on data usability and requirements</li><li>Maintain data confidentiality and security</li><li>Request new data access through proper channels</li></ul><p><strong>In financial services context:</strong> Think of your regulatory reporting teams, risk analytics, asset-liability management, customer analytics, and management information systems. These teams consume data from your enterprise data warehouse and depend on its accuracy for critical decisions.</p><h3>6. Data Governance Manager</h3><p><strong>Who they are:</strong> The senior professional who oversees the entire governance program and coordinates across all roles.</p><p><strong>What they do:</strong></p><ul><li>Develop and maintain governance policies and standards</li><li>Coordinate data governance council meetings</li><li>Track and report on data quality metrics</li><li>Manage governance tools and technology</li><li>Drive governance maturity improvements</li></ul><p><strong>The reality check:</strong> This role is the glue that holds everything together. Without a dedicated DG Manager, governance becomes everyone’s responsibility — which quickly becomes no one’s responsibility.</p><h3>The Key Insight: Roles Come From Different Areas of Expertise</h3><p>Here’s what took me years to fully appreciate, and it’s the insight that transformed my approach to data governance implementations:</p><blockquote><strong><em>Data Governance roles are not created in a vacuum. They are formed from different areas of expertise across the organization.</em></strong></blockquote><p>This might sound obvious, but the implications are profound. Let me explain.</p><p>When organizations try to “build” a data governance team from scratch, they often fail because they’re looking for unicorns — people who understand business, technology, compliance, and data management equally well.</p><p><strong>That’s the wrong approach.</strong></p><p>Instead, recognize that each DG role naturally draws from a specific area of expertise:</p><p>RoleArea of Expertise<strong>Data Owner</strong>Business Units — These are your department heads and senior managers who already make decisions about their domain<strong>Data Steward</strong>Business-IT Hybrid — Subject matter experts who understand both data and business context<strong>Data Custodian</strong>IT/Technical Teams — Your DBAs, infrastructure specialists, and security professionals<strong>Data Producers</strong>Source Systems — The applications and teams that generate data as part of operations<strong>Data Consumers</strong>Downstream Teams — Finance, Risk, Compliance, Analytics teams who use data for decisions<strong>DG Manager</strong>Enterprise Architecture or CDO Office — Someone who can coordinate across all domains</p><h3>Why This Matters</h3><p>When you understand that DG roles come from different expertise areas, several things become clear:</p><p><strong>1. You don’t need to hire an entirely new team.</strong> The people who should be Data Owners, Stewards, and Custodians are likely already in your organization. They’re just not formally recognized in those roles.</p><p><strong>2. Role conflicts become easier to resolve.</strong> When the DBA (Custodian) and the Business Head (Owner) disagree about data access, it’s not a power struggle — it’s a governance process. The Custodian implements what the Owner decides.</p><p><strong>3. Training becomes targeted.</strong> Instead of generic “data governance training,” you can provide role-specific guidance. Data Owners need to understand accountability and decision rights. Stewards need metadata management and quality rule definition. Custodians need security implementation.</p><p><strong>4. The DG Manager’s job becomes clearer.</strong> They’re not managing the data — they’re coordinating across all these different expertise areas to ensure governance actually happens.</p><h3>A Visual Framework</h3><p>To help my teams understand this concept, I developed a visual framework that shows how these roles interact:</p><p><strong>Governance Layer (Strategic Oversight)</strong></p><ul><li>Data Owner ←→ DG Manager</li><li>Sets policies, standards, and accountability</li></ul><p><strong>Data Lifecycle Layer (Operational Execution)</strong></p><ul><li>Producers → Custodians → Stewards → Consumers</li><li>Shows how data flows and who manages it at each stage</li></ul><p>The key visual insight is that <strong>Data Owners and DG Managers sit above the operational flow</strong>, providing oversight and governance, while the other roles operate within the data lifecycle.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*a_s5kuYnE-Ivkm-hA9REvw.png" /><figcaption>Data Governance Roles and Responsibilities</figcaption></figure><h3>Common Mistakes to Avoid</h3><p>Based on my implementation experience, here are the pitfalls I see most often:</p><h3>Mistake #1: Treating Data Ownership as IT Responsibility</h3><p>Data ownership is a <strong>business</strong> accountability. When IT “owns” data, nobody is truly accountable for its business meaning and quality.</p><h3>Mistake #2: Hiring Data Stewards Without Subject Matter Expertise</h3><p>A Data Steward for customer data should understand customer lifecycle, KYC requirements, and how customer information flows through your systems. Don’t hire generic “data people” — hire domain experts who can learn data management.</p><h3>Mistake #3: Confusing Data Custodians with Data Owners</h3><p>Just because the DBA controls access to the database doesn’t mean they should decide who gets access. Technical control ≠ Business accountability.</p><h3>Mistake #4: Ignoring Data Producers in Governance Discussions</h3><p>Your source systems and the teams operating them are critical stakeholders. If they don’t understand governance requirements, data quality issues will persist at the source.</p><h3>Mistake #5: Not Having a Dedicated DG Manager</h3><p>Part-time governance leadership produces part-time results. Someone needs to wake up every day thinking about data governance.</p><h3>Getting Started: Three Practical Steps</h3><p>If you’re looking to clarify data governance roles in your organization, here’s where I’d start:</p><h3>Step 1: Map Your Current State</h3><p>Identify who is <em>actually</em> making data decisions today, even informally. Who approves access requests? Who defines data quality rules? Who maintains your data dictionary (if you have one)? You’ll likely find that governance is happening — it’s just not formalized.</p><h3>Step 2: Align Roles to Expertise Areas</h3><p>Don’t fight organizational structure — work with it. Your Data Owners should come from business leadership. Your Custodians should come from IT. Your Stewards should be those rare individuals who bridge both worlds.</p><h3>Step 3: Start Small</h3><p>Don’t try to govern all data at once. Pick one critical data domain — customer data, financial transactions, or regulatory reporting data — and implement clear roles for that domain first. Learn, iterate, then expand.</p><h3>What’s Next: The Coordination Challenge</h3><p>Understanding the six roles is just the beginning. The real challenge is coordination: How do you get people from different expertise areas, with different priorities and incentives, to work together effectively?</p><p>That’s where the <strong>Data Governance Management Team</strong> comes in — the central coordination body that monitors, controls, and escalates governance activities across the organization.</p><p>In <strong>Part 2 of this series</strong>, I’ll dive deep into:</p><ul><li>How to structure a DG Management Team</li><li>The Monitor | Control | Escalate framework</li><li>A practical RACI matrix for governance activities</li><li>Tips for building governance culture</li></ul><h3>Your Turn</h3><p>I’d love to hear from fellow practitioners:</p><ul><li>How does your organization define data governance roles?</li><li>What challenges have you faced with role clarity?</li><li>Which role has been most difficult to operationalize?</li></ul><p>Drop a comment below or connect with me on LinkedIn. Let’s learn from each other’s experiences.</p><p><em>This article is Part 1 of a 2-part series on building effective data governance teams. Part 2 will cover the Data Governance Management Team and coordination frameworks.</em></p><p><strong>About the Author</strong></p><p><em>Umamaheswara Rao Pragada is a Principal Enterprise Architect specializing in Data &amp; AI at a leading financial institution, with over 20 years of experience in banking architecture and enterprise data systems. He has implemented enterprise data warehouses, metadata management platforms, and AI systems for regulatory compliance. Umamaheswara Rao is passionate about bridging the gap between data management theory and practical implementation.</em></p><p><em>Connect with Umamaheswara Rao Pragada on LinkedIn | Follow for more data governance insights</em></p><p><strong>Tags:</strong> #DataGovernance #DataManagement #EnterpriseArchitecture #Banking #FinancialServices #DataQuality #DMBOK #DataStrategy</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a4042351153c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>