It's the Chief Data and AI Officer, not the Chief AI and Data Officer

Walk the halls of any major international bank today and you'll notice a new title on the executive floor: Chief Data and AI Officer. The title has grown quickly across financial services in 2026. Several banks appointed Chief Analytics Officers recently, tasking them with scaling AI across the business while also owning governance and oversight. Banks in Asia, Australia and Ireland have each made comparable appointments this year, and a leading US bank now projects roughly $570 billion in global AI-related debt issuance for 2026 ( FinTech Magazine).

What's notable isn't just the growth of Chief Data and AI Officers as a title, it's the word order in the title which suggests a more important message: Data comes first. AI comes second. The ordering is deliberate, and it's important to understand why because the market noise might suggest otherwise.

Conferences are full of generative AI pilots, copilots, and "First of a Kind" (FOAK) projects. But the numbers on what these deliver are sobering. MIT's NANDA initiative reviewed roughly 300 enterprise generative AI deployments and found that 95% failed to produce any measurable P&L impact, despite an estimated $30–40 billion in enterprise investment behind them ( Fortune, via Yahoo Finance). Analysts also predict through 2026, organizations will abandon a majority of AI projects that aren't built on AI-ready data. A large share of generative AI proofs-of-concept get shelved for exactly this reason ( Gartner). The AI sparkle is alluring; the business impact is often, less so.

Here are 5 reasons why the priority (and the title) is rightly focused on data.

1. AI is only as good as the data and context underneath it. MIT's research points to a "learning gap" in enterprise integration rather than model quality as the real driver of failed pilots ( Fortune). A 2026 peer-reviewed study in the journal Forecast puts a number on the gap in finance: When researchers ran identical machine-learning models on raw versus cleaned and economically validated financial data, data preparation alone lifted accuracy by an average of 15.6 percentage points and specificity by 26.9 points across six modelling techniques — with data quality accounting for roughly half of the total performance variation otherwise attributed to algorithm choice ( MDPI, Forecast). Here is the striking number: Half of what looks like an "AI problem" is actually a data problem wearing an algorithm's clothes. A model layered on top of fragmented, siloed or poorly governed transactional data will underperform, or worse, produce confidently wrong outputs. Banks that chase AI capability before fixing data foundations end up automating their own mess faster.

But quality is just part of the story; context is the other half. A loan balance, a transaction flag or a risk score is just a number until it's anchored to who the customer is, what relationship they have with the bank and what happened immediately before and after. Two identical transaction amounts can mean entirely different things depending on account history, counterparty and timing; that’s context a raw data field can't account for. Models trained on clean but decontextualized data can still make confidently wrong decisions, because "clean" and "correctly understood" aren't the same thing. A model layered on top of fragmented, siloed or poorly governed transactional data will underperform, or worse, fail quietly in exactly the high-stakes moments (including credit decisions, fraud flags and AML alerts) where the context mattered most. Banks that chase AI capability before fixing both data quality and data context end up automating their own mess faster, just with more confidence behind it.

2. Data is the durable asset; AI models are transient. Model architectures get replaced on an 18-month cycle. The customer, transaction and risk data a bank holds compounds in value over years and outlasts any single model generation. Separate industry analysis puts the share of AI project failures traceable directly to data problems, rather than algorithms, at 70% or higher ( Turning Data Into Wisdom). Much of a bank’s durable assets are unstructured to begin with. Consider that credit memos, loan files, call transcripts, underwriting submissions and compliance filings are the kind of content legacy systems were never built to read. McKinsey notes that gen AI and agentic systems depend on exactly this material, and every time a document is transformed or reused across a workflow, small quality issues compound and become harder to trace back to their source ( McKinsey). Digitizing and making these files searchable isn't the same as making them usable or governable for AI. Bringing the unstructured estate under the same lineage, access and quality discipline as structured data is arguably the harder half of the CDAO's mandate, but it's where most of a bank's uncaptured value is actually sitting.

3. Regulatory accountability starts with data, not algorithms. In banking, the Basel Committee's BCBS 239 standard has been in force for global systemically important banks for a decade, requiring banks to prove their risk data is accurate, complete and timely ( Bank for International Settlements). More recently, MiFID II (Markets in Financial Instruments Directive II) joins BCBS 239 as a second consequential data governance mandate imposed on global banks. MiFID II demands transaction-level reporting accuracy, reference data integrity, and best execution documentation across every trade. BCBS 239 demands that the risk data feeding those reports is itself governed, lineage-traced, and auditable. Together they create a compounding obligation: a bank cannot satisfy MiFID II’s reporting requirements without the governed, reliable, traceable data infrastructure that BCBS 239 requires underneath it.

BearingPoint's 2025 Regulatory Reporting Study of 50 institutions found only 18% report full implementation of BCBS 239 principles, and just 24% have extensive data lineage documentation ( BearingPoint). Deloitte's own BCBS 239 benchmark survey found a similar gap between ambition and delivery: 68% of banks expect more effective business steering from implementing the principles, but only 21% say they've actually achieved it ( Deloitte). An executive who can't answer for data lineage has no credible answer on AI risk either.

4. FOAK projects reward data readiness, not AI ambition. Gartner's own survey work found 63% of organizations either lack, or aren't sure they have, the data management practices AI actually requires ( Gartner). The FOAK pilots that do reach production are almost always built on top of data platforms that were already mature. The AI layer is often the easier part once the data plumbing is tested.

5. Depth of data control and acumen are important signals to the board and financial markets, beyond just regulators. Naming the role Chief Data and AI Officer rather than Chief AI and Data Officer tells the board and financial markets this executive won't chase flashy pilots at the expense of governance and quality, a distinction that matters in a regulated industry where a single bad model output, or a BCBS 239 finding, or a MiFID II reference data discrepancy can trigger time consuming processes and audits, supervisory reviews, Pillar 2 capital add-on or a ESMA query.

As the FOAK hype cycle matures, expect banks to quietly redirect budget from AI showcases toward the less glamorous work of data quality, lineage, context and architecture. The title Chief Data and AI Officer isn't semantics; it's an important statement of where the real, durable value and work of AI transformation actually begins.

• 

  Al Setikas

  Al Setikas

  Enterprise AI Control Plane

  Collibra