Every AI roadmap we read this year opens with the same phrase: “AI-first.”
Scroll past the first slide, though, and you’ll usually find a data model that was never fully designed or implemented, and certainly not fit to be read by a machine looking to understand the organisation and how it operates most were built for a silo purpose by somebody that left three reorgs ago.
That gap, between the ambition on slide one and the programmatic data schema underneath it, is where most AI programmes quietly die. Not in the model. In the underlying data model.
It’s not a new problem wearing an AI costume. It’s the oldest problem in data management, showing up at the one moment nobody can afford to ignore it anymore.
The Number That Should Be on Every COO & CDOs Dashboard
Every analyst report I’ve read this year lands on the same uncomfortable number, give or take a few points.
Somewhere between 70% and 85% of failed AI and ML initiatives fail because of the data underneath them, not the model on top (AImultiple). Roughly one in ten organisations currently holds data of sufficient quality to run AI at scale. And Gartner projects that most AI projects lacking AI-ready data risk being abandoned before 2026 is out.
Figure 1. The failure pattern is consistent across every major 2025–2026 study: it’s the data, not the model.
Whatever the exact figure, every study agrees on the pattern: the technology is rarely why AI fails. The data foundation is. For a CDO, that’s not a statistic to file away, it’s a mandate. Data understanding, models & quality just became the single biggest determinant of whether your AI strategy earns its budget or becomes next year’s write-off.
Treat readiness as a gate, not a curve
For twenty years, data quality has been a maturity curve something you climb steadily, forever. AI doesn’t tolerate the curve. It needs a gate.
A human analyst spots a wrong number and quietly discounts it. A model trained on that same number propagates the error with total confidence, at scale, into every decision downstream. So before you fund a use case, test five things:
- Completeness: does the data represent the real, relevant population to support real understanding of the organisation, or just the easy 50%?
- Definitional consistency: does “customer” mean the same thing in every conversation, business process, report and source system?
- Lineage: can you trace a data point back to its origin?
- Currency: is it fresh enough for how fast the AI has to decide?
- Ownership: is someone actually accountable, or is the dataset orphaned?
We wrote more on why bad data compounds rather than just costs once in The Cost of Bad Data.
Zillow lost $881m. The AI was working perfectly.
The Zestimate model never malfunctioned it was well-validated and publicly accurate. What broke was the decision built on top of it: Zillow quietly turned an advisory price estimate into a binding purchase offer for a physical, illiquid asset.
The model had only ever seen a pandemic-era housing boom. When the market cooled in late 2021, it kept pricing offers at the peak, textbook concept drift. Zillow shut the programme down and cut roughly 25% of its workforce.
The lesson isn’t “validate your data.” It’s narrower: a model’s validation is scoped to the context it was validated for. Move it from advisory to transactional and that validation stops transferring – someone needs the authority to say so before it goes live.
The AI Accuracy Ceiling – Your Data Model Is Now the Reasoning Substrate for Every Agent You Deploy
This is the section most CDOs are underweight, because it sounds like classical data architecture. Truth is, that’s spot on. However, the harsh reality is that it’s become the single biggest lever on AI accuracy.
To be blunt, the data model becomes the machine readable definition of the organisation – so the right question to ask is whether the CDO is even the right custodian. Is this not a business problem that needs a COO or business facing Chief AI Officer at the helm?
If an agent querying a raw, undocumented schema has to guess what a column means from its name, these guesses compound into hallucinations.
Recent text-to-SQL benchmark work puts it in numbers: accuracy against a raw, unmodelled schema sits around 40%. Against a well-modelled semantic layer with clear business definitions, that climbs to 83–90%+. GraphRAG (knowledge-graph-grounded retrieval) shows the same pattern for anything that requires reasoning across entities rather than simple lookup.
Figure 2. The data model is the accuracy ceiling.
A semantic layer isn’t a nice-to-have anymore. It’s the interface between natural-language intent and your actual data. Without one, every agent you deploy is reverse-engineering your data model in real time, one prompt at a time and getting it wrong on scale, at expense, and in ways nobody notices until it falls over.
Three things follow for the organisation, CDO, COO and Chief AI Officer:
- Your data model is the reasoning substrate, not a build artefact.
Business terms, ownership and relationships need to be machine-readable metadata, not tribal knowledge in confluence or on a wiki.
- Composite and reference models matter more than ever. On our IRB Mortgages lineage pilot with Solidatus, the highest-value work wasn’t the scanning, it was mapping business terms, ownership and classification onto the physical model.
That’s exactly what an agent needs to reason reliably.
- Legacy entity-resolution debt becomes AI risk. If “customer” means something different in your CRM process vs your core banking platform, an agent won’t ask a colleague to clarify.
It’ll pick one, silently, and act.
We explored the commercial angle in Data Products Are Failing in Most Organisations, most “data products” are old datasets with a new label, still missing the modelling discipline that makes them reusable.
Governance Debt: The Bill AI Is Forcing You to Pay
Every organisation carries technical debt. Fewer talk about governance debt – years of undefined ownership, unresolved definitions and unmapped lineage, always due for delivery “next quarter”. It’s invisible on the balance sheet until something forces a valuation event. For most organisations, AI is that event.
Picture your AI stack as a building. The visible floors of strategy, GenAI use cases, dashboards sit on foundations of ownership, lineage and definitions. When those foundations were never properly poured, the building doesn’t collapse. It leans. Every new floor of AI ambition makes the lean harder to ignore.
Figure 3. Governance debt: cracking beneath the AI stack.
Three symptoms tell you it’s already dangerous: no single named owner per critical domain, definitions that exist in five places and agree in none, and lineage that was documented once and never maintained.
None of these are AI problems. They’re data fundamentals that AI has simply stopped letting you defer. The fix isn’t a governance programme restart, it’s quantifying the debt like a CFO would, and paying it down against the use cases actually queued for delivery, not an abstract maturity model.
It’s More than Data Models – Make Data Quality an AI Operating Model
The instinct in most data offices is to hire more DQ analysts when demand rises. That doesn’t scale, and increasingly it doesn’t need to. Treat quality as a domain AI can help operate, not just a domain AI depends on:
- Design-time governance, quality and lineage requirements embedded before a pipeline goes go-live, implemented & monitored by AI, not audited after.
- Continuous monitoring, catching drift and anomalies in near real time, not at the next scheduled review.
- Shared ownership, in our One Data Lifecycle model, the CDO owns structure, the CPO owns legitimate use, the CISO owns resilience. Quality sits at the intersection and fails the moment any one function treats it as someone else’s job.
- Human-in-the-loop, not human-in-every-loop: route the ambiguous cases to a person, not every record, let the DQ AI Operating Model .
The organisations moving fastest here don’t have the biggest DQ teams. They’ve engineered trust into the architecture metadata, controls, ownership instead of assuring it after the fact with manual sign-off.
The organisation succeeding in the future will implement AI heavy Data Management Operating Models, that manage all data real-time as business processes and AI Agents consume – true AI Enabled Data Management at scale.
Regulators Are Turning AI-Ready-Data Into an Evidential Requirement
The EU AI Act requires data governance controls for high-risk systems, and that obligation doesn’t stop with model builders, deployers carry direct responsibility for input data quality too. You can’t outsource it to your vendor.
The UK’s principles-based regime arrives at the same place through the ICO, FCA and MHRA: you can’t demonstrate fairness or contestability in a decision you can’t trace. In the US, state and sector regulators are increasingly treating poor data governance as evidence of inadequate AI risk management.
Sitting underneath all three: ISO/IEC 42001:2023, the first certifiable AI management system standard. Three clauses do the work – ongoing risk assessment as conditions change (Clause 6), impact assessment when deployment context shifts (Clause 8), and demonstrable top-management commitment (Clause 5).
Read against the Zillow case, those three clauses map almost exactly onto the three gaps that produced an $881m loss. A useful test: if your AI governance couldn’t pass a Clause 6 or 8 audit today, that’s the gap to close first.
The figures behind it are already real, not theoretical: US regulators issued roughly $4.3bn in financial penalties in 2024, banks accounting for 82% of that. Global AML fines rose 417% in the first half of 2025 alone.
The Bank of England’s 2024 AI survey found four of the top five perceived AI risks at UK institutions were data-related. Behind almost all of it sits a data quality symptom that was visible internally long before it reached a regulator.
What This Means from Monday Morning
For the CDO, COO & Chief AI Officer: stop treating data management and AI governance as adjacent workstreams.
Fund and sequence them as one programme, your regulators and your own agents will treat them that way regardless of your org chart.
For the Head of Data Architecture: the data model you maintain is no longer a build artefact for engineers.
It’s the reasoning substrate every AI Agent in the enterprise will query. Semantic clarity and named ownership aren’t documentation overhead they’re the accuracy ceiling for everything built on top.
A practical starting checklist ✏️
☐ Run an AI-readiness gate assessment against your top three use cases before the next funding cycle, not after.
☐ Find the one data domain where governance debt and AI ambition are both highest. Start there, not everywhere.
☐ Stand up a semantic layer and business glossary that’s machine-readable & actively maintained, not a confluence site, set of stale artefacts or poorly implemented data management tools.
☐ Map lineage end-to-end for one high-risk use case. Could you produce it for a regulator today?
☐ Name an accountable owner per critical domain, and check the name would survive a board meeting.
Where Quaylogic comes in
This is the work we do all day, every day: closing exactly this gap for COOs, Chief AI Officers, CDOs, Heads of Data Architecture and Data Governance leads from AI-readiness assessments and data lineage delivery, through to embedding the Trust Continuum operating model across data, privacy and security.
If you’re scoping what “AI-ready” actually means for your data estate, we’d like to hear where you’re stuck.
Get in touch with our team and let’s find the gap before your next funding cycle does. 🚀
About the Author
JAGTAR BACHRA, Managing Partner at Quaylogic
Jag co-founded Quaylogic with a clear vision: to simplify data management delivery and turn data into measurable business value. He has led some of the largest and most complex data programmes in Financial Services, consistently demonstrating the ability to align data with business strategy and deliver sustainable growth.
With 20 years of experience and a blue-chip consulting background, Jag has worked with senior executives up to CxO level, specialising in the full data management journey — from strategy through to hands-on delivery and embedding change into BAU. His track record is defined by measurable outcomes, commercial impact, and lasting transformation for clients.
What We Get Asked (FAQs)
Why do most AI projects actually fail?
Not because the model is weak; because the data underneath it is. Independent research puts data-related causes behind 70–85% of failed AI and ML initiatives. Poor completeness, inconsistent definitions, and missing lineage get encoded into the system’s behaviour at scale, long before anyone blames the model.
What is an AI-readiness data gate?
It’s a pass/fail check run before a use case is funded, not a clean-up exercise run after a pilot stalls. It tests completeness, definitional consistency, lineage, currency and named ownership. Organisations that treat this as a gate consistently outperform those that treat data quality as a parallel-track clean-up.
What’s the difference between data quality debt and governance debt?
Data quality debt is about the records: accuracy, completeness, freshness. Governance debt is structural: undefined ownership, unmapped lineage, definitions that disagree across systems. AI doesn’t create either problem, but it’s the valuation event that forces both onto the balance sheet.
Why does a semantic layer matter for AI accuracy?
Because an agent querying a raw, undocumented schema has to guess what a column means. Benchmark work on text-to-SQL generation shows accuracy jumping from around 40% on unmodelled schemas to 83–90%+ once a well-modelled semantic layer with clear business definitions sits in front of it.
Does ISO/IEC 42001 replace the EU AI Act or UK principles?
No — it complements them. The EU AI Act, UK principles and US patchwork define what outcome regulators want. ISO 42001 defines how to structure the operating model to get there, and it’s the one you can actually be audited against, through clauses covering risk assessment, impact assessment and top-management commitment.
Who’s responsible for AI data quality: the vendor or the organisation using the model?
The organisation deploying it. Under the EU AI Act, deployers carry direct responsibility for input data quality, performance monitoring and incident reporting, regardless of who built the underlying model. You can’t outsource that obligation to your AI vendor.
Where should a CDO start if governance debt feels overwhelming?
Don’t restart the whole governance programme. Quantify the debt like a CFO would, find the single data domain where governance debt and AI ambition are both highest, and pay it down against the use cases actually queued for delivery, not against an abstract maturity model.
