A model that knows nothing about your company will still answer your question. It will tell you the work gets faster and cheaper, which is the average of everything it has read, and nothing in the output will mark it as generic. That reversion to the mean is the real cost of data that is technically clean and semantically silent. Atlan closes it by carrying certified definitions, ownership, lineage and access policy into one Enterprise Data Graph that agents read at answer time, so the reply is specific to your business rather than to the industry you happen to be in.
| Area | What actually decides readiness |
|---|---|
| The unit of assessment | One question, one agent, one use case. Not the warehouse |
| The failure that hurts most | A fluent, plausible, company-agnostic answer that nobody flags |
| What clean data does not supply | Agreed meaning, entity relationships, provenance, access scope at runtime |
| The split nobody names | Training readiness and answering readiness want different things from the same table |
| The decay rate | Published measurements show analytics accuracy falling roughly 30 points in a month without maintenance |
Why clean data is not the same as AI-ready data
Data quality work has a clear definition of done. Nulls filled, duplicates collapsed, formats standardised, outliers investigated. When the job finishes, a data engineer can point at a table and say it is clean, and be right.
Gartner puts the limit of that work bluntly: “‘High-quality’ data, as judged by traditional data quality standards, does not equate to AI-ready data.” Quality is a property of the table. Readiness is a property of the relationship between a table and a question. The same warehouse can be ready for “what was net revenue last quarter” and not ready for “which accounts are at churn risk,” because the second needs a definition of churn nobody has certified and a join path nobody has approved.
That is why data quality for AI agents is a necessary condition and not a sufficient one. Every published definition now carries the same caveat, inherited from the same source: readiness is relative to use. Almost none of them explain what that changes about the work.
What changes when the consumer is an agent
A human analyst handed an ambiguous request asks a follow-up. An agent does not. It resolves the business concept to a physical field on its own, picks from hundreds of plausible candidates in a wide schema, writes valid SQL, and returns a number.
Nothing in that sequence fails loudly. The query runs. The number formats. The chart renders. If the agent resolved “active customer” to the wrong column, the only signal is that the figure looks slightly off to someone who already knew the answer, which defeats the point of asking.
Spider 2.0, presented at ICLR 2025, put a number on how hard that resolution is: 632 real enterprise text-to-SQL workflows against databases averaging 812 columns, on which a code agent built on o1-preview solved 21.3%, against 91.2% on the older academic benchmark. The gap is almost entirely the gap between a toy schema and a real one carrying decades of undocumented meaning.
The failure mode with a name: reversion to the mean
Ask a general-purpose model a company-specific question without company-specific context and you get the company-agnostic answer. Ask what the business value of a new capability is and it will say faster and cheaper, because that is what the internet says about every capability. Fluent, defensible, true of nobody in particular.
This is not hallucination in the usual sense. Nothing is invented. The model answers from the only context it has, which is everyone else’s. The valuable thing your business knows never entered the conversation, so it could not come out of it.
Anthropic published the clearest public measurement of the difference. Its internal analytics agent, asked questions against the company’s own data without skills and semantic grounding, did not exceed 21% accuracy on their evaluation set. With that context supplied, accuracy ran consistently above 95% in aggregate and around 99% in some domains. The model did not change between those numbers. What changed was whether it had been told what the company means by the things it was asked about.
Why nobody flags the average answer
A wrong number gets caught when someone recognises it. A generic answer gets nodded at, because it is not wrong. It is useless, and uselessness has no error code.
That is what makes the failure expensive at scale. Teams roll out an assistant, the answers read plausibly, adoption looks fine, and the thing never produces an insight anyone could not have got from a trade publication. The gap surfaces when someone compares the AI’s answer to what the business actually knows, by which point the deployment has a budget and a roadmap. Context-aware retrieval is the difference between a system answering from your world model and one answering from the median of everyone’s.
The unit of readiness is the question, not the table
Gartner’s public framing is unambiguous: “There is no way to make data AI-ready in general or in advance. The readiness of data for AI depends on how the data will be used.”
Most organizations read that sentence and run a warehouse-wide readiness programme anyway, because a warehouse is something you can staff a project against and a question is not. The result is months of work that improves the average and leaves the workflows the business cares about as broken as they were.
The more useful opening question is not whether the rows and columns are right. It is: what questions do we want our people to answer with AI, and what has to be true for those answers to be consistent, accurate and useful? Every readiness activity then has somewhere to be scored. Work that moves no named question is housekeeping, and it can wait.
Training readiness and answering readiness are different states
The distinction has no agreed vocabulary anywhere in the field, which is part of why it gets missed.
A training-readiness checklist tells you to remove outliers, deduplicate and balance. An answering-readiness checklist tells you to preserve the meaning, owner, freshness stamp and access scope of every field an agent might reach for. On outliers the two point in opposite directions: the anomaly a cleaning pass deletes is often the pattern a model needs to see, which is why Gartner’s definition insists representative data “may include poor-quality data, too.”
One public framework makes the split concrete, scoring 62 measurable requirements across five workload profiles. Training carries 50 of them, agents carry 37, and they are different sets rather than a subset. Most readiness content in circulation applies training-era logic to a reader asking an answering-era question. If your consumers are agents, knowledge architecture for AI agents matters more than corpus hygiene.
The five layers that decide whether data is AI-ready
Readiness is not one property. It is five, and a failure in any one of them produces a different flavour of wrong answer.
Semantics: what a number means here
A column named revenue tells an agent almost nothing. Does it include intercompany eliminations? Recognised or billed? Which of the four tables carrying that name does finance sign off on?
A semantic layer answers those questions in a form a machine can read, and the measured difference is large. dbt Labs ran the same questions through raw text-to-SQL and through a semantic layer in 2026: Claude Sonnet 4.6 scored 90.0% against 98.2%, GPT-5.3 Codex 84.1% against 100.0%. The more consequential finding was behavioural. Text-to-SQL failed silently and returned plausible wrong answers; the semantic layer failed loudly and refused rather than guessing.
That is the property you want. Not that it never fails, but that when it fails you find out. Systems of semantics and data contracts for AI are how that gets written down rather than remembered.
Ontology: how entities relate here
Semantics fixes what one thing means. Ontology fixes how things connect, which is what any question spanning more than one concept depends on. “Which enterprise accounts renewed late after a support escalation” is three entities and two relationships, and each has to resolve the same way twice for the answer to hold.
Researchers at data.world benchmarked GPT-4 answering questions directly against an enterprise SQL schema at 16% accuracy, and the same questions over a knowledge-graph representation of the same data at 54%. The underlying data did not improve between those runs. The relationships became explicit.
The distinction between ontology and a semantic layer matters here, as does the difference between a context graph and a knowledge graph. A static model of relationships helps; an active ontology that updates as the estate changes helps considerably more.
The context layer: what your company knows that nobody else does
This is the anti-reversion layer, holding what never fits in a schema: the instructions, working procedures, business knowledge and documentation that make an answer specific to you. Why the EMEA numbers restate every January. Which customer segment definition the board uses and which one the product team uses. What “at risk” means here rather than in the category.
Most of it lives in people’s heads, which is why tribal knowledge and institutional knowledge loss are AI problems and not just HR ones. A context layer for AI agents is where that knowledge becomes readable by a system, and context engineering is the discipline of deciding what reaches the model for a given question.
Gartner’s April 2026 guidance tells data and analytics leaders to “make the context layer the central brain for AI agents to deliver trusted intelligence,” which is the same observation reached independently.
The harness: your SQL and BI estate is part of the answer
There is a persistent assumption that AI replaces the analytics stack. In practice the stack is the machinery the answer runs on. The BI semantic models, the certified dashboards and the metric definitions already agreed are part of the system that turns a question into a number, and text-to-SQL for enterprise works far better inheriting that estate than improvising around it.
Readiness work that ignores what exists produces a parallel definition of revenue and a fresh argument about which one is right. Retrieval orchestration and protocols like MCP exist so an agent can call the governed thing instead of rebuilding it.
Deterministic assets and procedure graphs
The fifth layer is the one almost nobody writes about, and it has two halves.
Analytical logic you can read. When a model writes Python to produce a number, you get the number and lose the reasoning. The same logic in a notebook is an asset: versioned, reviewable, and callable by an agent before any interpretation happens. The shape that works is an agent calling a skill, the skill calling a deterministic asset for the data, and the model entering only at the point of explanation. Determinism is what makes an answer survive a second run, and that matters more than it sounds. Thinking Machines Lab ran 1,000 identical completions of an open model at temperature zero and got 80 distinct outputs, all agreeing through token 102 and diverging at token 103.
The procedure, not just the skill. Teaching an agent to do a task is not the same as giving it the procedure: the ordered steps, the decision points, the checks, and what to do when one fails. A procedure graph encodes how the work is actually done, retries included, so the system corrects itself rather than stopping at the first broken link. No published readiness checklist asks for it.
Both halves are what make decision traces possible, and traces are what turn one correct answer into something auditable.
What the standard definition gets right, and where it stops
Gartner’s public definition is the load-bearing source for almost every page on this subject, and it earns the position. Proving readiness is “a process and a practice based on availability of metadata to align, qualify and govern the data,” and one of the five recommended steps is to evolve metadata from passive to active, which describes exactly what active metadata management does.
Three things follow that most readers never get told.
If readiness is proved from metadata, it is measurable rather than declarable. A survey answer is not evidence. A metadata layer for AI that shows coverage, ownership and freshness is.
Readiness also expires. Gartner is explicit that AI-ready data is “not one and done,” and the decay is measured: Anthropic watched offline accuracy fall from roughly 95% to roughly 65% over a single month when the supporting skills went unmaintained. Context freshness is a readiness property.
And the boundary keeps moving. In September 2026 Gartner pushed it past the data entirely, arguing that AI-ready data “also requires AI-ready stakeholders who understand the value of trusted data.” Their route there is culture and literacy. The engineering route is to put the knowledge those stakeholders hold somewhere a system can reach.
How you can tell whether readiness is real
Self-assessment is unreliable here in a measurable way. A March 2026 survey of senior data and AI executives by Monte Carlo with CDO Magazine found 71% of organizations preparing to be AI-ready and 29% with a documented definition of what that means. The same survey found 61% reporting silent failures, where monitored metrics looked normal throughout a critical issue.
Three tests cut through that, and none needs an opinion.
Ask the same question twice. On unchanged data, count the distinct answers. The pass mark is one. Nothing in the published readiness literature asks for this, and it is the fastest way to surface a definition two systems disagree about.
Ask it as two different people. Run the identical question as two principals with different entitlements. The results should differ correctly, and the agent should appear in the audit log under its own identity rather than a shared service account. That is where agent access control stops being a policy document and becomes observable.
Ask for the reasoning. For any answer given last month, you should be able to produce the question, the definition it resolved to, the exact query, the policies that applied and the data version. If the logic lived inside a model call, none of that exists.
The AI readiness assessment page turns these into dimensions and scoring; the AI-ready data checklist turns them into items you can fail. If you already know the gaps, how to prepare enterprise data for AI agents is the build path.
Where this leaves the readiness programme
Gartner predicted that through 2026, organizations would abandon 60% of AI projects unsupported by AI-ready data, and the same research found 63% of organizations either lacking or unsure of the right data management practices, from a survey of 1,203 data management leaders. A separate 2026 study with Harvard Business Review Analytic Services put the share of enterprises calling their data completely ready for AI at 7%.
Those numbers usually get read as an argument for a bigger programme. Read alongside the definition, they argue for a narrower one. If readiness cannot be established in advance, a two-year estate-wide effort optimises something that was never the unit of measurement while the five questions the executive team actually asks stay unanswerable.
The organizations getting useful answers picked a question, made the meaning, relationships, context, machinery and procedures behind it explicit, checked that the answer reproduced, and did it again for the next one. That is slower to announce and faster to trust. What it leaves open is which questions are worth the work, and that is not a data decision.
FAQs about AI-ready data
1. Is clean data the same as AI-ready data?
No. Gartner states plainly that high-quality data, judged by traditional data quality standards, does not equate to AI-ready data. Cleaning removes duplicates, fills nulls and strips outliers. Readiness adds meaning, relationships, provenance and access scope, against a specific use rather than in the abstract.
2. Can we make all of our data AI-ready in advance?
No. Gartner’s position is that there is no way to make data AI-ready in general or in advance, because readiness depends on how the data will be used. Pick the questions the business actually asks and make the data behind those ready first.
3. What is the difference between AI readiness and AI-ready data?
AI readiness usually describes the organization: strategy, funding, skills and governance maturity. AI-ready data describes a property of data relative to a use case. The two are routinely conflated, and an organization can score well on one while failing the other.
4. Does AI-ready mean the same thing for model training and for AI agents?
No, and the two can pull in opposite directions. Training wants representative data, including outliers a cleaning pass would remove. An agent answering at runtime wants the meaning, owner, freshness stamp and access scope of every field it might reach for. One public framework scores 50 requirements for a training workload and 37 for an agent workload, and they are different sets.
5. How long does AI-ready data stay ready?
Not indefinitely. Gartner describes AI-ready data as a practice rather than a one-and-done project. Anthropic measured the decay: offline accuracy of its internal analytics agent drifted from around 95 percent at launch to around 65 percent over a month when the supporting context went unmaintained.
6. What role does metadata play in making data AI-ready?
Metadata is what readiness is proved with. Gartner describes proving AI readiness as a practice based on the availability of metadata to align, qualify and govern the data, and names evolving metadata from passive to active as one of the five steps. Without it there is nothing to inspect, so readiness can only be asserted.
Sources
-
Lack of AI-Ready Data Puts AI Projects at Risk, Gartner (26 February 2025)
-
How Anthropic Enables Self-Service Data Analytics With Claude, Anthropic (3 June 2026)
-
Semantic Layer vs Text-to-SQL: 2026 Benchmark, dbt Labs (7 April 2026)
-
Spider 2.0: Evaluating Language Models on Real-World Enterprise Text-to-SQL Workflows, ICLR 2025
-
Defeating Nondeterminism in LLM Inference, Thinking Machines Lab (10 September 2025)
-
The State of AI Reliability, Monte Carlo with CDO Magazine (March 2026)
