AI data readiness: how to assess it before you fund AI

A utility provider took a £2.7 billion capital request to its regulator and had it rejected, because its own data could not show how it had spent its money. AI data readiness is the same test, set by software instead of a regulator: can an AI system find the data a task needs, know what it means, trust that it is current, and keep it from the wrong people?

The test matters more now that AI takes actions, not only drafts. In October 2026, NVIDIA, IBM and KPMG all published work built on one idea: an AI agent, software that uses an AI model to plan a task and then act on it, is only as good as the company context it reads. In a KPMG survey, 58% of leaders named data readiness and access as one of their biggest obstacles to deploying agents.

The answer is not a multi-year clean-up: assess one workflow at a time, fix what it needs, then reuse the fix.

What is AI data readiness?

AI data readiness means the data one AI task needs can be found, understood, trusted and controlled by software, with no person filling the gaps. It is a property of a use, not of a whole company. The same customer table can be ready for a sales summary and unready for an agent that issues refunds.

Most company data was built for people. A finance analyst knows things the data does not say: that the sales system's revenue figure, say, ignores returns while the ledger's does not, or which spreadsheet is the real one. None of that knowledge sits next to the data, so an AI system never sees it.

Documents need the most care. Contracts, policies, emails and call notes are much of what AI assistants actually read, and IBM cites an estimate that less than 1% of such unstructured data is in a format AI can use directly. Three faults recur:

• Old versions that look current. Last year's travel policy sits beside this year's, and nothing tells the assistant which one applies.
• Duplicates across shared drives. The assistant quotes a copy that someone edited and never filed back.
• Scanned files. A signed contract held only as an image stays invisible until its text is extracted, while the unsigned draft is easy to find.

One analogy helps. An AI model is like a very able new hire on their first day: quick, well read, and unaware of how your firm works. The analogy breaks after day one. A new hire asks questions and learns for years. A model starts each task with only what the system hands it. Whatever your people know but never wrote down, the model does not know.

That is why readiness depends on the job. Snowflake's open AI-ready data framework uses separate profiles for an assistant that searches documents, for agents that query systems and act, and for training a model. It calls the agent profile the broadest and most demanding: it spans all six of its tests, with strict thresholds.

What changed: AI agents now act on company data

Until recently most enterprise AI drafted text for a person to check. Agents now query systems and carry out steps, so a data fault becomes a wrong action instead of a wrong draft.

The shift shows in three places.

• Work. KPMG's Global Tech Report predicts that digital assistants will account for 36% of core technology teams' capacity by 2027, up from 28% in 2025.
• Data work itself. On the Databricks platform, over 70% of pipelines built with its Spark Declarative Pipelines framework are now created by agents. A pipeline is the code that moves and reshapes data between systems. Databricks says this raises the stakes: every attempt an agent makes costs compute and time, and code review becomes a critical gate, a control someone has to staff and pay for.
• Products. NVIDIA's expanded Agent Toolkit includes an "auto-ontology" library that translates raw company data into business concepts such as customers, products and orders, and maps how they relate. An ontology is that map: the concepts your business uses and where the data for each one sits. IBM's data business argues for governing data close to where it is created, then feeding it to models while it is still current.

The common thread is context. Vendors now compete to own the map of your business, which matters for lock-in later.

The problem itself is not new. In July 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, and poor data quality came first in its list of causes. What changed is the price of getting it wrong.

How an AI system reads your data, step by step

Whether an employee asks an assistant a question or an agent starts a task, the same six steps run, and each one can fail on data.

1. The request becomes business concepts. "What is our best-selling product in mid-market?" turns into product, sales and customer segment. KPMG notes that the answer hinges on whether best-selling means units, revenue or margin. If nobody has written that down, the system guesses.
2. The system finds where each concept lives. The ontology points to the tables and documents that hold product, sales and segment data, often across several systems. Data outside the map does not exist for the AI.
3. It checks who is asking. Microsoft states that its Copilot assistant only surfaces company data that the user already has at least view permission for. The AI follows your access rules, so it is only as safe as they are.
4. It retrieves the relevant rows or passages. This pattern is called retrieval-augmented generation, or RAG: the system looks up the right material first, then hands it to the model with the question. Stale or duplicate records are handed over as if they were true.
5. The model writes the answer or takes the action. It cannot tell a wrong input from a right one. Snowflake's framework puts it plainly: when a model receives stale data, it produces a confident, wrong answer.
6. The system records what it used. This record is called lineage: where each fact came from and what was done to it on the way. Without it, nobody can explain a bad answer or prove a good one.

Many failures that look like a model problem start in steps one to four. A better model does not fix a missing definition or an open folder.

The AI data readiness framework: six tests

A usable assessment framework applies six tests to every data source a task needs. The names come from Snowflake's open framework. The tests work on any platform, and the checklist further down turns each one into a question for your team.

Score each test for each source as pass, partial or fail. A fail on clean or compliant should stop the project until it is fixed. A partial on the other four can often go ahead, provided a person reviews the output until the gap closes.

What weak data readiness costs in money, people, risk and scale

Weak data costs in four ways: AI projects that stall, skilled people who check data by hand, exposure when an AI system shows data to the wrong person, and use cases that cannot share the work.

Stalled projects. Gartner predicts that through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data. That is a share of the projects that lack ready data, not of all AI projects, and it is a forecast, not a count. IBM's 2025 CEO study found that just 16% of AI initiatives have reached enterprise scale. Two survey figures show where the gap sits, and what each means for the board:

• Practices are missing. In a 2024 Gartner survey of data management leaders, 63% of organisations either lacked the right data management practices for AI or were not sure they had them. For the board: ask who owns data for AI before asking which model to buy.
• Confidence is low. In a 2024 IBM survey, only 29% of technology leaders strongly agreed their data met the quality, access and security standards needed to scale generative AI. For the board: treat any AI plan that assumes clean data as unproven until that data has been tested.

Treat vendor numbers with care: IBM and KPMG both sell data services. The independent diginomica study, which interviewed 18 senior practitioners with no vendor funding, found only three with anything resembling a live agentic AI system. The rest described pilots that had not proven value and initiatives that were stalling.

People. In the same study, participants estimated that between 30% and 70% of professional time goes to manual data assembly, reconciliation and verification instead of analysis or decisions. Treat that as a range from interviews, not a measured benchmark. One utilities CIO described losing more than a thousand person days a year to reconciliation alone. Asked what share of their data they would pass to their chief executive without checking it first, participants said, in effect, none.

Risk. IBM's 2026 Cost of a Data Breach Report puts the global average cost of a data breach at US$4.99 million, and that is the average across all breaches, not the cost of AI oversharing alone. Under the EU AI Act, penalties can reach €35 million or 7% of a company's worldwide annual turnover, but that top tier is for banned AI practices, not for a data fault in a permitted system. Data protection and AI rules often ask firms to document where data came from and whether it was fit for purpose, which is lineage by another name.

Scale. Every AI use case that builds its own pipeline, its own definitions and its own access rules repeats the cost. Give the board one line to track: the cost and the weeks it takes to launch each new use case. The tenth should come in cheaper and faster than the first; if it does not, the data foundations are not being shared.

What people get wrong about AI-ready data

Four beliefs waste time and money on AI data readiness.

"Our reports work, so our data is ready"

Reports have a person between the data and the decision. That person corrects for known gaps without thinking. KPMG's guidance on AI-ready data is direct: data that works for reports and dashboards can create problems for AI, because definitions differ across the business and the rules for using data are stored apart from the data. Test the data against the AI task, not against last quarter's board pack.

"We must clean all our data first"

A whole-estate clean-up tends to become a multi-year programme with no payback until the end, and the target moves while you work. Readiness is set by the task, which is why Snowflake's framework scores data against a profile for each kind of workload. KPMG calls AI-ready data a no-regret investment "as long as the work is tied to business value". Clean what the first workflow needs, then what the next one needs.

"We need one central copy of everything first"

Moving all data into one platform before any AI use is slow and often unnecessary. Shaun Clowes, general manager of Confluent and data at IBM, argues for preparing usable data close to where it is produced: "The earlier you govern it, the better." NVIDIA says its new knowledge tools work without moving data or changing governance controls. Both are vendor claims. Test them on your own systems before you plan around them.

"The AI follows our permissions, so we are safe"

It does follow them, and that is the risk. Microsoft's guidance tells customers to use their permission models so the right people have the right access, because Copilot can surface anything a user has permission to view. If a finance folder is open to the whole company, a person may never stumble on the salary file inside it. An assistant that searches everything a user can open can find it in seconds. Review access on the sources in scope before launch, not after.

How to run an AI data readiness assessment

Assess one workflow, not the whole estate. A focused assessment ends in a clear decision: go, fix first, or stop.

1. Pick one decision with a known cost. Devavrat Shah, chief scientist of enterprise AI at Celonis, gives the same advice in an MIT Technology Review Insights podcast made with Celonis: start with a specific business decision, find its hardest steps and test whether AI improves them. Choose work where you can count hours, errors or delay today.
2. List every source the workflow touches. Systems, documents, spreadsheets and the people who correct them. The spreadsheets matter most, because they show where the official data is not trusted.
3. Score each source on the six tests with the people who use it every day, not only the data team.
4. Decide how the AI gets access. KPMG asks whether an agent inherits the access of the person using it or receives its own permissions. For agents that act, give them their own identity with the least access the task needs.
5. Measure today's manual cost. Count the hours spent finding, reconciling and checking data for this workflow. This is your baseline and most of your business case.
6. Set "good enough" for this use. Be strict where the AI acts without review, and looser where a person checks every output.
7. Decide: go, fix first or stop. Write down the fix cost, who owns it and what it saves. Then measure return at the level of the workflow, with full cost and outcome in view, as InfoWorld's pilot-to-production framework recommends. Stop is a valid answer: if the data fix costs more than the workflow can save, pick another workflow.

AI data readiness checklist: questions for your team and vendors

Use these questions in a steering meeting or a vendor pitch. A good team answers each one in a sentence. A vague answer marks a gap to close before the build starts.

• Which fields and documents does this task need, and how many are blank, duplicated or out of date?
• Can the system fetch that data in the time the task allows?
• Who owns the definition of each business term this AI will use, and where is it written down?
• How old can each source be before the answer is wrong, and what enforces that limit?
• Does the agent act with the user's access or its own, and who approved that access?
• Can we trace any answer back to the records it used, on the same day?
• What happens to the AI when a source system renames or changes a field?
• Where do copies, summaries and search indexes of our data sit, who owns them, and when are they deleted?
• If we change vendor, do our definitions and data maps come with us in a format we can read?

The risks, and the control for each

Each data risk below has a specific control you can require before launch.

• Oversharing. The AI shows a user data they can open but should not see. Control: review access on every source in scope, label confidential content, and give agents their own narrow identity.
• Stale data. The AI acts on last week's prices or stock. Control: set a freshness limit for each source, monitor it, and tell the AI to stop or flag when the limit is passed.
• Definition drift. A team changes what "active customer" means and the AI keeps the old meaning. Control: give each definition a named owner, and review and log every change.
• No trail. Nobody can explain a wrong answer to a customer, an auditor or a regulator. Control: log every source behind every answer or action, and keep the log as long as the decision must be defended.
• Copies spreading. Search indexes and summaries hold your data too. Control: give each one an owner, a refresh cycle and a deletion date, like any other data asset.
• Lock-in on the context layer. NVIDIA, Palantir, Databricks and IBM all now offer ways to describe your business to AI. Control: keep definitions and maps exportable, and write portability into the contract.
• Urgency from sellers. The diginomica researchers note that much of the research urging firms to move faster is funded by the people selling the platforms. Control: let your own pilot data, not a survey, set the pace.

Where Thriiv fits

Thriiv builds and runs custom AI systems, workflow automation and secure data pipelines for international enterprises. We find where the operating cost sits and put a quantified saving on the table before budget is committed. A working pilot follows in weeks, not quarters, and we host and run what we ship.

What we build runs in a private environment that we host and manage. Our page on how we host and run systems sets out what that covers. For the controls an agent needs once its data is ready, see agentic AI risk management.

What to do this quarter

Five actions put AI data readiness on a footing a board can track.

• Name one workflow with a measurable cost and run the six-test assessment on it.
• Ask for an access review of every source an AI assistant can already search, starting with finance, HR and legal content.
• Appoint a named owner for the ten business terms your AI use cases rely on most.
• Measure the hours your teams spend checking data before they trust it, and report it as a cost line.
• Make portability of definitions, data maps and logs a condition in every AI and data contract.

Readiness is not a state you reach once. It is a cost you manage, one workflow at a time. If you want a second view on one workflow, tell us where the cost sits.