Is Your Data Ready for AI? A Practical Check Before You Point AI at Your Numbers

Ask AI for last quarter's margin by service line and it will give you a clean, confident figure. Ask your finance team and you will get a different one. In a 2026 survey of 270 U.S. CFOs and finance leaders, 65% named exactly this, confident answers built on the wrong data, as their most common frustration with AI.¹ The cause is rarely the AI. It is whether your numbers agree with each other.

That is what data readiness for AI means: whether your business data is fit for the specific question you are asking AI to answer. This article gives you five checks you can run this week, shows what they look like in professional services, construction and finance, and describes the smallest fix that works. If you would rather talk it through, a free 30-minute Strategy Call with an AI adoption specialist ends with the first two or three workflows worth automating.

01. When AI gives a confident wrong answer

The scene you may already know. Someone on your team asks an AI tool for the margin on a service line. The answer comes back in seconds, formatted well, with no hedging. It does not match the report finance sent last week.

Neither number is a mistake in the usual sense. Each one is correct for the data it was given, and the two sets of data disagree about what "margin" includes, which month is closed, or which customers belong to the line. The AI did not create that disagreement. It hid it behind a fluent sentence.

It is common, and it costs time.

Datarails, a finance software company, surveyed 270 CFOs and finance leaders in July 2026, at U.S. organizations with more than 1,000 employees and annual revenue of $100 million or more. Teams reported spending 26% of their working week verifying or correcting AI output, 96% said at least 10% of their time goes to that checking, and 56% said team members had been given materially different outputs from the same prompt and the same data.¹ That is the survey's own sample, so read it as a signal, not a benchmark for your business.

The signal is still useful. Among teams whose top operational challenges were manual reporting and data consolidation, 86% reported confident answers based on the wrong data, against 65% of all respondents.¹ The messier the underlying numbers, the more often AI sounded sure and was wrong.

What "ready" is measured against. The international standard on data quality for analytics and machine learning, ISO/IEC 5259-1, frames data quality as something you assess and improve across the data life cycle, and as a question of whether data is fit for its intended purpose.² Data is not simply good or bad. It is fit or unfit for a particular question. Your customer list can be perfectly good for sending invoices and unfit for answering "which customers are active?" if three systems define "active" three ways.

What follows. The rest of this article gives you a way to check, in order: five questions you can ask about the data behind one AI question, what those questions turn up in three kinds of company, the smallest fix that works, and a habit for checking any AI answer against numbers you already trust.

02. What "data ready for AI" actually means

Ready is about a question, not a company.

Data readiness for AI means the data behind one question is accurate enough, defined clearly enough, current enough and permitted enough that an AI answer built on it is something a person can rely on after a quick check. Gartner puts the test bluntly: "if the data has issues, then the data is not ready for AI." It also stresses that AI-ready data is not "one and done", meaning readiness has to be maintained as data and use cases change.³ You can be ready for one question this month and not ready for the next.

It is not a warehouse project, and it is not a clean-up of everything.

"Data readiness" can sound like a long program: a new platform, a migration, a team of analysts. That is a different thing.

ISO/IEC 25012, the long-standing data quality model, describes fifteen quality characteristics viewed from two angles, inherent to the data itself and dependent on the system that holds it, and it notes that their importance differs between stakeholders.⁴ You do not need all fifteen perfect. You need the few that matter for the question in front of you.

How wide the gap is. Gartner reports that 63% of organizations either do not have, or are unsure they have, the right data management practices for AI, and predicts that through 2026 organizations will abandon 60% of AI projects that are not supported by AI-ready data. That second figure is a prediction, not a survey result.³

Closer to home, RSM's 2026 survey of 827 U.S. and 203 Canadian middle market executives found 34% naming data quality as the top inhibitor to deploying AI, ahead of security and privacy (30%) and legacy system integration (28%). Respondents could pick up to three answers, and RSM sells AI readiness assessments.⁵

Precisely, which sells data integrity software, surveyed more than 500 senior data and analytics leaders at large enterprises. Of those, 88% said they have the data readiness AI requires, yet 43% also named data readiness as one of their biggest AI challenges.⁶ Confidence and reality do not line up.

The same pattern shows up in the work. In our introduction to AI adoption we described a reporting project that surfaced 655 site records where about 412 were expected. The reporting did not cause the problem. It made years of duplicated records impossible to ignore, and any AI answering from those records would have inherited the error and sounded confident doing it.

That article gave the lesson in one paragraph. This one gives you the checks.

03. Five checks you can run this week

Pick one question your team asks AI today, or wants to ask: the margin on a service line, open receivables by customer, the status of jobs against budget. Run the five checks below against the data behind that one question. None needs a new tool.

Check 1: Is it the same number everywhere? Pull one figure from three places your people actually use, for example last month's revenue from the accounting system, the monthly report and a spreadsheet someone keeps. If the three disagree, AI may pick one and state it as fact. It is also the quickest check to run.

In the Datarails survey, just 4% of CFOs said they have a single source of truth for finance and operational data, and 23% rely on disconnected systems and manual reconciliation.¹ If your three numbers disagree, you have plenty of company.

Check 2: Who owns the definition? Choose three terms your business uses every week, such as "active customer", "billable hour" or "backlog". For each, name one person who decides what it means, and write the definition in a sentence. If two people give different answers, AI will inherit both.

Precisely's survey points the same direction: 71% of organizations with data governance programs reported high trust in their data, against 50% of those without one, and 29% said the biggest obstacle to high-quality data was measuring data quality in the first place.⁶ An owner and a written definition are the smallest form of governance, and they do not need a committee.

Check 3: Can you trace it to its source? Take one number in a report and walk it back to the transactions behind it. If that takes an afternoon, nobody can check an AI answer built on the same number.

Fragmented systems are why this is slow. Intuit's 2026 survey of 725 U.S. accounting and bookkeeping professionals found that only 41% say their tools are fully integrated, and that accountants lose an average of five hours a week moving, re-entering or reconciling data across disconnected systems.⁷ The sample is accountants in the United States and Intuit sells accounting software, but the mechanism is general: every manual hop between systems is a place for a number to change.

Check 4: Is it current, and does it say as of when? A figure with no date is a trap. AI will happily report last quarter's backlog as today's. Ask of any number: when was this last updated, and does the report say so?

Gartner's point that AI-ready data is not "one and done" applies directly here, because stale data is wrong data with a delay.³ If a report has no "as of" stamp, add one before pointing AI at it.

Check 5: Who may see it? Before AI reads a data set, decide who is allowed to see it and what must not leave your environment. Canada's federal, provincial and territorial privacy authorities jointly published principles for generative AI in December 2023. They include being open and transparent about how information is used, building safeguards for privacy rights, and limiting the sharing of personal, sensitive or confidential information.⁸ In practice that means a short list: which fields contain personal or confidential information, which of them AI may read, and which must stay out. It is also why we connect AI to company systems read-only first, with a dedicated account limited to the records the question needs.

What a pass looks like. You pass when you can answer all five for one question.

• Three systems agree on the number, or one is named as the source.
• A person owns each definition.
• You can trace the figure to transactions in reasonable time.
• The data carries an "as of" date.
• You know who may see it.

Do not build a score or a maturity model out of this. One question that passes is worth more than a company-wide rating, and it gives you the template for the next question.

04. What this looks like in your kind of company

The five checks fail in recognizable ways depending on the business. These are patterns, not a measure of your company.

Professional services. Hours sit in a time tracking tool, project status sits in another, and the invoice sits in the accounting system. Ask AI for realization by client and the answer depends on which of those it reads and how "billable" is defined. Accounting firms are one example of a professional services business, and Intuit's survey found 30% naming manual data cleanup as the top blocker to doing more advisory work, ahead of staffing shortages at 24%.⁷

If that is your week, the place to start is the checks above, on one client report. See AI for professional services.

Construction and specialty contractors. Job cost lives in the accounting system, quantities and hours come from the field, and change orders live in email and phone calls. A 2026 survey of more than 1,000 construction leaders by ServiceTitan, which sells contractor software, found 38% reporting measurable business impact from AI, and only 20% of contractors operating on a single platform.⁹

With data split across platforms, an AI estimate or job-status answer is only as good as whichever source it reached first. Start with one question, such as committed cost against budget on one job. See AI for contractors.

Finance teams. If your month-end numbers are assembled from several systems, each with its own formats and mappings, an AI summary will inherit every mismatch. KPMG's 2026 global survey of 1,013 senior finance leaders in 20 countries found 36% citing data quality as both their top barrier and their top opportunity. Technology and financial services organizations make up 58% of that sample, so adoption figures in the report skew toward sectors further along.¹⁰

Gartner's May 2026 finance research, from a March survey of 204 finance leaders, found 63% of finance organizations saying AI implementation was slower than expected in 2025, and recommended that CFOs strengthen data, talent, process and governance foundations.¹¹ Start with one report the board sees every month. See AI for finance teams.

What the three share. In all of them, the fix is the same size: one question, one agreed source, one owner and a date. That is the next section.

05. The smallest fix that works

Fix the data for one question, not the whole company.

Choose the question your team asks AI most often, or the one you most want to ask. Fix only what that question touches. You will learn more from making one answer trustworthy than from a company-wide assessment, and the work is small enough to finish.

Four moves. All four change how people work before they change any technology, and you can start all four this week.

1. Name the owner: one person accountable for the question and the numbers behind it.
2. Write the definitions: a sentence each for the two or three terms the question depends on.
3. Point the question at the one agreed source, and tell the team that is where AI reads from for this question.
4. Add an "as of" date to the data and to every answer.

Why scope matters more than size. Real enterprise data is harder for AI than tidy test data. Spider 2.0, a peer-reviewed academic benchmark presented at ICLR 2025, tested AI agents on 632 real-world problems. The problems required answering questions by querying enterprise databases, many with more than 1,000 columns. The authors' agent, built on a leading AI model of the time, solved only 21.3% of the tasks, against 91.2% on the earlier Spider 1.0 benchmark and 73.0% on BIRD, another established benchmark.¹²

Spider 2.0 is a research benchmark, not a result for a business like yours. A reasonable reading is that the wider and messier the data AI has to search, the harder the answer is to trust, and that a narrow question on a clean, agreed source is the safer place to start.

Who does what. Three roles cover it. In a smaller leadership team one person may hold all three, and that works too.

• The owner or CEO picks the question, because it should be one that matters to the business.
• The COO or head of operations owns the process behind it: where the data is entered, by whom and how often.
• The CFO or controller owns the definitions and the check against trusted numbers, since finance is often where the trusted numbers live.

How this fits our service. Data readiness for AI is one of the services in the Plan stage of our work, described on our data readiness for AI page. It starts the way this section does, with one question and one source, and a person at your company stays accountable for the result. Connections to your systems start read-only and every answer is checked against numbers you already trust before anyone relies on it, which is the subject of the next section. Where AI drafts or produces something, a person signs off before it goes anywhere.

06. Checking an AI answer against numbers you trust

Test the data, not the model. Before you rely on an AI answer, check it against one number you already trust: a bank balance, a closed month, a signed contract total. If the AI cannot reproduce a figure you know, you have learned something useful about the data, and you learned it before the answer reached a decision.

Three tests you can run today.

1. Reproduce a known figure: ask for something you already have a trusted answer to, such as last closed month's revenue, and compare.
2. Ask the same question twice, in different words, and compare the answers. In the Datarails survey, 56% of respondents reported team members getting materially different outputs from the same prompt and data, so a difference between two runs is a signal, not noise.¹
3. Ask the AI to show the records or documents behind the figure, then open one and check it. An answer that cannot point to its source is not an answer you can sign off.

When to stop trusting. If two of the three tests fail, treat it as a data problem, not a model problem. Stop, go back to the five checks, and fix the failing one before asking again. Asking more questions of data that disagrees with itself produces more confident wrong answers, not fewer.

A person stays accountable.

None of this removes the need for human sign-off. NIST's AI Risk Management Framework calls for policies that define roles and responsibilities for human and AI oversight, and for human oversight processes that are defined, assessed and documented.¹³ For you, that means a named person who looks at the answer, runs the check and takes responsibility before anything goes to a client, a lender or a board. The check is short. Skipping it is how a wrong number ends up in front of someone who trusted it.

Where this fits in the series. This is the second article in our AI adoption series. The first, our introduction to AI adoption, explains Plan, Build and Run.

Where to read next. Why we connect AI read-only first explains how connections to your systems start, and what is checked before anyone relies on an answer.

This content may be AI-generated and may contain inaccuracies. It is not professional, legal, or financial advice. Verify independently before relying on it.