Every AI tool, from a forecasting model to an agent that reads invoices, is limited by the records you feed it. Before you spend on software, it is worth an honest hour on data readiness for AI: is the information the AI would need complete, consistent, reachable and safe to use? This checklist lets you answer that yourself, department by department, without a data scientist in the room.
Work through it with the person who runs the department you have in mind. Tick each line only if you could prove it to an outsider.
Why data readiness for AI decides the outcome
Analysts have been blunt about this. In early 2025, Gartner reported that 63% of organizations either did not have, or were unsure whether they had, the right data management practices for AI, and predicted that organizations would abandon 60% of AI projects unsupported by AI-ready data.
Poor data costs money even without AI. Gartner research has put the average cost of poor data quality at at least $12.9 million a year for the organizations it studied. AI tends to make the problem visible faster, because a model trained on messy records produces confident answers that are wrong.
The good news: you rarely need perfect data. You need the right data, for one specific job, in good enough shape. That is what the checklist tests.
Before you start: pick one job
Data is only "ready" for something. Write down the single task you want AI to help with, such as forecasting next quarter's demand, matching supplier invoices or answering customer questions about order status. Then run the checklist against that task alone.
The job changes what counts as ready. A demand forecast leans on clean sales history by product and week. An invoice-matching agent needs purchase orders, delivery notes and supplier records that line up. A customer-question assistant needs current policies and product details written down somewhere it can read. Same company, three very different readiness answers.
Part 1: Does the data exist?
The records this task needs are captured in a system, not only in people's heads or paper files.
There is enough history to show patterns. For a forecast, that usually means several full cycles of your busy and quiet seasons.
The outcome you care about is recorded. If you want to predict late payments, your records show when each invoice was actually paid, not just when it was due.
Important events are written down: price changes, promotions, stock-outs, system switches. These explain jumps in the numbers.
Part 2: Is it accurate and consistent?
The same customer, product or supplier appears once, not three times under slightly different names.
Units, currencies and date formats are consistent, or the differences are documented.
Required fields are actually filled in. Spot-check 50 recent records and count the blanks.
Someone has corrected known errors at the source, rather than in a spreadsheet copy that nobody else sees.
Two departments asked the same question (how much did we sell last month?) would give the same answer.
Part 3: Can you get to it?
You know which systems hold the data: ERP, CRM, accounting, shared drives, inboxes.
Data can be exported or read through an API, without a vendor charging per request or refusing access.
Records in different systems can be linked, for example by a shared customer number or order number.
Documents the AI needs to read, such as contracts or emails, are stored digitally and can be searched, not only as scanned images.
Part 4: Is it safe and lawful to use?
You know which fields contain personal data about customers, staff or patients.
You have a clear legal basis for using that personal data for this new purpose.
You plan to use only the fields the task needs. The UK Information Commissioner's Office, in its guidance on AI and data protection, treats data minimization and accuracy as core principles that apply to AI as to any other processing.
Access is limited by role, and you can see who looked at what.
Any AI service you use has training on your data switched off, and your records stay in an account you control.
Part 5: Does someone own it?
Each key data set has a named owner who decides how it is defined and fixes it when it is wrong.
Key terms are written down. "Active customer" or "on-time delivery" means one thing across the business.
There is a routine for keeping data clean after the project ends, not a one-off tidy-up.
Someone will check whether the AI's output still makes sense as your business changes. The US NIST AI Risk Management Framework describes this as a continuous cycle of governing, mapping, measuring and managing AI risk across the life of the system, rather than a single sign-off.
Score your data readiness
Count your ticks across all five parts. There are 22 lines in total.
Ticks | What it usually means | Sensible next step |
|---|---|---|
18 to 22 | Your data is in good shape for this task | Run a small pilot on real records and measure the result |
12 to 17 | Usable, with specific gaps | Fix the gaps that block this one task, then pilot. Do not try to clean everything first |
6 to 11 | The task is likely to struggle | Start with the data foundation: one system of record, agreed definitions, an owner |
0 to 5 | Not ready for this task yet | Pick a different task with better data, or build the system that captures it first |
Pay most attention to Parts 1 and 4. Missing history cannot be invented, and a privacy problem can stop a project outright. Gaps in Parts 2, 3 and 5 take effort, but they can usually be fixed while the project moves forward.
Common gaps and what to do about each
"Our numbers live in spreadsheets"
Very common, and not fatal. Find the spreadsheet that people actually trust, agree that it is the source, and move it into a proper system over time. If the records really belong in an ERP or a custom tool, that build often comes before the AI. Our custom software development work frequently starts here.
"Our history is short or patchy"
Some tasks need less history than others. An agent that reads incoming invoices relies on the documents themselves more than on years of records. A demand forecast needs history, so start capturing it properly now and pick a different first project.
"Nobody agrees what the numbers mean"
Get the department heads in a room and write the definitions down. It is unglamorous work, and it often fixes reporting arguments long before any AI arrives.
"Our data is spread across too many systems"
Draw a simple map: which system holds customers, which holds orders, which holds invoices, and how they link. Often a shared order or customer number is enough to join them. Where nothing links, a small integration that copies records into one place each night may be all the first project needs. That kind of plumbing is a natural fit for workflow automation.
"We're worried about where the data goes"
That is the right worry. Ask any supplier where your data is stored, whether it is used to train their models, and who can see it. Your records can stay in your own cloud account while the AI works on them.
From checklist to forecast
Once the data passes for a task, the payoff can be quick. Predictive analytics built on clean history gives you forecasts for demand, cash or capacity with a confidence range, in the dashboards your leadership team already reads. For document-heavy tasks, AI agents can read, check and draft, with a person approving what goes out.
Related reading: AI Readiness Checklist: 10 Questions Before You Invest · How to Scope an AI Pilot Project That Proves Its Value
Get a second pair of eyes
If you would rather not score yourself, our free 5-day AI readiness audit looks at one department with you. You get a plain verdict on whether your data is ready, the three tasks where AI would save the most hours, and a fixed-price plan for the first one that is yours to keep.
