What data readiness for AI means for one use case
Data readiness for AI is not a state your whole company reaches. Your data is ready for one specific task when every source that task depends on passes five checks: the tool can reach it, you are allowed to use it this way and only the right people can see it, it is current, it is structured enough for the task, and it is correct with a named owner. Run the check per source before you buy or build, and end each one with a decision: ready, fix first, leave out, or stop and ask whether you may use it at all.
For most small and mid-size businesses, the data is not in a warehouse. It lives in an inbox, a CRM, a price list in a spreadsheet, procedures in a shared drive, and the bookkeeping system.
Academic work, such as a 2024 survey by Hiniduma, Byna, and Bez (opens in a new tab), catalogs ways to evaluate data used to train models. Most small businesses are not training models; they are pointing a purchased tool at records they already keep, which calls for a narrower check.
Why bad data produces confident wrong answers
NIST's AI Risk Management Framework (opens in a new tab) lists data problems among the risks that set AI apart from traditional software. Its Appendix B notes that data used to build an AI system may not be a true or appropriate representation of the intended use, and that training datasets can become stale relative to where the system is deployed. Those points concern training data, but the same problem applies when a purchased tool reads your documents. NIST notes that the framework is being revised (opens in a new tab) as of October 2026.
A tool that reads an outdated price list will quote the outdated price as confidently as a current one. NIST's Generative AI Profile (opens in a new tab) calls confidently stated but false content "confabulation." Good data does not eliminate that risk, but bad data hands the tool wrong facts to repeat.
Start from the task and list its inputs
Write the task in one sentence, then list every source it reads and what the output needs from each. For a tool that drafts replies to quote requests: the request itself, the price list, the rules for where you work and what you won't take on, examples of good past replies, and the schedule. Anything the task does not read is out of scope, which keeps the check to an afternoon. Do it with the person who does the task and whoever manages the systems; between them they usually know where each fact really lives. If you have not chosen the task yet, the scoring sheet in AI use case prioritization ranks candidates, with data readiness as one of six factors.
The five-question AI data readiness checklist
Run each source through these five questions; notes on each follow the table.
| Question | What to check | Ready looks like | Not ready looks like | First fix |
|---|---|---|---|---|
| Reachable | Can the tool get to it through an export, connector, or API, in a format it can read? | Lives in a system with export or a supported connector | Lives in someone's head, on paper, or in scans nobody can search | Export it, or write it down in one place |
| Permissioned | Are you allowed to use it this way, and will the tool show it only to people who should see it? | The use fits what customers and contracts were told; sharing settings are deliberate | Customer data collected under promises that did not cover this use; a shared drive open to everyone | Check privacy promises and contracts; fix sharing before connecting |
| Current | When was it last updated, and who updates it? | Updated on a known schedule, with a date on it | Several versions, no dates | Pick one source of truth and date it |
| Structured enough | Does the task need consistent fields, or can it work from documents? | Consistent wording for what the task uses: service names, units, areas | The same thing written five ways; key facts buried in long notes | Standardize only the fields this task uses |
| Correct and owned | Does someone own its accuracy, and how often is it wrong? | A named owner and a way to report errors | Nobody knows who maintains it | Name an owner; spot-check a sample |
Reachable
If a fact lives only in a person's memory, no tool can use it. That is common for the rules that matter most: which jobs you decline, which neighborhoods you serve, when you charge a trip fee. Writing those rules down is often the most valuable readiness work, with or without AI.
Permissioned
Permission has two meanings here. Inside the business, tools that read your files generally follow the access settings you already have. Microsoft's Copilot documentation (opens in a new tab), for example, says Copilot only surfaces organizational data that the user has at least view permission for, and advises using permission models to control access. A tool that respects permissions will still surface whatever your sharing settings over-expose, so tidy the shared drive before you connect it.
Outside the business, the question is whether you may use the data this way at all. FTC technology staff wrote in February 2024 (opens in a new tab) that a company adopting more permissive data practices, such as using consumer data for AI training, through a quiet retroactive change to its terms or privacy policy may be acting unfairly or deceptively. Before you point a tool at customer records, read what your privacy policy and client contracts promised, and what the vendor's terms say it does with the data; the AI vendor evaluation checklist lists the documents to ask for. Whether you may is a question for counsel; this article is not legal advice.
Current
A price list with no date, or three versions in three folders, is not ready. Choose one source of truth, date it, and put a review on the calendar.
Structured enough
Most first uses do not need a database. They need consistency in the few fields the task depends on. If one service appears under five names, or prices are sometimes per hour and sometimes per job without saying which, standardize those fields and leave the rest alone.
Correct and owned
Each source needs a person responsible for its accuracy and a way for staff to report errors. Spot-check a sample of 20 records or pages. If you find errors you cannot explain, fix the process that creates them before you add a tool that repeats them.
Turn the answers into a decision for each source
End every source with one of four decisions, checked in this order:
| Decision | When it applies |
|---|---|
| Stop and ask | You cannot say whether you may use the source this way. Do not connect it until someone with authority, and counsel where needed, has answered. |
| Fix first | It fails a question, and the task cannot produce a usable draft without it. |
| Leave out | It fails a question, and the first version can work without it. |
| Ready | It passes all five questions. Include it only if the first version actually needs it. |
Fewer, cleaner sources make a first version easier to review and to trust.
Worked example: a quote-drafting assistant
This example is hypothetical. A home-services company with eight staff wants an assistant to draft first replies to quote requests from its website form, with the office manager reviewing every draft before sending. Copy the table, one row per source, and write a reason in each cell rather than a bare yes or no.
| Source | Reachable | Permissioned | Current | Structured enough | Correct and owned | Decision |
|---|---|---|---|---|---|---|
| Web form submissions | Yes | Yes, collected to answer the request | Yes | Yes, form fields | Yes | Ready |
| Price list spreadsheet | Yes | Yes | No, last edited over a year ago | Yes | Partly, edited by several people with no review | Fix first: update, date, name an owner |
| Service-area and job rules | No, in the owner's head | Yes | Unknown | No | No | Fix first: write a one-page rule set |
| Past quote emails | Yes, by mailbox export | Partly, they contain addresses and private notes | Mixed, older prices | No, free text | No | Leave out of the first version |
| Crew calendar | Yes | Yes | Yes | Yes | Yes, the dispatcher | Ready, but left out: first replies will not offer appointment times |
| [Your source] |
The result is a narrower first version: the form submission, the updated price list, and the written rules, with a person approving every draft. Past quotes stay out until someone selects and cleans a set of good examples. The check runs again whenever a source changes owner or format.
Signs to pause before you start
Pause if you cannot name the task, if the key facts exist only in people's heads and nobody has time to write them down, or if you cannot tell whether customer data may be used this way. Pause, too, if staff already paste this data into unapproved tools: the task matters, and shadow AI explains how to give it an approved path instead of a ban. A one-page AI acceptable use policy is where that path gets written down. Projects that skip these questions tend to stall when the system meets real inputs, the pattern described in why AI pilots stall.
How to get data AI-ready without a data project
If nothing above says pause, work through the sources marked Fix first, starting from the First fix column in the checklist. Three rules keep it from turning into a data project:
- Fix only what the first use case needs. Readiness for one task does not require readiness for every task.
- Leave a source out rather than delay the whole project for it.
- Test with real inputs, including deliberately bad ones. How to red-team an AI system covers data poisoning and other tests worth running before launch.
A readiness and governance gap assessment is part of the AI Adoption Assessment that starts Cleland & Co.'s AI Strategic Advisor engagement. If you already know the use case, the five questions above are the place to begin.
