“Is our data ready for AI?” has no answer until someone says what for. The same data can support lookups of internal procedures perfectly well and be useless for forecasting next quarter’s demand. Readiness is not a grade the company gets once: it is an assessment made for a specific task, with specific sources, for specific people.
Start with the task, not the technology
The first conversation should not be about models, platforms or vendors. It should describe a real task:
- what is done today (answering a question about a procedure, checking fields on a document, explaining why a metric moved);
- who does it, how often and how long it takes;
- where that person looks for the information;
- what happens when the answer is wrong, and who notices.
A well-described task already rules out a lot of ideas. If nobody can say where the correct information lives, AI will not find it either. If a mistake has serious consequences and there is no way to check the answer, the project needs human review from the start, or it should not begin.
Sources, quality, owners and permissions
With the task defined, list the sources behind it and assess each one on four points.
Source. Where does the information live: the ERP database, a Power BI model, a document folder, the ticketing system, spreadsheets? Is there an official version, or does each department keep its own copy? Are there old documents that contradict the current ones?
Quality. Is the data complete for the period that matters? Are the fields the task relies on filled in consistently? Do the documents contain searchable text, or are they scanned images that would need text recognition first? For metrics, is there a written definition of the calculation rule?
Owner. Who is accountable for the content and decides when it changes? A source with no owner goes stale without anyone noticing, and any answer built on it goes stale too.
Permissions. Who can see what today? If finance sees data that sales does not, an assistant that draws on both sources has to enforce the same separation for every user. Access granted to “everyone” on a network share for convenience becomes a leak once an assistant starts answering from it.
Measure the current process first
Without a baseline, there is no way to tell later whether AI helped. Before any pilot, record how the task works today:
- average time to complete the task or get the answer;
- how often the answer has to be corrected or redone;
- what the current process costs, including people’s time;
- a set of real cases with the correct outcome, validated by someone who knows the subject.
That set of cases is what you will later use to evaluate the pilot. Split it in two: one part used during development and another held back for the final evaluation, so the test is not run on the same questions used to tune the solution.
When better data or BI solves it without AI
A good readiness assessment sometimes concludes that AI is not the next step. A few signs:
- the information exists, but it is scattered across spreadsheets that nobody reconciles. A consolidated database and a dashboard solve it;
- the same question keeps coming up because the procedure is outdated or poorly written. Revising the document solves it;
- the team asks for “natural language questions,” but the questions are always the same five. A report with filters solves it, at a lower operating cost;
- the numbers disagree across reports. Until the rule is defined, any assistant will simply repeat the disagreement with more confidence.
These improvements are not wasted effort: they build the foundation an AI project will need anyway.
Be careful with built-in AI features in data platforms. Copilot in Power BI and Fabric data agents, for example, require paid capacity and administrator configuration, which in some regions includes allowing data to be processed outside the capacity’s region. The requirements change often; check Microsoft’s current documentation before counting on them in a plan.
Readiness checklist for a task
- The task is described: who does it, how often, how long it takes and what a mistake costs.
- The sources behind the task are listed, with the official version identified.
- Each source has an owner accountable for its content.
- Quality has been checked on the fields and the period the task uses.
- Current permissions are mapped and make sense for the people who will use the solution.
- The personal data involved is identified, with purpose and legal basis reviewed by the legal team.
- There is a baseline for time, rework and cost of the current process.
- There is a set of cases with correct outcomes, validated by experts and split between development and evaluation.
- Someone has assessed whether a data, document or BI improvement would solve it without AI.
- It is clear who reviews the answers at the start and what stops the operation if quality drops.
Unanswered items do not block the conversation, but they show what needs to happen before the pilot.
Personal data and what not to send to models
Brazil’s data protection law (LGPD) still applies when data passes through an AI model. The principles of purpose limitation and necessity apply: use only the personal data the task actually needs. Sensitive data, such as health, biometrics and racial or ethnic origin, is subject to stricter rules. If the service processes data outside Brazil, that is an international transfer, which has its own requirements. Decisions about legal basis and transfers belong to the legal team and the data protection officer; the technical side needs to know exactly which data flows and where. A solid personal data mapping is the starting point.
A few practical rules:
- never send passwords, access keys, tokens or connection strings to a model;
- do not paste internal documents or customer data into personal-use tools that lack a corporate agreement covering retention, use for training and processing location;
- prefer aggregated or pseudonymized data when the task does not depend on who the person is;
- record which sources feed the solution and who approved each one, so a source can be removed later.
Want to assess a use case?
If you have a candidate task and the question is whether your data can support it, an assessment maps the sources, permissions and baseline, and compares AI with simpler alternatives. Learn about our enterprise data and AI service or request an assessment.