How to Prepare Your Business Data for AI

Data preparation for AI is workflow-scoped, not company-wide. For each workflow you want to automate, you need three things: access (the AI system can reach the data programmatically), consistency (the data does not contradict itself in ways that change the outcome), and safety (personal and sensitive data is minimized and controlled).

That is a much smaller job than the "fix all our data first" project many companies assume they need - and which frequently stalls AI adoption for a year. Prepare the data for one pilot, ship the pilot, and let each subsequent use case pull the next piece of data work behind it.

A workflow's data is ready when a capable new employee could do the task using only that data - without asking a colleague for missing context. That is a practical test worth taking literally: if a person would need to ask "which of these two spreadsheets is current?" or "what does status 3 mean?", an AI system has the same problem, and unlike the employee it will not always know to ask.

Readiness is therefore about resolving ambiguity, not achieving perfection. Typos, gaps, and noise are usually tolerable; contradictions, undocumented codes, and unknown currency of information are not.

For the workflow you want to automate, list every place its information lives: business systems (CRM, ERP, help desk), shared documents and spreadsheets, email threads and chat conversations, PDFs and scans, and - the one every company has - knowledge that exists only in someone's head.

For each source, note two things: can it be exported or accessed through an API, and who controls access. The undocumented-knowledge category matters most: if the task depends on rules nobody wrote down, capturing those rules in a short reference document is data preparation, and it is often the single highest-value hour of the whole project.

Good enough for the pilot, judged by outcomes: run the workflow's real cases through the data and see whether the answer is derivable. If an AI drafting quotes needs price lists, the current price list must be identifiable as current - the retired ones can stay in the archive as long as they are distinguishable.

Two rules keep this proportionate. First, fix forward: establish a clean way of recording things from today, rather than repairing years of history that the workflow may never touch. Second, clean by exception: run the pilot, watch where it fails, and fix the specific data those failures point at. Failure-driven cleanup is dramatically cheaper than speculative cleanup.

Modern AI handles most digital document formats directly, so conversion effort is smaller than commonly feared. The practical priorities are: prefer digital originals over photos of paper; scan remaining paper at readable quality; use consistent, descriptive file names and folder structure so retrieval can find the right document; and designate one authoritative location per document type instead of copies scattered across drives and inboxes.

Structure inside documents helps too. Templates with labeled fields, consistent headings, and dated versions all raise extraction accuracy - and they cost nothing when adopted for documents created from now on.

An AI workflow usually needs to read from at least one system and often write to another. Check three levels, in order of preference: a documented API, a scheduled export (CSV or report), or - as a last resort - the vendor's own AI or integration marketplace. If a critical system offers none of these, that is important to learn during preparation, not mid-build; it may change which workflow you pilot first.

Also confirm permissions early: who can approve API credentials or a new integration, and does your plan tier include it. Waiting two weeks for an access approval is a common and entirely avoidable pilot delay.

If the workflow touches personal data - customers, employees, applicants - design the treatment before the build. The working principles under Brazil's LGPD are: minimize (send the AI system only the fields the task needs, not the whole record), redact identifiers where the task works without them, control access to sensitive stores and log who reads them, and know where data goes when an external AI service processes it.

Also verify your AI provider's terms on training: reputable providers offer business tiers that contractually exclude your data from model training - confirm this in writing rather than assuming it. This section describes common engineering practice, not legal advice; your counsel or DPO should validate the specifics for your data.

Day 1-2: map the sources for your chosen workflow and mark each as API, export, or manual. Day 3: take ten real, recent cases of the task and check whether the recorded data alone answers them - note every gap and every ambiguity. Day 4: capture the undocumented rules in a one-to-two page reference document, reviewed by the person who knows them. Day 5: classify what is personal or sensitive and write down the minimization plan.

The output is a short readiness memo: what is available, what is ambiguous, what must be fixed before the pilot, and what can be fixed forward. That memo turns "our data is a mess" - a feeling - into a work list, which is something you can actually finish.

Three patterns reliably waste months. Building the platform first: starting a data warehouse or lake project before any use case exists to justify or shape it. Perfection paralysis: refusing to pilot until historical data is clean, when the pilot itself is the best tool for finding what needs cleaning. And ignoring the human layer: automating on top of data whose meaning lives in one person's head, without capturing it first.

The common thread is sequencing. Data work done in service of a specific, shipping workflow compounds; data work done speculatively usually has to be redone when the real requirements arrive.

ConsultatechAI runs data readiness as part of its opportunity assessment: mapping sources, testing real cases against the recorded data, documenting the undocumented rules, and defining the LGPD-aware minimization plan for each candidate workflow - in Portuguese and English.

The deliverable is the readiness memo and a scoped preparation plan per pilot: what to fix now, what to fix forward, and what to leave alone. It is deliberately the opposite of a platform pitch - the goal is the smallest data effort that makes the first workflows trustworthy.