Data

Preparing your data before an AI project: the checklist

The quality of an agent's results depends first on the quality of the data you give it. A short checklist to go through before you start.

You don't need a data warehouse to benefit from AI. A well-prepared export is often enough. Here are the points to check before handing a file to an agent.

Structure

  • A single header row, with explicit column names.
  • One piece of information per column (no “amount + currency” in the same cell).
  • No subtotal rows or merged cells in the middle of the data.

Consistency

  • Dates in a single format.
  • Amounts without currency symbols, with the same decimal separator.
  • Customer or supplier names written the same way from one row to the next.

Completeness

Spot the columns that are often empty and decide whether they are needed for the mission. An agent can flag missing values, but it must not invent them: it's better to know in advance what's missing.

Privacy

Remove anything that isn't useful for the mission, especially personal information: social insurance numbers, dates of birth, personal contact details. The principle is simple: only share what is necessary.

An example of a well-prepared mission

“Here is the sales export from July to September (one row per order: date, customer, region, product, amount before taxes). Calculate growth by region compared with the previous quarter and flag the abnormal weeks.” The mission specifies the content of the file, the expected calculation and the deliverable.

All resources

Automation 2 min read

Automating invoice checks: a five-step method

Duplicates, unusual amounts, miscalculated taxes: a simple method for handing the first check of supplier invoices to an agent, without losing control.

AI agents 2 min read

AI agent or chatbot: what changes for an SMB

An assistant answers a question. An agent takes on a task from start to finish, and is accountable for it. The difference is measured in hours recovered.

Business Intelligence 2 min read

Useful KPIs rather than decorative dashboards

A good dashboard fits on one screen, answers specific questions and triggers decisions. How to choose the right indicators and keep them alive.