Short answer

Governing AI in a small organisation does not require a committee or a fifty-page document. It requires an honest table: for each tool, what the AI does, what it never does, where it runs and which person reviews what it proposes. That table clarifies more than any generic policy. It shows that in some tools the AI only transcribes, in others it prepares a record that someone validates and in others, deliberately, there is no AI.

Key points

  • Governing AI in a small organisation requires an honest table: for each tool, what the AI does, what it never does, where it runs and who reviews it.
  • Writing down what the AI never does, such as scoring applications or replying to a client, stops its scope from growing without anyone deciding it.
  • Whatever is handed to an external model leaves the organisation, and removing the name is not anonymisation: sector, size, dates and job titles also identify people.
  • Every model output should arrive flagged as a proposal, with its evidence, until a person accepts, corrects or rejects it, with who and when recorded.
  • Sometimes the right answer is not to use AI, as with a whistleblowing channel, where trust rests on a person handling the case.

Four columns that change everything

The first column describes the specific task: transcribing a CV into a record, extracting the purpose, deadline and amount from a call for proposals, proposing a classification for a document. The narrower the task, the easier it is to review and to replace.

The second sets out what the AI never does: score applications, decide whether the organisation enters a tender, reply to a client, promise results. Writing this down stops the scope from growing without anyone deciding it. The other two columns, where it runs and who reviews, turn the table into a risk map.

Where it runs matters as much as what it does

Whatever is handed to an external model leaves the organisation. That is why it is worth distinguishing between a model running on the organisation’s own hardware, one running in a vendor’s account under contract and an external one that only receives public text, such as that of an official call for proposals. A layer of your own between the application and the model makes it possible to switch vendor without rewriting the product, and logging each call with its cost tells you how much it costs and when it fails.

Removing the name is not anonymisation. Sector, size, dates and job titles identify someone just as surely as a surname. Minimising what is sent and aggregating data with minimum thresholds offers more protection than superficial deletion.

Flag the proposal until someone reviews it

Everything that comes out of a model should arrive flagged as a proposal, with its evidence alongside: the fragment of the document from which an amount was extracted, the paragraph a date came from. The flag disappears when a person accepts, corrects or rejects it, and who did so and when is recorded.

If the output does not match the expected format, a retry is advisable and, if it fails again, a manual review queue. A system that falls back to a person is more reliable than one that improvises.

When the right answer is not to use AI

A whistleblowing channel should not depend on a model for anything: the trust of the person making a report rests on a human being handling the case. A tool that reconstructs billable hours from the calendar can classify with rules and keywords, which are predictable and auditable.

Recruitment tools deserve special attention: the EU AI Act classifies systems that filter or evaluate applications as high-risk. Designing the tool so that the AI only transcribes, without scoring, ranking or extracting age, sex or nationality, is a design decision that must be documented and does not replace each organisation’s own legal analysis.

Frequently asked questions

How do I document AI use across my organisation's tools without setting up a committee?

With a four-column table for each tool: the specific task the AI performs, what it never does, where it runs and which person reviews what it proposes. That table clarifies more than a generic policy and works as a risk map. It shows, for example, that in some tools the AI only transcribes, in others it prepares a record that someone validates and in others, deliberately, there is no AI.

Is removing names enough before sending data to an external AI model?

No. Removing the name is not anonymisation: sector, size, dates and job titles identify someone just as surely as a surname. Minimising what is sent and aggregating data with minimum thresholds offers more protection than superficial deletion. It also matters whether the model runs on the organisation's own hardware, in a vendor's account under contract or in an external service that only receives public text.

Can a recruitment tool use AI to filter job applications?

This case requires special attention: the EU AI Act classifies systems that filter or evaluate applications as high-risk. One possible design decision is for the AI only to transcribe, without scoring, ranking or extracting age, sex or nationality. That decision must be documented and does not replace each organisation's own legal analysis.

What should the system do when the AI returns output in the wrong format?

Retry once and, if it fails again, send the case to a manual review queue. A system that falls back to a person is more reliable than one that improvises. In any case, everything a model produces should arrive flagged as a proposal until a person accepts, corrects or rejects it.