The method · Astérisme

Each process described, tested, validated, measured.

AI validated before it is used. The method comes from computerised system validation in the biopharmaceutical sector: a system is documented before it is built, tested on real cases, validated by the people who use it, then reviewed. Applied to AI in a business, it comes down to one process sheet per process.

We work on site, with your teams: we install the AI, test it on your own files and stay until it is up and running.

One process sheet per process, nothing connected without being written down.

The process sheet says what the AI does, with which data, who reviews and what has changed. It also names the provider and the model, chosen with you. It lives in your tools, in open format.

Example process sheet, for illustration only.

FP-03 · v1.2 Validated on 14.10.2026

Reminders for unpaid invoices

Objective: Send reminders at D+7 and D+21, never chasing a client who has already paid.
Inputs: Due dates from the invoicing software, payment history.
Handed over to AI: Draft the reminder from the template and the client file.
Human check: The head of accounting validates. No automatic sending after the second reminder.
Excluded data: Bank details, ongoing disputes.
AI provider: Mistral Small, open-weight model, hosted in Belgium on a server dedicated to the engagement.
Acceptance threshold: No reminder to a client who has already paid. Fewer than 2% substantive corrections, demonstrated by zero substantive corrections across a hundred and fifty files. Set in writing before the test.
Test set (kept apart from the build): A hundred and fifty files from three real working weeks, never used to build the process.
Confidence threshold: If the amount or payment status is in doubt, the AI flags it, proposes no reminder, and the head of accounting decides.
Drift monitoring: Share of corrected reminders tracked every month. Each quarter, twenty reminders already sent are re-read blind by a second person. Above 3% corrections, the sheet is reviewed before its scheduled date.
Signatories: Process owner (the head of accounting) and lead user (an accounts clerk).
Log: v1.0 Created
v1.1 Softer tone for clients of more than five years
v1.2 Added the ‘part payment’ case

Inherited from the biopharmaceutical sector.

Write before you build

A written specification comes before each process: objective, inputs, what the AI does, what it never does.

Test on your real files

A documented test, with expected and actual results, over real working weeks. The error rate is measured and recorded in the process sheet. The acceptable error threshold is set in writing before the test. Test files are not used to build the process. When the AI is not sure enough of an answer, it says so and a person decides. The number of files tested follows the threshold: the more demanding it is, the larger the test set.

Have your team validate it

Your business lead signs off the validation, not us. Whoever builds it does not validate their own work.

Version everything

Each process sheet has a version and a change log.

Review regularly

An AI model can change its behaviour without warning. Each process sheet has its next review date. After go-live, results are monitored: if they drop, the sheet is reviewed before its scheduled date. The correction rate is not enough: over time, reviewers correct less. A sample is therefore re-read blind every quarter by another person.

Keep or stop

After go-live, the gain is measured on your own figures. A process that brings nothing is stopped, and this is written in the process sheet.

For each process, six steps.

Assessment

Data, professional rules, hosting, signatories named.

Approved process sheet

Error threshold and stop criterion signed before anything is built.

Verified configuration

Model, instructions and connections in line with the sheet.

Test on your files

Results measured against the threshold, signed by your team.

Go-live

Team trained, first review dated.

Follow-up

Kept, corrected or stopped at each review.

The number of signatories follows the size of your team. The full course of the engagement.

What stays with you, signed.

The register of authorised tools

Every AI tool used in the business, with its use, its data and its validation date. A tool missing from the register is not authorised.

File

One validation file per process

The process sheet, the cases tested, expected and actual results, the error rate, and your business lead's signature.

Incident

The incident procedure

Who alerts whom, how the error is corrected and logged, when the process sheet is reviewed.

Your business software already has its own AI assistant? We test it on your files with the same method, and it goes into the register if it passes validation.

Three possible places. The choice is made process by process.

Public AI

Claude, ChatGPT, Mistral, Gemini: the most powerful models, with nothing to install.

Data: Sent to the provider, through a business plan whose terms we read with you.
Cost: Pay as you go.
For: Texts with no sensitive data: adverts, job ads, market watch.

Open models, on a dedicated server in Belgium

Open-weight models, such as Mistral, Qwen or Gemma, installed on a server we manage.

Data: Stays in Belgium. It does not pass through any major AI provider.
Cost: Monthly flat fee, on quotation.
For: Client correspondence, minutes, internal documents.

Local, on your premises

A dedicated server installed in your offices, after an audit of your needs and your infrastructure.

Data: Never leaves the company.
Cost: Hardware purchase, then maintenance.
For: Confidential files, professional secrecy, health data.

The same firm can write its adverts with a public AI and keep its debt recovery files local. Each process sheet states where the AI runs, and why.