Prepare a workplace AI use before deployment

Start from a real need, match safeguards to the level of risk and organise a limited pilot before any decision to scale.

In 30 seconds

Essential points

  • A useful pilot begins with a bounded, measurable and easily verifiable task.
  • Vigilance increases with sensitive data, system autonomy and the consequences of error.
  • Human oversight, stop criteria and post-deployment indicators must be defined before the first test.
Decision filter

Does this AI use deserve a pilot?

Technology is not a need in itself. These questions help determine whether the proposed pilot addresses an identifiable work problem.

Which problem are we trying to solve?
Describe an observable work difficulty without starting from a tool that has already been selected.
Which task will be assisted?
Identify a precise activity, its steps, the people involved and the expected result.
Which data will be used?
Review their sensitivity, origin, hosting, access rights and possible reuse.
Who will verify the output?
Name a person with the time, competence, information and authority required.
How will usefulness be measured?
Track quality, error, time, verification workload and effects on real work.
How can the pilot be stopped?
Define fallback arrangements, stop criteria and the reversibility of the process.
Three levels of vigilance

Match safeguards to the real workplace use

The level depends on the task, data, system autonomy and people affected — not only on the model or product name.

Level 1 · Lower risk

Document assistance

Reading, summarising, classifying or reformulating non-sensitive content without a direct effect on an individual decision.

  • verify sources;
  • exclude sensitive data from uncontrolled tools;
  • review before reuse.
View examples
  • evidence scanning;
  • reformulating training material;
  • structuring a bibliography;
  • preparing a document outline.
Level 2 · Supervised professional use

Support for professional quality

Producing an alert, proposal or structure that must be reviewed and discussed by a qualified professional.

  • anonymised cases;
  • human reference standard;
  • documented errors;
  • final professional validation.
View examples
  • second reading of recommendations;
  • drafting support for risk-assessment documents;
  • training vignettes;
  • structuring field observations.
Level 3 · Sensitive use

Health data or work organisation

Processing sensitive data, influencing an individual decision, monitoring activity or changing how work is organised.

  • formal assessment;
  • medical, legal and organisational governance;
  • secure environment;
  • information and involvement of affected stakeholders.
View examples
  • health-data analysis;
  • collective signal detection;
  • worker scoring or surveillance;
  • automated scheduling affecting workers.
SMEs and first pilots

Run a limited and reversible experiment

A small organisation can begin without complex infrastructure, provided that the scope is narrow and the rules are explicit from the start.

1

Select a bounded task

Prefer an activity that is repetitive, well understood and easy to verify.

2

Avoid sensitive data

Begin with public, fictional or genuinely anonymised material.

3

Name a pilot owner

One person must be able to change, suspend or stop the experiment.

4

Test several cases

Compare output with usual practice and retain examples of failure.

5

Measure more than time

Track quality, verification workload, autonomy and user experience.

6

Make an explicit decision

Continue, redesign or stop: scaling must not be automatic.

Minimum pilot record

What should be defined before testing begins

This checklist is a starting point for discussion. Sensitive uses require a more complete assessment and the involvement of relevant occupational, legal and worker-representation expertise.

Problem identified Task concerned People affected Data used Tool and provider Pilot owner Verification process Quality indicators Occupational-health indicators Pilot duration Stop criteria Documented final decision

Assess the project before deployment

Review likely effects on work organisation and people.

Plan post-deployment monitoring

Define now what will be observed once the system is used.

Review law and governance

Connect data protection, the EU AI Act, responsibility and worker involvement.

Written by Dr Charles Broutin, occupational physician and AI representative for the French Society of Occupational Health.

Search

Enter at least two characters.