Your team needs AI training if it must learn to choose use cases and check responses; an integration provider becomes relevant if it must build a process connecting its information, the deliverable and its validation. We advise examining a real task to determine whether the blockage concerns the judgement to be exercised or the work to be organised. To choose a service, describe the work that is stuck in the team’s own words. What document needs to be produced? What information is missing? Where is a decision waiting to be made? We advise starting with this specific situation: it provides a subject to examine and a deliverable to discuss with the service provider.

Ask the person who carries out a use case to describe it. Keep the initial request, the resulting draft, the corrections and the selected version. Together, these provide a basis for discussing the actual work: what the person understands, what they check and what they revise before considering the document usable.

Key takeaways

  • We recommend expressing a training objective as an ability to put into practice. Choosing a suitable task, providing useful context, examining a response and explaining a decision are objectives to discuss with the team. Ask for exercises in which participants must justify their choices, using situations close to their work.
  • A useful scoping question is to ask where the work should go after the draft stage. Specify the recipient, the expected format and the decision to be made. If several people are involved, describe their handovers. You will then have a process to examine with those responsible for it.
  • Validation deserves an explicit role. Choose the person with the necessary professional judgement, then specify what they must check before accepting the document. Also ask how they report an error and send the work back for correction. Simply receiving a draft does not define this responsibility.
  • Examine the exit documentation and request a data retrieval exercise. The designated person on your team must be able to explain what they retrieve, where they find it and how they resume the work. Assess this exit procedure before expanding the work entrusted to the integration or its scope of operation.

Contents

How do you choose the first use case to examine?

The first use case to examine must have an identifiable deliverable and a person capable of judging its quality. We favour a task the team can describe in full, from the initial information to the final decision, so that corrections, responsibilities and the conditions for continuing the trial can be discussed precisely.

Choose a first use case for which you can recognise an acceptable output. A routine document, an identified recipient and a person capable of checking it provide good starting points for the examination. Favour a scope that the team can explain in full, through to the point when it accepts, corrects or rejects the prepared work. The expected deliverable deserves a description before the technical choice is made. Write down the essential sections, the elements requiring attribution and the information that must remain visible for review. Add an acceptable example and explain why you selected it. The service provider can then discuss an expectation that your team can justify.

To decide whether to continue a trial, examine the draft and the necessary corrections with its user. Ask whether the important information remains verifiable and whether the person knows when to reject the output. Select criteria the team can apply to the next trial, then explain the decision to continue, make corrections or stop. Keep a brief record of the decision: the need selected, the expected deliverable, the person responsible and the condition for continuing. This note can be used to examine a proposal or postpone the project. We advise revisiting it when the team requests a new use case, to decide again which work deserves to be entrusted.

How do you recognise a need for AI training?

A need for AI training can be recognised in the questions of judgement the team must learn to address: choosing a task, providing useful context and checking a response. We recommend examining these abilities using work documents, then setting objectives that participants can explain and put into practice themselves.

HENUSSE presents training as work to develop understanding, intended for executives and their teams. The stated scope covers how AI works, its costs, its limitations and ways to use it while preserving the company’s freedom of choice.1 The training offered by HENUSSE aims to provide a framework for decision-making. The choice of tool remains open, and the training can be taken independently of an integration project: the service described allows participants to work through the decision without committing to sign up for an installation afterwards.1

The HENUSSE training day format connects the concepts covered to work situations. The stated programme combines practice, comparison and writing the first safeguards. These activities describe the content planned for the day, moving from concepts to the uses being examined.1 In a fictional example, a team is unsure how to check a response prepared from an internal document. The training uses this document as an exercise: participants find the relevant passages, identify an insufficiently supported conclusion and explain the correction. The expected learning deliverable is their reasoning when checking the response.

A survey published in April 2025 covers 319 knowledge workers, who shared 936 personal examples of using generative AI in their work. In this self-reported survey, greater confidence in AI is associated with less critical thinking. Greater confidence in one’s own abilities is associated with more critical thinking. These associations concern the uses studied and do not demonstrate a cause-and-effect relationship.2 We recommend expressing a training objective as an ability to put into practice. Choosing a suitable task, providing useful context, examining a response and explaining a decision are objectives to discuss with the team. Ask for exercises in which participants must justify their choices, using situations close to their work.

In its training approach, HENUSSE distinguishes between the executive’s responsibilities and those of the team. The executive makes decisions about time, budget and rules; the team works on using AI, checking its outputs and identifying situations to exclude. Shared rules connect these responsibilities.1

When does AI integration become relevant?

AI integration becomes relevant when the work to be done requires an explicit process connecting the information, the prepared document and its validation. We advise describing the steps and responsibilities before building, including an acceptable output, situations requiring a stop and a way to resume the task if necessary.

To build a new workflow, HENUSSE states that it first defines the steps, data, validation and how to revert if the result is unsuitable. The tool prepares and passes on the work. A person remains responsible for validating the deliverable and making the next decision.3 A useful scoping question is to ask where the work should go after the draft stage. Specify the recipient, the expected format and the decision to be made. If several people are involved, describe their handovers. You will then have a process to examine with those responsible for it.

In a fictional example, a team prepares a meeting summary from its raw notes. The draft organises the decisions, the people responsible and the points requiring clarification. The person responsible for reviewing it checks the attributions and leaves uncertain points open before sending the summary to the participants concerned.

What information should you prepare to scope an integration?

To scope an integration, prepare the information used to carry out the task, its origin and the people who can correct it. Add an example of an acceptable deliverable and the validation rules, so you can discuss with the service provider work that your team can describe and check through to handover.

To prepare the working context, select information whose origin and responsible person are known to the team. State who can correct it and how to identify the version to use. A useful information base should also be discussed from this perspective: which documents are authoritative for this task, and who decides when they contradict each other?

A research publication revised on 5 May 2023 describes remote attacks against applications that integrate AI: hostile instructions are placed in data that the application is likely to retrieve. The risks examined include data theft and information contamination. The attack described can therefore occur through the data consulted, without a direct interface with the application.4 Prepare situations in which work must stop: missing information, a contradiction or a request outside the scope. Describe the expected response in each situation. We advise examining these stopping points as carefully as the usual process, to establish who the case goes back to and what clarification that person requests.

Involve the person who carries out the task, the person who validates it and the person familiar with the documents used in defining the scope. Each must be able to describe their responsibility and difficulties. We recommend making their disagreements explicit among the scope decisions to be resolved, so that the proposed process rests on choices they jointly take responsibility for.

What should you ask a service provider before starting?

Ask the service provider for a written scope, an expected deliverable, each person’s responsibilities and the criteria for accepting the work. We also recommend examining the information used, the documentation handed over and the exit procedure, so that your team can assess the proposal and explain how it will resume the task after the assignment.

HENUSSE states that its integrations are self-hosted in the client’s environment by default, with the database, rules and history kept there. If the workflow calls an external engine, the information sent is specified in the quotation. The setup is presented as documented and reversible, including in this case.3 Favour a demonstration that your team can examine step by step. Ask to see the information used, the document prepared, the points still to be checked and the expected validation. A demonstration then becomes an opportunity to discuss working choices and the limits of the proposed scope with the people concerned.

Validation deserves an explicit role. Choose the person with the necessary professional judgement, then specify what they must check before accepting the document. Also ask how they report an error and send the work back for correction. Simply receiving a draft does not define this responsibility. Before commissioning an integration, ask who will carry out the work and who will answer questions after handover. Ask for clarification on the elements your team will need to maintain and the changes it will be able to make itself. These answers give a place to the people who will use the process and those who will develop it further.

How do you assess the possibility of changing service providers?

The possibility of changing service providers is assessed by examining the documentation, data retrieval and your team’s ability to resume the work. We advise requesting an exit exercise and having the person who will be responsible explain the procedure before expanding the scope entrusted to the integration.

In the HENUSSE integration method, validation and exit procedures are described in writing. This description states who checks the deliverable and who makes the decision. It also specifies how to retrieve the data and stop the workflow while retaining the company’s knowledge.3 Examine the exit documentation and request a data retrieval exercise. The designated person on your team must be able to explain what they retrieve, where they find it and how they resume the work. Assess this exit procedure before expanding the work entrusted to the integration or its scope of operation.

Resuming work manually deserves a description the team can use. Specify where to find the information and how to continue the task when a workflow is stopped. Have this description reviewed by someone familiar with the daily work. Their review should focus on the actions needed to prepare, check and hand over the expected deliverable.

The 2023 NIST AI RMF 1.0 framework provides for post-deployment monitoring plans in its Manage 4.1 category. Their content includes evaluating user feedback, appeal and override mechanisms, decommissioning, incident response, recovery and change management.5 A change to a document, internal rule or recipient deserves a fresh check of the process. Designate the person who will examine these changes and the situations to revisit. We advise keeping previously reviewed cases to discuss differences with the team, then recording the decision made on each change examined.

Frequently asked questions

How can you tell whether AI training has been useful?

AI training is useful if the team is better able to justify its choices and check the prepared work, because these abilities directly address the need for judgement. To evaluate it, ask participants to choose a task, explain the information needed and show how they accept, correct or reject a response.

Can training and integration complement each other?

Training and integration can complement each other when your team needs to learn how to judge use cases and have a process for carrying them out. We advise giving each service its own objective and deliverable, then choosing their order according to the decisions the team must make before continuing the work.

Who should validate the work prepared by AI?

Responsibility for validating work prepared by AI should rest with a person who has the professional judgement needed to accept or reject it. We recommend assigning this responsibility explicitly, specifying the points to check and establishing how the person requests a correction or clarification before making their decision.

When is it better to postpone an AI integration?

It is better to postpone an AI integration when the team cannot yet define the expected deliverable or designate the person who will validate it, because the scope still needs to be settled. We recommend clarifying the task, the information needed and the acceptance criteria, then revisiting the decision on that basis.

Sources

  1. Formations IA pour dirigeants et équipes
  2. Lee et al. (2025)
  3. Intégrations IA opérationnelles
  4. Greshake et al. (2023)
  5. AI RMF Core

Théo Hénusse, Founder of HENUSSE