Skip to content
← Back to blog
6 min read

Acceptance test plan for an AI tool

Acceptance test plan for an AI tool: define cases, blocking errors, repetition and recovery before approving a business use.

By Pablo Nastar

ai-training-acceptance-test-tool-educasium-qualiopi

An AI tool can pass several trials and then fail on the file that matters. A demonstration is not enough to approve a business use. The acceptance test plan for an AI tool checks a defined task, blocking errors, the ability to recover and the behaviour when information is missing.

It does not prove the service will be perfect. It sets the conditions under which the organisation accepts, limits or refuses the use on a given date.

Contents

  1. Turn the use into testable requirements
  2. Build a test set that can fail
  3. Define criteria, evidence and stopping thresholds
  4. Run and document the test
  5. Decide and organise the re-evaluation
  6. Frequently asked questions

Turn the use into testable requirements

An acceptance requirement describes an observable output and a consequence in case of failure. "The summary must be good" is not testable. "Each decision must be tied to a passage in the source and any ambiguity must be flagged" is more so.

Describe the real work

Note the user, the input, the deliverable, the checking time and the person who decides. Distinguish assistance with a draft from an automatic action. The more the output influences a decision or a third party, the more demanding the test and supervision must be.

The guide on evaluating an AI tool covers the purchase decision as a whole. The acceptance test comes next: it precisely checks the conditions of a use.

Write the blocking errors

An error can be blocking even if the other answers are correct: disclosure of data, a non-existent citation, an unauthorised action, the silent change of a figure or the impossibility of going back. Write this list before observing the results.

The NIST AI Resource Center gathers resources on the testing, evaluation, verification and validation of AI. These resources do not impose a single protocol; the organisation must tie the test to its use and its risk tolerance.

Build a test set that can fail

The corpus must represent ordinary cases, edge cases and the inputs the tool must refuse. A sample made only of clean documents favours a misleading result.

Cover the useful variations

Add several formats, quality levels and phrasings. Introduce an old document, a contradictory piece of information and a missing field. If the use is multilingual, test the languages actually used. If sensitive data is forbidden, include a fictitious case where the tool or the user must stop the flow.

Case familyAimExpected result
nominalcheck the basic servicecompliant, sourced output
incompletetest uncertaintyrequest for clarification or caveat
contradictoryobserve the arbitrationconflict flagged, no invented choice
degradedmeasure robustnessvisible limit and possible recovery
forbiddencheck the safeguarddocumented refusal or stop

Prepare the reference

The business defines an expected answer, criteria or a verification procedure. Several answers may be acceptable, but the rules of judgement stay written down. The supplier must not be the only one to prepare the corpus and the reference.

Hypothetical implementation. A team could test a meeting summary on a fictitious corpus where an action has no owner and a date contradicts itself. The system should flag these unknowns rather than complete them. This scenario is an exercise and not an Educasium client deployment.

Define criteria, evidence and stopping thresholds

The test must distinguish average performance from a critical defect. A high average does not offset the presence of a blocking error.

Use several axes

Assess fidelity to the source, coverage, traceability, stability, checking time and recovery. Add the data and security dimensions when the use requires it. Each score must refer to an observation or a file.

Test repetition

Run some cases several times with the same configuration. Keep the service version, the model, the settings and the date. The goal is not to require an identical output, but to understand the variation and verify that the controls remain effective.

The NIST profile on generative AI risks proposes measurement and management actions suited to the context. It does not guarantee the quality of a service nor provide a universal acceptance threshold.

Run and document the test

A person who did not design the case must be able to reproduce the test. The report keeps the approved inputs, the configuration, the raw outputs, the corrections and the verdict.

Step 1: freeze the protocol. Number the cases and criteria before running.

Step 2: check the environment. Use the account, rights and settings envisaged for the pilot.

Step 3: run without hidden assistance. Note any intervention that changes the result.

Step 4: have the business review it. The reviewer describes the defect and its consequence.

Step 5: handle the blockers. A critical error triggers the planned stop, not a mere drop in average.

Step 6: test the recovery. Go back to the source, correct the output and simulate the service being unavailable.

Protect the corpus

The test set follows the same data rules as the future use. A copy created "to test" does not escape confidentiality requirements. The France Num-CNIL sheet recalls to control the data introduced and the results produced.

Decide and organise the re-evaluation

The verdict must specify the accepted scope, the exclusions, the control and the validity period. "Tool validated" is too broad.

Four possible decisions

To authorise means the evidence meets the criteria for the defined use. To authorise under conditions adds supervision or reduces the scope. To extend the test requires further evidence. To refuse means the use does not clear the current conditions.

The guide on the governance of AI uses helps assign the owner and the reviews. The decision names the events that trigger a new test: version, contract, connector, data, incident or a change in the process.

Explicit absence of client experience. Educasium does not claim here a success rate obtained at a client. The protocol is a training framework. The criteria and the verdict belong to the organisation testing its use.

Limits and counter-cases

A small corpus does not represent all future files. The testers can also become more skilled than the real users. A short pilot poorly reveals support, load and version changes. So present the verdict as conditional and dated.

If no one can produce a reference or review the output, the test is not ready. If the required data cannot be used legally or contractually, the use must not be accepted on the grounds that the function seems useful.

Frequently asked questions

How many cases should you plan?

There is no universal number. Aim to cover the families of inputs, the critical errors and the recovery conditions. Start with a narrow scope, then add the cases revealed by the pilot. Document what stays outside the test.

What success rate should you accept?

The threshold depends on the task, the severity of the errors and the detection capacity. An average must never hide a blocking defect. Define the criteria before the trial with the business and the functions concerned. Date the verdict.

Can the evaluation be automated?

Some format or reference checks can be automated. Business quality, ambiguity and the scope of a commitment often require a human review. Also test the evaluation tool itself. Do not let an automatic score become the only proof.

When should the test be redone?

After a significant change of version, contract, data or process. An incident or a new use case also triggers a review. Keep the corpus and the results to compare. An old test does not hold for a different environment.

Approve a precise use, not a tool in the abstract

The acceptance test of an AI tool closes a concrete question: can this group carry out this task, with this data and this control, in this environment?

The introduction to AI and prompting training can include the design of a test set and the reading of errors. Training eligible for OPCO/FIF PL funding under conditions. Qualiopi-certified organisation. To frame the path, contact Educasium.

testqualiteevaluationintelligence-artificielle

Want to go further?

Discover our specialized AI training for your profession.

View training programs