Verify before circulating

What evidence should you require before using an AI response in a business decision?

1960 € per session, excl. VAT

In short

This seven-hour day, excluding breaks, establishes a process for checking AI responses used in documentary research. Participants break down claims, verify references, compare evidence and document uncertainties. They produce a validation grid and measure the quality of a response before and after correction.

Why this training?

A fluently written response can mix accurate facts, weak inferences and non-existent references. In documentary work, these errors become difficult to spot when a synthesis circulates without its sources or the conditions under which it is valid. Intuitive review is not enough: decisive claims must be identified, their evidence located, and a distinction made between absence of evidence and contradiction. A shared protocol makes judgements explicit and focuses effort on business risks. Continuing case study: The fictitious organisation Avelcor Équipements is preparing a purchasing brief comparing three filtration systems. A generated synthesis contains references that cannot be found, performance figures taken out of context and documents from different versions. The documentation service must determine which conclusions can be circulated using an entirely fictitious dossier. Transfer to the workplace: Apply the grid to an upcoming anonymised research brief with a second reader, record disagreements and compare the percentages of substantiated claims before and after revision. Adjust the checks according to the errors observed.

Objectives of this training

  • Break down a response into verifiable claims and classify them by criticality.
  • Check the existence, relevance and currency of cited references.
  • Check claims against source passages and recalculate decisive figures.
  • Characterise uncertainties and decide between validation, correction and escalation.
  • Measure documentary quality before and after revision.

Who is it for and what are the prerequisites?

Roles and tasks concerned

  • Information specialists, information monitoring officers and analysts producing source-based syntheses.
  • Business managers validating AI-assisted research briefs or comparisons.

Prior skills

  • Use generative AI regularly and know how to formulate a request with appropriate context.
  • Know the documents and validation criteria relevant to your work.
  • Be able to consult references and perform simple calculations in a spreadsheet.

What you will work on

  1. Mapping errors and their consequences

    Objective: Prioritise checks according to the intended use. Fabrication, omission, outdated information, confusion over scope; distinguishing fluent writing from accuracy.

  2. Building the claims register

    Objective: Make every conclusion checkable. Facts, interpretations, recommendations; granularity of claims and levels of criticality.

  3. Auditing references

    Objective: Distinguish a plausible citation from a source that provides evidence. Author, version, date, primary source, independence; an existing reference does not automatically prove the claim.

  4. Comparing evidence

    Objective: Resolve factual and quantitative discrepancies. Independent cross-checking, contradictions, units, test conditions and limitations of comparison.

  5. Correcting and deciding

    Objective: Produce a usable brief without concealing its limitations. Validation rules, withdrawal of claims, cautious wording and consultation with an expert.

  6. Measuring and calibrating the review

    Objective: Assess the quality achieved and the consistency of judgements. Proportion of substantiated claims, remaining critical errors and disagreements between readers; limitations of an overall score.

What you take away

  • Claims register with criticality, evidence and status.
  • Reliability grid with decision rules.
  • Corrected and annotated comparative brief.
  • Before-and-after measurement table and review protocol.

Method and assessment

  • Short instructional inputs followed by immediate practice on the continuing case study.
  • Individual work followed by work in pairs, with cross-review of the outputs.
  • Guided group discussion: what works, what needs correcting and what remains a human decision.
  • Preparation: Identify a type of response in which an error could affect a business decision. Arrange access to the tools authorised by the organisation; use only fictitious or anonymised cases.

Observable criteria: Every decisive claim has a status and a locatable source passage, or an explicit statement that evidence is absent. ; References are checked for existence, version and scope. ; Decisive figures are recalculated with units and assumptions documented. ; Any unresolved critical contradiction blocks circulation of the conclusion concerned. ; Before-and-after indicators use explicit and comparable denominators..

Frequently asked questions

Can AI verify its own response?

It can suggest checks, but its self-assessment does not replace independently consulted evidence.

Does every sentence need to be checked?

Break down the factual claims, then strengthen checks on those that determine the decision.

What should you do if a source remains inaccessible?

Mark the claim as unverified and seek alternative evidence or suspend its use.

Is a high score enough to allow circulation?

No: a single critical error may require a correction, even with a good overall result.

Last updated: