Evaluate AI for bounded preparation and checking tasks, with evidence-linked review before consequential use.

Do not assume that a system capable of extracting text or drafting an estimate is proven to produce an acceptable complete file.

Potential pilot tasks include document classification, factual extraction, photograph organisation, draft narratives, cross-document discrepancy checks and suggested estimate items. FM describes using or experimenting with AI to summarise consultant material and prepare file narratives subject to human checking. Cotality describes automated estimate auditing against configured carrier guidelines. These are different functions and should receive different tests. (fm.com)

For each automated task, specify:

• The permitted inputs and outputs.
• Which fields must have source evidence.
• What constitutes an unsupported answer.
• When the system must abstain or request clarification.
• Who checks the result and can override it.
• Which downstream actions are prohibited without authorisation.

NIST’s Generative AI Profile identifies confabulation, including false content and misleading explanations or citations, as a risk. Consequently, a generated explanation is not itself evidence that an estimate item is correct. (nvlpubs.nist.gov)

Begin with a narrow workflow and measure total handling effort. Expand only when the evidence supports the next scope—not because the demonstration looks fluent.