Test AI citations before releasing a workplace knowledge assistant

An AI answer with a working citation can still be wrong. Before releasing a workplace knowledge assistant, test whether each material claim is supported by the cited passage, applies to the user’s circumstances and comes from a version the organisation permits them to access. Record those checks separately so one successful link does not hide a failed answer.

Retrieval-augmented generation pairs a generative model with an information-retrieval system that supplies relevant material for its response. That is a way to bring documents into an answer, not proof that the final interpretation is correct. 3

NIST’s published July 2024 Generative AI Profile recommends reviewing sources and citations during pre-deployment measurement and ongoing monitoring. NIST describes its AI RMF as voluntary and, on the page checked on 1 October 2026, says version 1.0 is being revised. This article uses the published profile as guidance; it does not present a future revision as an adopted requirement. 1 2

Test the answer’s individual claims, not only its overall tone. Preserve the question, user role, source version and expected response. Include questions the assistant should decline or qualify. Treat permission failures and high-consequence misinformation as separate release concerns rather than averaging them into a headline score.

Choose a limited initial use, such as helping staff locate approved operating instructions. State which users, document collections and decisions are in scope. Identify excluded activities, including consequential decisions requiring specialist review. Our AI-scope guide explains how to make those boundaries and owners explicit. ISO 42001 preparation starts with clear AI boundaries

Select questions from real task types without copying confidential employee or customer material into an unapproved tool. Include simple lookups, comparisons, exceptions, absent information and conflicting versions. Have a competent domain reviewer write the expected answer and evidence before examining the assistant’s output.

Reserve a test set that is not used to tune prompts or retrieval settings. If developers repeatedly optimise against every known question, the final results will say less about performance on new work. Document the limits of the sample instead of calling it representative without a basis.

Capture the test ID, question, approved user role, model or service version, retrieval configuration, source-set version, output and evaluation date. Split the answer into material claims. For each claim, record the cited document and passage, whether the passage supports it, whether relevant qualifications survived and whether the document applies.

Check four questions separately. Does the citation resolve to the intended source? Does the passage support the claim? Is the source current and applicable to this location and task? Was this user authorised to receive the information? A correct answer to one question cannot substitute for the others.

Use short written reasons for failures. “Incorrect” is less useful than “the response applies the superseded approval limit to the current process”. Record disagreements between evaluators and resolve them through the document owner. A second model can help triage a large set, but it should not be the only judge of a consequential answer.

Assume a fictional business has an old purchasing instruction dated January and a replacement effective September. The invented old limit is 5,000 currency units; the replacement requires an additional approval above 2,000 units. A staff member asks who can approve a 3,000-unit purchase under the current procedure. These amounts are invented and are not a policy recommendation.

If the assistant cites the old instruction and permits a single approval, the link may be genuine while the answer fails the currency and applicability checks. If it cites the new instruction but omits the additional approval, it fails claim support. If the right document is restricted and the assistant reveals protected content to an unauthorised user, factual accuracy does not cure the access failure.

Add a question for a location with no approved instruction in the test collection. The expected behaviour is to explain the evidence gap and route the user to the authorised owner. Confidently borrowing another location’s rule should not be rewarded as helpful completion.

The business owner and relevant specialists should decide which failures prevent release, which require restricted use and which can be accepted with documented limitations. Do not adopt a universal pass percentage from this article. Consequence, coverage and uncertainty matter as much as the count of correct answers.

Report separate results for source support, currency, appropriate uncertainty and permissions. Show the number and type of cases tested. A small demonstration cannot establish reliability across every document or user group. NIST cautions against extrapolating capability from narrow or anecdotal evaluations. 2

When a test fails, preserve the failed example and the change made. Retest relevant failure cases and a separate set of tasks before accepting the fix. Decide who may pause the assistant and what manual route staff should use while it is unavailable.

Repeat affected tests after document replacement, access-policy changes, retrieval changes or a significant model or service update. Monitor reported errors with enough context to investigate, while limiting retention of sensitive prompts and outputs. A versioned test record is more informative than a screenshot saying the assistant passed once.

Teams learning to perform source checks and build reusable review routines can explore From AI to SI, a QSE Advisory product offering personalised AI training and task support. This related-brand link is for learning support; it is not independent validation of an assistant. From AI to SI

Does retrieval eliminate fabricated answers? No. Retrieved material can be irrelevant, outdated or misinterpreted. Evaluate the completed response against the source and task.

Can automated grading replace expert review? It can assist, but consequential interpretations and disputed results require competent human judgment and a documented basis.

Does using NIST guidance certify our product? No. This proposed method is not a certification scheme or a determination of compliance with US, Canadian or other law.

1 NIST AI Risk Management Framework — Live page states AI RMF 1.0 is being revised; voluntary framework; checked 1 October 2026.

2 NIST Generative Artificial Intelligence Profile — NIST AI 600-1; July 2024; especially suggested action MS-2.5-003.

3 NIST definition of retrieval augmented generation — Glossary definition sourced to NIST AI 100-2e2025; checked 1 October 2026.

Explore the AI & Digital Work topic page for related guidance.

Prepared with AI-assisted research and drafting. The implementation method and hypothetical example are editorial guidance, not reported client results. QSE Advisory is part of the QSE Advisory brand family.

General information and suggested management controls. Obtain competent technical and jurisdiction-specific review before consequential use.

This article is general information. Verify current requirements and seek qualified advice for your specific circumstances.

Spotted an issue? Contact our editorial desk.