An answer is only as good as its source
A recurring question reaches the commercial management: how does the approval of special terms work? The information is somewhere in internal documents, but nobody knows immediately where. Using AI could find matching passages and suggest a short answer. Its value would at first be orientation, not an automatic decision. The example is fictional. No real knowledge base was connected and no answer quality was measured.
With internal rules in particular, a fluent answer quickly sounds binding. NIST points out that generative AI can also produce false reasoning and false source references [Q1]. A mere “with source” in the text is therefore not enough. The source must actually exist and fit the question. The conclusion for this issue: a trial without a verifiable document version is too weak to derive a working rule from.
One case: special terms in a fictional business
Synthetic input: A fictional document titled “Special terms leaflet”, version 2 of 1 October 2026, contains three sentences in the section “Approval”: “Special terms of up to 5 % are checked by the commercial management. Above that, the managing director decides. Before any commitment, the calculation must be stored in the case file.” The internal question is: “Who may approve a 7 % discount for a new project?” The figures are solely part of the fictional case. They are no recommendation for real discounts and no digiFORMER rule.
Possible instruction: “Answer the question only with information from the leaflet provided. Name title, version and the exact section. If the source is missing, contradictory or outdated, say so. Do not make an approval decision.” A first trial could work with the one fictional leaflet and a deliberately different old version. The old version would be marked as test material so that it is clear whether the search finds the newer rule.
Illustrated output, not actually generated: “According to leaflet version 2, the managing director decides at 7 %. The calculation must be in the case file before any commitment. Source: Special terms leaflet, version 2, section Approval.” A good answer would add that it only restates the rule at hand. It would neither approve the discount nor claim that the calculation is already stored. If the current version is missing, “I cannot verify the applicable rule” would be the better output.
Error check: The reviewing person opens the cited source. They check the threshold, the responsible role and the condition on the calculation one by one. An error would be if the answer extends the 5 % limit to 7 % or presents an older version as valid. An invented section number would also be an error. The NIST risk of false sources makes this check necessary [Q1]. With documents that others can edit, a further risk arises: OWASP describes how injected instructions in documents can influence an AI application [Q3]. For a real set-up, access to sources and answers would have to be restricted technically and checked; that is not confirmed here.
Conditions and limits
The business needs a small, maintained selection of approved documents. Every source needs a recognisable version and a responsible person. If two rules contradict each other, the conflict must stay visible. An AI system must not derive a permission from the content of a document. Before real use, check whether the chosen tool reliably respects existing access rights. This issue promises no finished integration.
For personal data, the business must check before a real test which data it may process and which tool is suitable. The GDPR requires data minimisation and a lawful basis [Q2]. This is a general note, not a legal assessment of the example. If current documents are missing, document maintenance comes first. An AI search cannot derive a reliable rule from sources that do not exist.
Your small next step
Take a fictional leaflet with a version date and ask a question whose answer is clearly in it. Then check every partial claim against the source. Add a question without a verifiable answer. If the system then guesses, the check has failed. The Potential Analysis can prioritise several such processes.
Fictional handbook; no knowledge base was connected and no answer quality was measured. No statement is made about the access rights of a specific tool.
- [Q1] NIST, Generative Artificial Intelligence Profile, section 2.2. Supports the risk of false statements and sources.
https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf(checked 11 November 2026) - [Q2] European Union, General Data Protection Regulation, Article 5(1)(c) and Article 6. Supports data minimisation and lawful bases; not an assessment of a particular case.
https://eur-lex.europa.eu/eli/reg/2016/679/oj(checked 11 November 2026) - [Q3] OWASP GenAI Security Project, LLM01:2025 Prompt Injection. Supports the risk of manipulating input; no guarantee of protection.
https://genai.owasp.org/llmrisk/llm01-prompt-injection/(checked 11 November 2026)