AI Assistant Mistakes That Leak Bad Answers (And How To Fix Them)
The specific mistakes that cause internal AI assistants to produce wrong answers — and the engineering and process fixes for each.
As the founder of Perceptra, a Mumbai digital growth studio, I work with real businesses on these challenges every week. This guide is written for owners and decision-makers, not engineers.
Why bad answers from an internal AI are worse than no answers
Mistake 1: Training on outdated documents without a currency check
A policy document from 2021 that has been superseded but never formally retired will produce confident wrong answers about the current policy every time it is retrieved. The system does not know the document is outdated — it answers from whatever it found.
The fix: Every document in the knowledge base should have a confirmed review date and an assigned owner. Implement a periodic currency audit (quarterly at minimum) where each document is confirmed as still current or removed/replaced.
Mistake 2: No "I don't know" pathway
A system without an explicit "I cannot find this in the available documents" response path will either refuse all queries without clear answers (overly restrictive) or fall back to general LLM knowledge to fill gaps (dangerous — this produces plausible but fabricated company-specific answers).
The fix: Configure the system prompt explicitly: "If the answer cannot be found in the retrieved documents, say 'I cannot find this in the available documents. Please contact [human point of contact].' Do not use general knowledge to fill gaps."
Mistake 3: No citation requirement
An answer without a source citation cannot be verified. A wrong answer with a citation can be caught by the user who checks the source and finds the answer does not match the document. A wrong answer without a citation gets acted on.
The fix: Require every answer to include the source document name, section, and where possible the specific page or clause. "Based on your HR Policy Manual, Section 4.2, updated March 2025: your annual casual leave entitlement is..." is verifiable. "You get 15 days of casual leave per year" is not.
Mistake 4: Retrieving too many loosely relevant chunks
Some RAG implementations retrieve 10–20 document chunks for every query and provide all of them to the LLM, hoping the LLM will identify the relevant subset. When the retrieved chunks are of variable relevance, the LLM may synthesise them into a blended answer that incorporates content from multiple policies inappropriately.
The fix: Tune retrieval to retrieve fewer, more precisely relevant chunks (typically 3–5) with a high confidence threshold, rather than a large, low-confidence set. Measure retrieval precision during testing with representative queries.
Mistake 5: No feedback loop for identifying bad answers
Without a mechanism for users to flag wrong answers, bad answers accumulate undetected and erode team trust over time until the system is effectively abandoned.
The fix: Build a feedback mechanism into the interface — a simple thumbs-down button, a "this answer seems wrong" button, or a direct channel to report incorrect answers. Review flagged answers weekly during the first month, monthly thereafter.
Frequently asked questions
Investigate the root cause — is the source document outdated? Is the retrieval finding an incorrect document? Is the LLM synthesising retrieved content incorrectly? Address the root cause, not just the symptom. A wrong answer is a system diagnostic, not just a service failure.
This depends on the stakes of the use case. For low-stakes information retrieval (finding a meeting room booking procedure), 5–10% error rate may be acceptable with a clear verification pathway. For high-stakes use cases (legal clause retrieval, financial policy), aim for less than 2% with mandatory source citation and human review for any answer being acted on in a material way.
Yes — adding missing documents that would have answered incorrectly-handled queries, improving retrieval precision through prompt and parameter tuning, and refining the document chunking strategy all improve performance iteratively without requiring a full system rebuild.
Ready to Build
This For Your Business?
Book a strategy session. We scope your first project in 30 minutes, no jargon, no obligation.