Why Generic AI Gives Your Team Wrong Answers — And What To Do About It
Why generic AI consistently gives wrong answers to company-specific questions — and the specific solution that eliminates this problem for Mumbai teams.
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.
The mechanism behind the wrong answers
The three types of wrong answers generic AI produces
Type 1: Fabricated specifics. "What is our standard notice period?" — an AI trained on internet data might answer "the typical notice period for companies in your industry is 30–90 days," confidently presenting generic industry information as if it were your company's specific policy. Team members who do not realise this is a guess may act on it incorrectly.
Type 2: Outdated public information. If your company has any publicly available information (a website, press releases, public filings), generic AI may have incorporated this into its training data — but at a training cutoff date. Team members asking about current policies may receive answers based on a description of your company from 2022, presented as current.
Type 3: Plausible but wrong policy synthesis. "How many casual leave days do employees get?" — if generic AI recognises the question as company-specific, it may synthesise an answer from what it knows about typical Indian employment practices, producing an answer that is plausible for a company of your type but completely wrong for your specific policy.
Why this is worse than simply not having an AI
The specific danger is not wrong answers in isolation — it is confident, well-structured wrong answers that team members trust because they came from an AI system. A team member who asks a colleague and receives a clearly uncertain "I'm not sure, let me check" knows to verify. A team member who receives a confident, formatted AI answer with no indication of uncertainty does not receive the same signal to verify.
The specific solution: a custom AI assistant trained on your documents
The solution is not better prompting of generic AI (no amount of prompting gives generic AI access to your documents it does not have) or more frequent reminders not to trust generic AI (behaviour changes are hard to sustain). The solution is providing your team with an AI system that actually has access to your documents — and consistently tells them when it cannot find the answer in those documents rather than guessing.
The interim measure while a custom system is being built
Until a custom AI system is deployed, establish a clear team policy: generic AI tools can be used for general tasks (writing, public research, coding) but are explicitly not authorised for queries about company-specific policies, pricing, or procedures. Provide the manual lookup path (the specific document and how to access it) for the most frequently asked company-specific questions. This reduces harm from wrong answers while the longer-term solution is built.
Frequently asked questions
Partially — uploading documents to ChatGPT's conversation context improves accuracy significantly for those specific documents. The limitations: document upload has size limits that make it impractical for a full policy library; it is a per-conversation manual step rather than a persistent knowledge base; the uploaded documents may enter OpenAI's training pipeline depending on the product tier being used; and it does not provide the access control, citation quality, or integration capability of a properly built custom system.
Ask directly, and create an environment where team members feel safe being honest. Some informal observation — watching over a team member's shoulder, reviewing what AI tools they are using and for what — may also be informative.
No — the fundamental issue is not model capability but information access. However good generic AI models become, they will not have access to your company's internal, confidential documents unless you explicitly provide them through an appropriate, privacy-respecting architecture. Better models improve general capability; they do not solve the company-specific information access problem.
Ready to Build
This For Your Business?
Book a strategy session. We scope your first project in 30 minutes, no jargon, no obligation.