An employee asks the AI: "What is the daily allowance limit for business travel abroad?" The answer comes fast and fluent. But is it right? From which policy, from which version?
Fluency is not accuracy
Language models can describe things they do not know just as fluently. In casual chat that is a small problem; in a business decision it is a big risk. A decision based on the wrong policy can affect a whole process.
How sourced answers work
The organization's documents are first split into parts and made searchable. When a question arrives, the system first finds the relevant parts, builds the answer only from them, and writes the document and page next to every piece of information. When the user taps the source, the document opens on that page.
This approach has three results:
- Verifiability. The user can check the answer in seconds.
- Honesty. If the documents do not contain the answer, the system says so instead of guessing.
- Respect for permissions. The system answers only from documents the user can access.
Tables and scanned documents
Much of an organization's documentation is not plain text: tables, scanned contracts, forms. A good system must read these pages too, re-reading difficult pages with vision models. Otherwise the most critical information stays on the hardest page to read.
Trust is a result, not a feature
Employees do not trust AI from day one. Trust forms when they check the source a few times and see that the answer was right. Showing sources is the only way that trust can be built.