Why AgOS says it does not know

The cost of a confident guess

Language models are good at producing answers that sound right. For general questions that is often fine. For questions about your own company, it is the worst failure there is: a made-up refund window, an invented deadline or a policy that was never written, delivered in the same calm voice as a correct answer.

Once people catch an assistant doing that, they stop trusting it, including on the answers it got right. Every answer has to be checked, and an assistant whose every answer must be checked is not saving anyone time.

Cite it, or say you do not know

AgOS answers only from what it can cite: the documents and pages in Knowledge, and what it has recorded from your connected tools. Every answer lists the sections it used. If nothing it can cite answers the question, it says so, plainly.

Citations do two jobs. They let anyone check an answer in seconds. And they hold the answer to its sources: a claim that cannot point at a passage does not belong in the answer.

A miss is a signal

When AgOS cannot answer, or a search turns up nothing, the question is recorded rather than thrown away. Questions that keep coming up unanswered are grouped together, and each group becomes a suggestion: this is the page your team should write.

That turns the assistant’s weak spots into a to-do list. The fifth person to ask about the travel policy is the reason it finally gets written down, and from then on everyone gets an answer.

The same rule for edits

The rule holds when AgOS edits a page, too. Improve with AI can make a page read better but may not add, drop or change a fact. If any number, link or code block from the original goes missing or changes, the rewrite is rejected. A page is the record, so the record does not get rewritten by accident.

Try it

Add three to five documents, ask something they answer, then ask something they do not. The second answer is the one to judge an assistant by. Get started.