The case for an AI that says “I don’t know”
Chatbots bluff because bluffing is what their scoreboard rewards. Change the scoreboard and the honest answer becomes the productive one.
Ask a product chatbot something its documentation doesn’t cover and you will usually get an answer anyway — assembled from adjacent material, delivered in the same even voice it uses when it actually knows. The system isn’t malfunctioning. It is doing precisely what it was built and measured to do, which is the uncomfortable part.
Language models complete plausibly by nature; left alone, they would rather produce a likely-sounding answer than a gap. Turning that instinct into an honest “I don’t know” takes deliberate engineering. And most deployments have little incentive to do that engineering, because of how they’re scored.
The scoreboard problem
The support-automation world grades itself on deflection: the share of questions the bot resolved without a human. As a cost metric it makes sense. As the primary metric it quietly defines success as “the human never got involved” — which a wrong answer achieves just as well as a right one, provided the asker doesn’t notice in time to complain.
The wrong answer’s cost is real; it just lands later and off the books. The customer configures their deployment around a capability that doesn’t exist. The seller repeats the bot’s claim in a call and gets corrected by a prospect who read your docs more carefully. By the time the damage surfaces, no one connects it to the chat transcript from three weeks earlier, so the scoreboard stays green. A system graded this way will drift toward confident guessing forever, because the metric can’t see the difference.
What a good decline looks like
The fix starts with a rule and gets interesting in the plumbing. The rule, for Deputy: if a claim isn’t in the briefing — the owner-approved record it answers from — it doesn’t get said. No adjacent-material assembly, no even-voiced improvisation.
But a bare decline just relocates the dead end, so the decline has to do work. When a question runs past the briefing, Deputy tells the asker plainly that this one needs the owner, pings you with the question, and marks it pending — visibly, so the asker knows a real answer is in flight rather than abandoned. The two failure modes people actually hate — the confident guess and the void — are both off the table. What remains is a small, honest state: someone who knows has been asked.
The loop that closes
Then the loop closes, and this is the half the deflection scoreboard never priced in. The question reaches you wherever you are; you answer once, out loud if you like — thirty seconds into your phone in a hallway. Deputy cleans up the transcript, delivers the answer to whoever asked in your framing, and then does the part a human assistant would forget: it finds the line in your briefing that your answer just made stale, strikes it, and hands you the diff for approval.
Jordan asked a question
“What happens if the site loses power mid-sync?”
Escalated to you · Deputy wasn’t sure
Answer once — Deputy handles it from here
The economics of the decline flip at that moment. The question that escaped the briefing didn’t cost an interruption; it purchased a briefing line. You were going to be asked it regardless — the loop just means you were asked once, at a moment you chose, by something that wrote the answer down. Same ping next month, any phrasing — answered instantly, cited to the line you approved, by a system that no longer needs you for it. Every reply you send teaches it, which is why the escalations thin out as the briefing fills in. A guessing bot never gets better this way, because a guess generates no question, no answer, and no line.
Where the boundary sits
One boundary worth drawing precisely: escalation is for questions you want answered once and never again — the recurring, gladly-delegated kind we mapped in the thesis post. Decisions still route to you as decisions; a stand-in that tried to absorb those wouldn’t be honest, it would be presumptuous. And yes, an escalated answer is slower than an instant one — minutes or hours instead of seconds. We take that trade without much agonizing: one slow answer that’s right, then fast forever, against fast answers you have to independently verify each time. The second option isn’t actually fast.
An answerer that guesses never gets better. One that asks you builds the briefing that retires the question.
Join the waitlistWe'll let you know the moment Deputy's open to you.
Or ask the live Deputy on the landing page