When the Assistant is not sure
'I do not know' is a correct answer here, not a failure to produce one. The Assistant sits on top of a system that sends mail, places calls and moves deals, so a confident invention does not stop at being wrong — it gets acted on. When it is unsure the Assistant says which part is uncertain, gives you what it does have, and offers the fastest way to settle it.
The argument, plainly#
A chat model that answers a general question badly wastes a minute. An assistant wired into a business answers a question that becomes an email to a customer, a follow-up somebody chases, or a price nobody agreed. The expensive part is not the wrong sentence; it is the action taken on it, by a person who reasonably assumed the sentence was checked.
So the trade is not 'accuracy versus helpfulness'. It is a small, visible cost — you learn the thing is not recorded — against a large, invisible one: a fabricated detail that reads exactly like a real one and is discovered by a customer rather than by you. A system that never admits doubt has not removed the doubt; it has moved it out of sight.
This is why an admission is treated here as a first-class result rather than a degraded one. The reasoning is set out at length in Why uncertainty is a valid answer from a business agent; the same discipline shows up as safe refusals in sales, where declining to answer a question about terms is better behaviour than answering it inventively.
The same principle, elsewhere in Connect#
- Memory returns an empty tier rather than dropping it. Asked what applies to one piece of work, memory answers with all four tiers — workspace, channel, endpoint, contact — including the ones holding nothing. 'Nothing is set here' is a visible answer instead of a silence you have to interpret.
- A send is uncertain until the provider acknowledges it. Connect will not call a message sent because it handed it over; the state stays uncertain until there is evidence.
- The capability vocabulary has five words, not two. Something can be available, operator-only, foundation, not yet, or reference — because collapsing 'the code exists' into 'you can use it' is the overclaim that wastes somebody's afternoon.
- A refusal is logged as a decision. The decision log holds what was refused and under which rule, because a refusal is a choice your business made.
None of that is decoration. Each one is a place where a system could have produced a confident-looking answer cheaply and instead produced an accurate one that costs a sentence of explanation.
What you get instead of a guess#
| What it says | What it means | Your next move |
|---|---|---|
| Nothing in this workspace records that | The lookup ran and found nothing | Write it down — as memory, knowledge or a follow-up — so the next answer is different |
| I can see X but not Y | Partial evidence, honestly split | Act on X; settle Y from the source it names |
| Two sources disagree | A real conflict, surfaced rather than resolved | When sources disagree |
| This is outside what I may do | Authority, not knowledge | What the Assistant may not do |
| A question raised for a person | It has put the question where a human will see it | Answer it in Needs You |
The last row is the one people miss. An open question can be handed to a person rather than left in a chat window — it appears in the queue a person already works, ranked with everything else waiting on a human, and it drains when it is answered.
Making the next answer better#
When the Assistant says nothing is recorded, decide whether that is a gap in the data or a gap in reality.
Result If the fact exists only in somebody's head, the fix is to record it. If it does not exist at all, the answer was already right.
Put durable business statements into Knowledge, and situation-specific ones into Memory at the narrowest tier that applies.
Result Knowledge grounds answers everywhere; a contact-tier memory shapes work on that one person without affecting anybody else.
Re-ask afterwards.
Result The same question now returns an answer with a source you can open. That loop — ask, discover the gap, fill it, re-ask — is how a workspace gets good at answering itself.
Questions#
Is a hedged answer just the Assistant being unhelpful?
A hedge that names what is missing is more useful than a confident answer you cannot check, because it tells you exactly what to do next. A hedge that names nothing is the unhelpful kind — if you get one, ask which part it is unsure about, and it will say.
Does uncertainty mean something is broken?
Usually not. The commonest cause is that nobody has written the thing down: a conversation on a channel Connect does not carry, a decision taken in a meeting, a document nobody added as a knowledge source. That is a data gap, not a fault.
Can I tell how sure it is?
Read the wording and the evidence together. A claim with a source you can open is checkable; a claim with no lookup behind it in the folded step detail has nothing behind it. Where an answer came from covers reading that detail.