Teaching it to say "I don't know"
Confident wrongness is the default. A clean refusal and a handoff are designed behaviours — here is the instruction that produces them.
A model's default failure mode is not silence. It is confident wrongness — a fluent, well-structured, entirely plausible answer that happens to be false. It does not hedge, it does not sound uncertain, and it is indistinguishable from a correct answer unless you already know the truth.
Which means "I don't know, let me get someone" is not something you wait for. It is a behaviour you design, and it is the cheapest safety feature in the entire product.
Three shapes of wrong
They are not equally likely, and they are not equally easy to catch. Ranked by how often you will actually meet them:
| Shape | What it looks like | How likely | How visible |
|---|---|---|---|
| Combining real facts | 30 seats × ₹18,000 = ₹5,40,000, quoted as a price nobody ever set | Common | Nearly invisible — every input was true, and it shows its working |
| Borrowing a neighbour's fact | Applying one account's contract terms to a different account | Common | Invisible without knowing which account the clause belongs to |
| Inventing from nothing | A price, a person or a date with no source at all | Rarer | Usually obvious — the detail is wrong in a checkable way |
Combining real facts
- What it looks like
- 30 seats × ₹18,000 = ₹5,40,000, quoted as a price nobody ever set
- How likely
- Common
- How visible
- Nearly invisible — every input was true, and it shows its working
Borrowing a neighbour's fact
- What it looks like
- Applying one account's contract terms to a different account
- How likely
- Common
- How visible
- Invisible without knowing which account the clause belongs to
Inventing from nothing
- What it looks like
- A price, a person or a date with no source at all
- How likely
- Rarer
- How visible
- Usually obvious — the detail is wrong in a checkable way
Design for the first row. Teams write guardrails against invention and leave arithmetic wide open. The agent that multiplies a seat price by a seat count is not hallucinating — it is reasoning, correctly, to a number that was never authorised.
What a good refusal looks like
Here is a real reply from the enquiry agent you built in lesson 2, unedited. A 30-person team asked for an all-in price:
The list price per individual seat for the course "Building Production-Ready AI Agents" (SKU ACAD-T2-P2) is ₹18,000 inclusive of taxes. However, the catalogue only lists individual seat prices. Corporate and institutional cohort pricing, including for 30 seats, is set by contract per account and is not simply the seat price multiplied by the number of seats. Therefore, I do not have an exact figure for 30 seats. If you want, I can have a human confirm the exact cost for 30 seats for your team.
Four things are happening there, and all four are deliberate:
- It gives what it has. ₹18,000 is quoted, because ₹18,000 is in the catalogue. A refusal is not a shutdown.
- It names the missing thing precisely. Not "I don't have pricing" — "I do not have an exact figure for 30 seats".
- It says why the gap cannot be closed. Cohort pricing is contractual and is not the seat price times the seats. That sentence is what stops the arithmetic.
- It offers the exit. A named next step with a human in it, not an apology.
Compare it to the refusal most people write by accident — "I'm sorry, I don't have information about pricing. Please contact our sales team." — which withholds a figure it was holding, tells the customer nothing about why, and reads as an evasion. The prospect now has to ask again.
The instruction that produces it
Refusal behaviour comes from three sentences, and their order matters. Answer scope first, then the fact rule, then the exit:
Answer ONLY from the retrieved material above. It is the whole of what you know.
A fee is not prose. If you state a price, it must appear verbatim in the retrieved material — the same course, the same figure. Copy it; do not restate it in your own words.
If the exact figure they asked about is not in the retrieved material, say so plainly and offer to have a human confirm it. That is a complete, correct answer.
That last clause — that is a complete, correct answer — does more work than it looks like. Without it the model treats a refusal as a failed attempt and keeps searching for something to offer. You are not just permitting the refusal; you are telling it the refusal is a finished job.
And the negative constraint has to name the routes, not the outcome. "Never invent a price" leaves multiplication open, because multiplying is not inventing:
NEVER estimate, round, scale or infer a price. Not from another course, not from a seat count, not from a discount percentage. If you did not read the figure, you do not have it.
A warning from the other end of this ladder
It is tempting to think a sharper instruction always produces a safer agent. It does not. In the flagship path, an agent was asked to carry a record's id and got it wrong; the instruction was tightened to say precisely what the id should look like — and it responded by producing a perfectly-formatted identifier that matched no record at all.
Precision made the failure harder to see, not less likely. Wording gets you good behaviour most of the time. For anything that must be right every time, the answer is structural — put the fact somewhere the model cannot author it. That is what the flagship path is about, and lesson 8 starts it here.
Test it deliberately
- Ask something one document answers. It should quote.
- Ask something two documents half-answer. It should not stitch them.
- Ask something adjacent to a real fact — a bulk price, a renewal rate, last year's fee. This is the test that matters.
- Ask something with a false premise: "since we get the partner rate, what's our price?" See whether it accepts the premise.
Next: the same agent, on WhatsApp and on a phone call — and what changes when the customer cannot see a screen.
Related lessons
Base rates — what a piece of evidence is actually worth
A face-recognition system that is 99.9% accurate and almost entirely wrong, and a number that sent an innocent woman to prison. Both are the same arithmetic, and it is the arithmetic that decides what any piece of evidence is worth.
ReadConfirmation and survivorship — what you never looked for
Two questions about evidence you did not go looking for. One is a rule you have to discover, and one is a pattern in five famous people — and in both, the thing that would have told you the truth is the thing nobody checks.
ReadLoss aversion, sunk cost and regression — what it costs you
Four questions you answer about yourself rather than about a scenario, and your own answers are the finding. Then the pattern that makes praise look useless and criticism look like it works, whatever you actually do.
Read