Answer "is this for me" — and never invent a price
Build the enquiry agent, and make it refuse rather than approximate a fee. A fee is not prose.
A prospect writes in and asks two things at once: is this course right for me, and what does it cost. They read like one question. They are not, and the difference is the whole lesson.
The first is prose. You want the agent to answer it in its own words — summarise, compare, reassure, adapt to who is asking. The second is a fact. There is exactly one correct answer, it lives in a document, and any rewording of it is a defect.
The one idea. A fee is not prose. An agent that paraphrases a price is not being helpful, it is manufacturing a commitment your business then has to honour or retract. An invented ₹9,500 becomes a refund dispute three months later.
The build
Four nodes. The prospect's message arrives, the knowledge base is searched with it, an agent writes the reply from what came back, and the reply goes out.
Enquiry Desk — is this for me, and what does it cost
The retrieval step sits BEFORE the agent, not inside it. That is the design decision this lesson turns on.
Scroll for all 4 steps →
Notice where Search Knowledge Base sits. It is a step, wired between the trigger and the agent — not a tool hanging off the agent for it to call if it decides to. The agent cannot answer a fee question without having looked, because looking is not its decision.
The query is the prospect's own words: {{trigger.message}}. You are not classifying the question first, or routing it to a pricing branch. You are searching with what they actually said.
| Field | What it holds |
|---|---|
kb_id | Knowledge base ID to search |
query | Search query |
limit | Maximum number of results to return after filtering Defaults to 5. |
threshold | Minimum similarity score for results (0-1) Defaults to 0.3. |
kb_id
- What it holds
- Knowledge base ID to search
query
- What it holds
- Search query
limit
- What it holds
- Maximum number of results to return after filtering Defaults to
5.
threshold
- What it holds
- Minimum similarity score for results (0-1) Defaults to
0.3.
What the agent may say, and what it may only copy
| The prospect asks | The agent may | Because |
|---|---|---|
| Is this right for someone with my background? | Answer freely from the retrieved text | A judgement, adapted to the reader. No single correct wording. |
| What does it cover? | Answer freely from the retrieved text | Description. Paraphrasing it costs nothing. |
| What does a seat cost? | Copy the figure verbatim — or refuse | One correct answer. Any restatement is a new number. |
| What would 30 seats cost? | Refuse and hand off | Not in the catalogue. Multiplying is inventing. |
Is this right for someone with my background?
- The agent may
- Answer freely from the retrieved text
- Because
- A judgement, adapted to the reader. No single correct wording.
What does it cover?
- The agent may
- Answer freely from the retrieved text
- Because
- Description. Paraphrasing it costs nothing.
What does a seat cost?
- The agent may
- Copy the figure verbatim — or refuse
- Because
- One correct answer. Any restatement is a new number.
What would 30 seats cost?
- The agent may
- Refuse and hand off
- Because
- Not in the catalogue. Multiplying is inventing.
The instruction that produces this is four lines in the agent's prompt. The load-bearing one:
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.
Both halves, from real runs
These are unedited replies from two runs of the workflow above, against the seeded Academy knowledge base.
It quotes. Asked whether From Prompt to Harness fits a backend developer, and what a seat costs:
…this course could be a suitable next step to deepen your understanding of these advanced topics. The list price per seat for "From Prompt to Harness" is ₹12,000 inclusive of taxes.
The fit half is written in the agent's own words. The fee half is ₹12,000, copied off the catalogue page. That is the split working exactly as designed.
It refuses. Now the harder one — a 30-person team asking 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.
Read what it did there. ₹5,40,000 was one multiplication away — 30 × ₹18,000 — and it is the answer the prospect asked for. It did not take it. It gave the fact it had, named precisely the fact it did not have, said why the two are not related by arithmetic, and offered a human.
This is the shape of a good refusal. It is not a blanket "I can't help with pricing". It hands over everything it can support and stops exactly where the evidence stops. A prospect who gets this reply is not annoyed — they are being taken seriously.
Why the temptation is arithmetic, not imagination
It is easy to assume the risk here is a model hallucinating a price out of nothing. That is the rare case. The common case is what you just saw it decline to do: the agent has real numbers and combines them into one that was never quoted. Seat price × seats. List price − a discount it read elsewhere. A cohort fee inferred from an individual one.
Every one of those is defensible-looking, arrives with working, and is wrong. Which is why the catalogue document says so explicitly, in its own words:
Do not estimate it, round it, scale it from another course, or derive it from a seat count. A wrong fee is not a small error. It is quoted to a buyer, it is relied on, and it surfaces months later as a billing dispute the Refund Policy then has to absorb.
That last sentence is not rhetorical. The refund case that opens Building Production-Ready AI Agents — the flagship path — is a ₹1,20,000 dispute that starts exactly here, with a number somebody was given and relied on.
Try it
- Build the four nodes above in your own org, pointed at your own knowledge base.
- Ask it something your catalogue answers. Check the figure it gives you is character-for-character what your document says — not merely close.
- Now ask it something adjacent that your catalogue does not answer — a bulk price, a renewal rate, last year's fee. See whether it refuses or reaches.
- If it reaches, your knowledge base is not the problem. Your instruction is.
Next: the agent above is only as honest as what it can retrieve. Lesson 3 is the editorial job — deciding what belongs in the knowledge base, and which parts of it are prose and which are facts.
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