The most important part of an AI assistant is what it refuses to do
Most teams try to make an AI assistant handle more. In a business that deals with money, contracts and regulated processes, the value lies in what it is forbidden to do. How we draw that line and enforce it.
The usual instinct with an AI assistant is to expand what it can do: answer more questions, handle more of the conversation, take more work off the team. When the assistant talks to real customers about money and regulated processes, that is the wrong thing to optimize first. The work that matters is deciding, precisely, what it must never do on its own.
We organize that decision into three tiers.
Three tiers
Tier one is automation with no AI. Timed reminders, confirmations and messages triggered by a change of stage. It is predictable, easy to audit, and it delivers most of the value.
Tier two lets the AI answer on its own, but only safe, factual questions such as “where is the meeting point?” or “what should I bring?”. These have one correct answer that does not depend on who is asking, and an occasional mistake costs very little.
Tier three always goes to a person. Money, cancellations, refunds, legal questions, anything involving a minor, and any customer who sounds upset. The AI does not attempt these.
| Tier | Who handles it | Examples | Cost of a mistake |
|---|---|---|---|
| 1. Automation | Rules, no AI | Reminders, confirmations, stage messages | Very low, behavior is predictable |
| 2. AI answers | The AI, alone | “Where is the meeting point?”, “What should I bring?” | Low, a factual slip |
| 3. Escalation | A person, always | Money, refunds, cancellations, legal questions, minors, upset customers | High, so it is never risked |
The design effort is not in tier two. It is in drawing the boundary of tier three and enforcing it.
The rules that matter are prohibitions
For the assistant we built into a regulated business, the important rules describe what it cannot do:
- It never quotes a price. Prices carry commitments and vary by situation. A confident wrong number is worse than no answer.
- It never signs as a person or implies that it is human.
- It never changes a customer’s status. It cannot mark someone as approved, bend a rule, or make a promise the business would have to keep.
- It always escalates the sensitive categories, even when it could probably answer.
None of these is a weakness to apologize for. They are the reason the assistant is safe to put in front of customers. An owner can leave it running because it cannot misquote a price or invent a policy. That build is described in the boating school platform case study.
What this looks like in practice
A message arrives at 11 pm: “I need to cancel and get my money back.”
An assistant with no limits reads the refund policy and replies. It does so confidently, possibly incorrectly, to a customer who is already frustrated, at an hour when nobody is watching.
The tiered system does something less impressive and much safer. It recognizes a refund and a cancellation as tier three, sends a short acknowledgment that a person will follow up, and flags the conversation for the morning. Nothing irreversible happens overnight. That uneventful outcome is the feature.
“It could probably handle it” is the risk
Current models are good enough to answer most sensitive questions plausibly. That is exactly the danger. Being right most of the time about a refund, a legal detail or an anxious customer is a liability, because the failures land on the cases that matter most, in a customer’s inbox, under your company’s name.
Escalation is not the model failing. It is the system working as designed. The goal was never for the AI to handle everything. It was to handle the safe majority reliably and step aside for the rest.
An assistant that is allowed to do anything is a demonstration. One that knows, by design, what it must not do is a product. Where a mistake has a real cost, the guardrails are not something to add at the end. They are what makes the assistant fit to ship.
For the next step, deciding when an AI may act automatically and when it must ask first, see Auto, needs approval, manual.
Working on something in this territory?
If any of this sounds like the problem in front of you, tell us where you are and where you'd like to be. We'll come back with a plan within two business days.