The most important part of an AI assistant is what it refuses to do

The default instinct with an AI assistant is to expand what it can do — answer more questions, handle more of the conversation, take more off the team. In a business where the AI talks to real customers about money and regulated processes, that's the wrong thing to optimize first. The work that matters is deciding, precisely, what it must never do on its own.

We think about it in three tiers.

Three tiers

Tier one is pure automation — no AI. Timed reminders, stage-based messages, confirmations. Deterministic, predictable, and most of the value.

Tier two lets the AI answer for itself, but only safe, factual questions: "where's the meeting point?", "what should I bring?". These have a correct answer that doesn't depend on who's asking, and getting one wrong costs nothing.

Tier three is the line: anything sensitive goes to a human, immediately. Money, cancellations, refunds, legal questions, anything involving a minor, anyone who sounds upset. The AI doesn't attempt these.

The design work isn't in tier two. It's in drawing tier three's boundary, and enforcing it.

The rules that matter are the negative ones

For the assistant we built into a regulated business, the important rules are the things it can't 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 it's human.
• It never advances someone's status — it can't move a customer to "approved," bend a rule, or make a promise the business has to keep.
• It always escalates the sensitive categories, even when it's fairly sure it could answer.

None of that is a limitation to apologize for. It's the reason the tool is safe to point at real customers — an assistant that can't misquote a price or invent a policy is one the owner can leave running. That build is written up in the boating-school platform case study.

"It could probably handle it" is the trap

Modern models are good enough that they usually could answer the sensitive questions plausibly. That's the danger. "Usually right" on a refund, a legal detail, or an anxious customer is a liability, because the failures land on the cases that matter most — in your customer's inbox, with your name on them.

Escalation isn't the model failing. It's the system working. The goal was never for the AI to handle everything; it was to handle the safe majority flawlessly and get out of the way for the rest.

An AI assistant that's allowed to do anything is a demo. One that knows, by design, what it must not do is a product. Where being wrong has a cost, the guardrails aren't a detail you add at the end — they're what makes it shippable.