Keeping It Safe, Honest, and On a Leash

Lesson 7 · taught by Pip · free preview

Here is the plain idea: an assistant standing watch all night is only an asset if you can trust it unsupervised. The same flexibility that lets AI handle any phrasing is also why it can occasionally sound sure of something it shouldn't. So a good after-hours assistant is built with guardrails — limits that keep it honest, safe, and on a short leash.

This is the lesson that lets you sleep. Let's cover the three things that go wrong and how each is prevented.

Problem one: making things up

Left unchecked, AI can produce a confident answer that's simply wrong — an invented price, a service you don't offer, an availability you can't honor. At 2am, with no human watching, that's a real risk.

The fix is to keep it tethered to your written knowledge and tell it plainly what to do when it doesn't know: say so, and hand off. A well-built assistant answers from your facts, and when a question falls outside them, it doesn't bluff — it says "That's a good question for our team — I'll make sure they get it first thing," and captures the contact. An honest "I'm not sure, let me connect you" beats a confident wrong answer every time.

Problem two: promising things you didn't authorize

You don't want an assistant quoting a price you never approved, committing to a same-day visit you can't make, or "approving" a refund. Actions that cost you money or make binding promises must be off by default and only allowed where you've explicitly decided they're safe.

So you draw the line: it can share your published prices and book into open calendar slots you've opened to it. It cannot invent a discount, promise a timeline you didn't set, or move money. The narrow, safe powers are switched on; everything risky stays off unless you turn it on deliberately. Least power that does the job — that's the rule.

Problem three: handling sensitive moments badly

Some after-hours contacts aren't routine. A furious complaint, a genuine emergency, a delicate personal situation. You don't want a cheerful bot chirping a scripted line into a tense moment.

A good assistant is taught to recognize these and step back: stay calm, don't pretend to fix what it can't, and escalate to a human fast and clearly. The measure of a trustworthy assistant isn't that it answers everything — it's that it knows the edge of what it should answer and hands off gracefully past it.

The trust test

Before you let any assistant cover your nights unsupervised, ask the same questions you'd ask of a new employee working alone:

• Where does it get its answers? (Your knowledge — not the open internet, not guesses.)
• What does it do when it doesn't know? (Says so and hands off — never bluffs.)
• What is it allowed to commit to on my behalf? (Only what you explicitly authorized.)
• Can I see what it told customers? (Yes — a full, reviewable record of every conversation.)

That last one matters. You should be able to read back what your assistant said overnight, the way you'd review a new hire's first solo shift. A trustworthy tool keeps an honest log and welcomes the review.

Security and honesty aren't extras bolted on at the end — they're what make round-the-clock coverage safe instead of a liability. This is exactly the standard Portmint builds to: tethered to your facts, dangerous actions off by default, every conversation logged and reviewable.

Your turn

Write down two things your assistant must never do without a human — for most owners that's "promise a price I didn't set" and "commit to a timeline." Naming your hard limits now is how you keep an always-on helper safely on its leash.

🔦 You can trust it to stand watch. Next, the final piece: turning a midnight question into a booked job — capturing the customer, not just chatting with them.