I locked my phone in a hotel safe in Bali for a week and let the agents run the business. Here's every decision they made.

A confession before the week starts. We spend so much time building these systems for clients that theirs were, hands down, MUCH better than ours. Plumbers with the leaky taps. So the fortnight before Bali was a few late nights bringing our own setup up to the standard we'd been shipping to everyone else. If you've ever been too busy fixing other people's problems to fix your own, you know exactly how that felt.

Then two things you need to know, because they're the reason it worked and not luck.

First: we turned the ads down. Every channel to maintenance levels. I was expecting some level of failure. You cover every angle and you still assume something will break, and I'd rather dull a bad week than let a stranger down at full volume. Assume the best, plan for the worst. We don't like letting people down, and a lead that lands during a failure is a person who got ignored.

Second: how Amber is built, because this isn't "we switched on a chatbot." We own the brain. The prompt, the rules, the knowledge, the memory of what Firestarter is doing this month. That's ours, not a vendor's. Then we give that brain access to the tools: the CRM, the calendar, the phones, the inboxes. Not so she can do more. So that we can see everything she does. When the intelligence lives in your systems instead of inside someone's platform, failures get caught, checked and learned from, rather than buried where nobody's looking.

Not a FAQ with a smile. The latest articles, including this one. The pitch we're running this month. The ideas we're testing, what's cooking in R&D that we're allowed to talk about, and a fair bit that is, frankly, our secret sauce. If the thing answering your question can't speak frankly across everything we're doing, and notice what's particularly piqued your interest, then what is the point of it?

And we don't trust any one system to mark its own homework. Every day, our audit counts what actually came in: every call, every text, every chat, every email, every DM, by channel. What each person said, and where they are now. Then it checks that against what the CRM has. Real numbers, from the source. Not the dashboards Meta and Google hand you, which are always, without exception, generous to Meta and Google.

Hold that thought. It matters on Thursday.

A great day, which is a strange thing to write about a day I spent underwater.

Eleven inbound calls, every one answered on the first ring, and five new inbound leads across the channels.

Amber asked each of them what they were after. Not what we sell, what they were after. Took names and numbers, and put them into the next available slots in my calendar, by text, with a one-line note each. One wanted pricing on the spot. She gave him the honest version, that a number without looking at his list is a guess, and booked him anyway.

Two existing clients rang: routed, message taken, tagged for the team. One recruiter: declined politely in under twenty seconds, which is faster than I've ever managed and considerably more polite.

First thing I noticed, reading this back from a beach: she doesn't argue. She books.

Six more inbound calls, four more inbound leads, all handled. But the one I'll remember was web chat.

Someone who'd read The Cost of Slow at eleven o'clock on a Tuesday night (which tells you when business owners actually think about this stuff) typed:

She answered in four seconds. Nine messages. He asked about his own response times, she asked one question back, he admitted "about a day," she told him what that costs, in his numbers, not ours, and he booked. Wednesday, 10am.

I was asleep. The business wasn't. Read that again.

Eight calls handled. Three inbound leads, booked. And vendor day.

Four separate calls from people selling us things: SEO, a lead list, "AI transformation consulting," and, no word of a lie, office cleaning. Two of them. She took details from the ones that might matter, declined the rest outright, and not one reached me. I did the sums later: that's about forty-five minutes of my day that used to belong to people selling me office cleaning.

Also Wednesday: a no-show. Monday's 10am booking didn't dial in. Amber texted him at 10:07, he replied "sorry, mad morning," she offered two new slots, he took Friday. Nobody on my team touched it. The old version of this story is a lost booking and a follow-up email someone means to send on Thursday.

Six calls. Nine inbound leads, the biggest day of the week, on ads turned down to a trickle. All nine booked into the next available slots without a human involved.

And one thing that actually needed the founder: an existing client with an invoicing issue. Unhappy tone, specific numbers. Not a sales conversation, not a routing job. A judgement call.

Amber didn't try to fix it. She didn't apologise for things she couldn't verify, didn't promise anything, didn't ask him to hold. She said the person who could sort this properly was me, that she'd get it to me directly, and took the details. Then she flagged it: one message, tagged so it couldn't be missed, with the client's own words in it. It was the only thing all week that came through the wall of the hotel safe. I sorted it from a café. Ten minutes.

That's the whole game right there. She knew what wasn't hers.

The skill everyone's terrified an AI lacks, knowing when to stop, turned out to be the one she was best at.

Every existing client got their weekly report on time.

The conversion tracking that feeds each client's sales back to their ad accounts (the thing that decides whether their ads learn from real revenue or from guesses) ran at 100% uptime for the week.

The follow-up sequences kept running on schedule: every quote and enquiry from the previous fortnight got its next touch without anyone remembering to send it. One of them, a lead from August who'd gone completely quiet, booked. Follow-up number five. The statistics say that's where the sales are, and nobody I've ever employed has reliably made it to five.

You know I always give you the caveat. Here it is.

One lead came in through web chat, gave a name, a number and an email, and never made it into the CRM. The conversation happened, the details were exchanged, and the record wasn't written. A wiring fault, not a judgement fault. But the lead doesn't care about the distinction, and neither do I.

Here's why it's in this article instead of buried: the daily audit caught it. Every channel counted at the source, then checked against what the CRM holds, and that day the two numbers didn't match. The missing lead was added, contacted, reported, and the fault fixed. Amber now doesn't say "you're in the system" until the system confirms it.

That audit exists because of a rule we hold harder than any other:

AI loves to please. Left to report on itself, it will tell you a story of success every single time, and it will believe it. So the counting is done by a separate system that doesn't care about anyone's feelings, and the AI only ever sees the answer. That's the difference between an agent you can leave in charge for a week and one you can't leave for an afternoon.

Not that AI works. I knew that. Three things I didn't expect.

The decisions I was scared to delegate were the ones she handled best. Escalate or don't, decline or engage, book or qualify. Every one was right, because every one had a rule behind it that I'd written down instead of carrying in my head for thirty years.

The time she gave back wasn't the calls. It was the noise. The cleaning companies, the recruiters, the "just checking in," all the stuff that used to arrive between the calls that mattered.

And this: I've been going deeper on AI lately and I've come back with real concerns. The severe unprofitability underneath it, the heavily subsidised cost of every token, a bubble that may well pop next year. That's a whole article on its own, and it's coming. But it sharpened one thing rather than blurring it. The only AI worth building is the kind that makes a business money: answers the lead, books the call, catches its own mistakes. Not the kind that automates low-grade repetitive work for the sake of saying you did.

Profit first. Everything else is a hobby with a token bill. Hmmm.