The Machine Room

Start with the Boring Pipes

Nothing in Daniel Okafor's first month of automation would have impressed anyone at a technology conference. He did not build a chatbot. He did not ask an AI to design a single menu. He connected four pieces of software that moved money and appointments from one place to another, and then he left them alone.

That order was deliberate, and I wish I had followed it in 2021. Payments and bookings are the most stable tasks in almost any Run Log. They happen the same way every time, the decisions inside them are few and already written down, and when they break somebody notices within a day, because either the money does not arrive or a client sits alone on a video call. Delivery, the thing I automated first, is the opposite. Its decisions hide. Mine hid in a billing address.

Daniel's stack in June looked like this, priced in pounds per month. The products are illustrations, not recommendations. Any equivalent does the same job.

• Recurring card billing through his payment processor: no subscription, about 1.5% plus 20p per charge, roughly £32 across his clients
• Accounting software that pulls those payments in and matches them: £33
• A booking page clients use to choose their own call slots: £10
• An intake form that feeds new requests into his queue: £20
• An automation service that joins the four together: £16

That comes to £111 a month.

He then went back to the times in his Run Log. Chasing late payments had taken him about four hours a month, and he had found this the worst part of the business, worse than any brief. Matching payments to invoices took two. Arranging calls by email ran to five hours, because restaurant owners answer messages at midnight and then disappear for three days. Copying requests out of his inbox into a list took another three. In total, fourteen hours a month.

£111 divided by 14 is just under £8 for each hour returned. His Revenue per Hour of Attention was well above £50, so the arithmetic is not close.

There is a cost, and it is not on the invoice. The first fortnight took him nine evenings of setup, and in the second week the booking page offered a client in Leeds a slot at 3 a.m. because Daniel had left his laptop set to Lagos time after a visit home. The client booked it anyway. She wrote: "Assumed you were very keen."

A Fast First Draft, Never the Last

Generative AI earns its place further up the Run Log, where the work is made of words rather than money. I use it daily for three jobs: drafting, summarising and the first pass at support replies. In each case it hands me something rough and quick, and I fix it. The speed is the entire benefit. The judgement stays where it was.

Daniel's clearest gain came from summaries. He held about twenty client calls a month, and after each one he spent around 25 minutes turning his scrawled notes into a brief for the queue. Now the call is transcribed, an AI tool drafts the brief in a couple of minutes, and he spends eight minutes correcting it against what the client actually said. Twenty calls at 25 minutes is 500 minutes. Twenty at ten is 200. Five hours a month returned, for a tool costing £18.

Support is where I lean on it hardest. My template catalogue draws about sixty emails a week, most of them variations on a dozen questions, and the answers already sit in my Run Log. The AI drafts a reply from that document; I read it, adjust it and send it. Writing each reply myself took about nine minutes. Checking a draft takes three. Sixty emails at nine minutes is 540 minutes a week. At three, it is 180. Six hours saved, every week, provided I never skip the reading.

I nearly did, once. In March 2024, late on a Thursday, I approved a batch of drafts by skimming the first line of each. One told a buyer in Bristol that refunds were available for sixty days. My policy is fourteen. The model had found "60" elsewhere in the document (it was the number of templates in a bundle) and attached it to the wrong sentence with complete confidence. I caught it the next morning, before she asked for her money, and only because the reply sat in a review folder rather than going out on its own.

That folder is the guardrail. Nothing an AI writes reaches a customer until a named human has read all of it. In my business the named human is me. The time saved comes from the typing. The reading is where the work still is.

On my monitor there is a yellow note, slightly curled now, that says: read it as if you bought the thing.

Fail Loudly

The worst automation failures make no noise at all. A broken payment rule does not ring a bell. It simply stops collecting, and you find out six weeks later when the bank balance looks thin. The second worst kind makes plenty of noise, but only in somebody else's inbox.

Mine was the second kind. In August 2022 I set up an automation to invoice the handful of firms that bought team licences for my templates each quarter. It watched my accounting software for any invoice marked "updated" and emailed it to the client. What I had missed is that sending an invoice also updates it: the software stamps it with the date it went out. So the automation sent the invoice, saw an update, and sent it again. It ran on a short delay, which was the only mercy. Between Saturday lunchtime and Sunday evening, a three-partner accountancy firm in Bath received the same invoice for £1,180 thirty times.

I was at a cousin's wedding in Nashik with my phone switched to silent. The automation reported nothing, because from its point of view nothing had gone wrong. On Sunday night the firm's office manager wrote: "We will pay it. Please make it stop."

They did pay, and renewed the following year, which says more about their patience than my systems. I lost nothing but some dignity. I could have lost the client.

Since then every automation I run follows three rules, and Daniel built his stack on the same ones.

• Every failure sends an alert to a person. An error goes to my phone as a message, not to a log file I never open.
• Every action that touches a customer has a ceiling. No automation may send more than two emails to the same address in a day. A third attempt stops the whole thing and tells me.
• Every automation reports when it works, too. A summary arrives each Monday listing what ran and how often. Silence from a rule that normally fires forty times a week is itself an alert.

The ceiling is the one people skip, because it feels like distrust of your own work. It is. Daniel's alerts go to a separate phone number he keeps for the business, and in his first year it buzzed eleven times. Ten were trivial. The eleventh was a client's card being charged twice.

I still have the thirtieth copy of that invoice. It is pinned beside the door, numbered in red biro, the way the firm's office manager forwarded it to me.

What the Machines Still Get Wrong

Anger comes first. When a customer writes in fury, an AI draft apologises the way a hotel lift does: politely, at length, and without going anywhere. In January 2025 a practice manager at a solicitors' firm in Croydon wrote that one of my contract templates had cost her a morning, because a clause about notice periods could be read two ways. The draft I was offered began "We sincerely apologise for any inconvenience" and then gave her three bullet points. She would have read it as a form letter, because it was one. I deleted it and wrote four sentences of my own. I admitted that the clause was ambiguous, said which reading I had meant, and sent the corrected file within the hour. She bought two more templates that spring.

Facts come second. A model states a figure with the same calm whether it is right or wrong, and the sixty-day refund was a mild case. Daniel once asked for the standard bleed on a printed menu (the margin of extra artwork that gets trimmed off). It said 5 mm. His printer in Salford wanted 3 mm. Nothing came of it. A wrong VAT threshold in one of Meera's client emails would be a different matter.

Prices come third, and here the machine is least useful precisely because it sounds most reasonable. It can suggest a figure. It cannot know that this client struggled to renew last year, or that you want this job badly enough to take less. Pricing is judgement, and judgement belongs in the Founder bin.

My rule sits in my Run Log in plain words. Three kinds of output never reach a client until I have read every word against a source: anything replying to someone upset, anything containing a number, and anything that commits me to money. In practice that covers most of what goes out. I have made my peace with that. The machine saves the typing. I still do the reading.

This Week

List every automation and every AI task you run, or plan to run. Beside each one, write three things: who reads its output before a customer sees it, what happens when it breaks, and where the alert goes. Any line you cannot fill in is a failure that has not happened yet. Then put thirty minutes in your diary for the first Monday of every month and go through the list again. That is the audit. While you do it, mark the steps that need a skilled human but do not need you.

Daniel's first list ran to nine automations. Two had a blank where the checker's name should have been, and in pencil beside both he had written: someone who isn't me.