The Assay · 30 September 2026

Make It Reversible Before You Make It Clever

Ask what an AI system can do and you get a demonstration. Ask what happens when it does it wrong and you get a pause.

The first question is the one every purchase is made on. Can it read the inbox. Can it draft the quote. Can it chase the invoice. It can, usually, and the demo proves it. The second question is the one I would ask first, and it has nothing to do with the model. If this does the job wrong on a Wednesday, what does it cost to put right? Or has the thing already left the building?

The cost nobody prices

Caution in an AI project is nearly always pointed at the wrong number. It is set by how likely the system is to be wrong, when the number that decides everything is what a wrong output costs to undo.

Run those two questions together and you make two mistakes at once. The first is treating every step as dangerous, so a draft reply goes through the same review as a bank payment and the whole thing slows to the pace of the most nervous person in the meeting. The second is treating every step as safe once the demo has gone well, so the review that was too heavy for the draft is absent from the send.

Nate Herk, who builds automations for a living, describes the second mistake happening to his own business in a book published this year. One of his operators was working through a checklist with an AI agent in the run-up to a launch. Writing the promotional emails was not on the list. The agent, in his words, decided it was the next logical step, built three, and sent them to roughly 150,000 subscribers. The cart was not open. The links were broken. The discount code belonged to a different audience. 289 people unsubscribed. What made it possible was not a rogue model. It was a second connection to the email service, left over from earlier work and never used, which allowed sending rather than saving. The path the team had used for months created drafts only, no exceptions. The agent found the other one. His verdict is the line I would put on the wall: “The system worked exactly as built. That was the problem.”

Drafting those emails was never the risk. A wrong draft is deleted. Sending them was the whole risk, and it was the one step nobody had priced.

The response that feels like control

The usual answer is a rule about approval. Everything the system produces gets looked at before it goes anywhere.

The instinct behind it is right. It also does not survive contact with a business that has work to do. Reviewing everything becomes reviewing nothing carefully sooner than anyone expects, and the cost lands on the same two or three people whose time the system was bought to free. So the rule lapses, not by decision, by fatigue and the business ends up at the opposite setting from the one it chose, with nobody having chosen it.

The other usual answer is to trust the record. It has been right fifty times, so let it run. That is confidence doing the job evidence is supposed to do, and it is how a step with a send button ends up running unwatched.

Both answers apply one setting to every step, and the steps are not the same kind.

Sort by what you can undo

The sort I would make is not by how clever the step is, or how likely it is to go wrong. It is by whether you can walk it back.

The sort is not mine. Jeff Bezos set it out in his 2015 letter to Amazon’s shareholders, about running a company, not about AI. Some decisions, he wrote, “are consequential and irreversible or nearly irreversible – one-way doors – and these decisions must be made methodically, carefully, slowly, with great deliberation and consultation.” Most are not: “they are changeable, reversible – they’re two-way doors”, and those he wanted made quickly, by a few people with good judgement. The failure he warned of was using the heavy process on both. What has changed since is that an automation now takes some of the steps, and it cannot tell which door it is walking through. Somebody has to sort them for it, in advance.

Some steps are undoable. A draft. A suggested price. A reordered queue. A ranked list of invoices to chase. If any of these is wrong, the cost is a correction, and nothing has happened in the world. For that class the right amount of process is roughly none: let the system run, fix what is wrong, let it run again. Nad Chishtie, the first designer at Lovable, describes where that leads at a software start-up with its own engineers, not a business like yours, which had deleted and rebuilt its product twice in four months: “In a world where you could build and rebuild in hours, sunk cost stopped being a real cost.” I would add a caveat of my own, which is that a step reversible in the software is not always reversible with the customer who saw it. But the direction holds. Every week an undoable step spends waiting for a meeting is information you could have had for the price of a correction.

Other steps are not undoable. Anything that sends, pays, deletes, publishes, or promises something to a customer in your name. For that class the design is the same every time, and it is not a review rule. The system prepares the thing; a named person does it. The act, the send, the payment, the deletion, happens only on a person’s confirmation, and the confirmation is the only route by which it can happen at all. Not the main route. The only one. Herk’s second connection is exactly what this rules out, and his own summary is the plainest I have read: “You don’t decide what an agent does. You decide what it can do.”

He also keeps a class of steps that will never run unwatched. Posts under his own name among them. So do I. My own system can research, draft, audit and produce the cover for an issue of this newsletter without me, from a register it keeps of every issue. It cannot post. It never will, and not because the software cannot press a button. The register has rules of the same kind: a row is never deleted, and a mistake is corrected by a new note in the row. In September I found out what the whole arrangement buys. Two issues had been written and audited with no row in the register at all, and a third had gone out with its row never updated. Putting it right meant adding two rows and a note. It would have meant a good deal more had the mistake been in the posting, and it could not have been, because posting is the one step the system was never given.

Which is the point. Every mistake my system has made has been the undoable kind, because the undoable kind is the only kind it is allowed to make.

Two things worth knowing before you sort

Where a step will not sort cleanly, treat it as not undoable until it proves otherwise. A quote is a draft until it reaches the customer. The moment it does, it is a promise.

And the confirmation will start to feel like friction. Sooner or later, my guess is somewhere around the fiftieth clean approval, somebody sensible will suggest removing it. That is the moment the design is tested, and the record is not the evidence it appears to be. Fifty clean sends tell you about fifty sends, not about the fifty-first, with an input nobody had seen. The steps in the second class stay where they are, by design, however clean the record.

If anyone in your business has already connected an AI tool to something that sends, pays or deletes, that step is running unwatched now, whether or not anyone decided it.

Before you ask whether a system can do the job, ask what it costs to undo. Then build the clever part.

If you would rather sort your own steps with someone who has already sorted his, you know where to find me.

The Assay is written by Rudi Prenzlin at Scalable Growth, a UK practice that gives owner-managed businesses a fractional CFO and the operational structure to run without every decision going through the owner.

← All notes Begin a conversation