The Assay · 23 September 2026

The Steps Were Never the Job

An AI agent can now do the visible half of a job convincingly. It reads the message, classifies it, drafts the reply, updates the record, moves the thing along. Watch it for ten minutes and it looks like the job.

Then look at the person it was bought to relieve.

They are not doing less. They are doing something different and, on most days, worse. Sitting between the agent and every case it had no rule for, answering exceptions as they arrive rather than at a pace they set. The work moved. The interruptions multiplied. Nobody got their afternoon back.

The steps were never the job

The steps are the visible part, which is exactly why they are the part that got copied.

The job is the deciding. When to make an exception and when to hold the line. Which customer gets the careful answer and which gets the quick one. What good enough looks like on a Friday afternoon with three jobs outstanding. When to stop and ask somebody rather than press on. Those are rules, real ones, consistent enough that the people around them can predict the answer, and nowhere in the business are they written down. They live in one or two heads and get applied all week without anybody calling them rules.

An agent cannot copy that, and not because the technology is weak. There is nothing there to copy.

Vin Vashishta wrote about the first-principles version of this at the end of August. His subject was whether AI is anywhere near general intelligence, not business automation. The observation is that a system configured closely enough to a person will do things the way a person does them, without being able to do what a person can. And that the resemblance is what makes it look like general intelligence when it is not. None of it supplies the logic and reasoning a person brings. He says that we automated the wrong logic.

Which I read across to a business of your size as this: the agent copied the doing, because the doing was the part on display. The deciding was never on display. So it stayed exactly where it was.

And here is the part that costs money. Automating the steps does not spread the dependence around. It concentrates it. Before, the person who knew the rules did the work and made the calls at the same time, in the same head, at their own pace. Now the work happens somewhere else, faster, and every case the rules do not cover arrives at that same desk as an interruption with a queue behind it.

The response that feels like progress

So you buy another agent. One for quotes, one for the inbox, one for the weekly report.

This is not foolish, and I want to be fair about it, because each one genuinely does something. Each demos well. Each removes real keystrokes. Anyone who has watched a good agent handle a routine case understands immediately why people buy a second.

What it produces is a faster business with the same shape. Every agent sends its exceptions to the same handful of people, and the more of the routine work you automate, the higher the proportion of what reaches those people is hard. You have not reduced the load on your best judgement. You have concentrated it and removed the easy cases that used to sit between the difficult ones.

Split the job before you automate it

The rule I work to is two lines long.

Automate where the rule is already explicit. Watch where it is not.

Explicit means somebody could write it on a card and a new starter could apply it on Monday: orders under this value go straight through, this customer is on these terms, anything past this date escalates. Plenty of work is like this and nobody notices, because obvious rules do not feel like knowledge. Where the rule is explicit, an agent can hold it, and the work actually leaves the person rather than changing shape.

Where the rule is not explicit, do not automate it. And do not go away and write it all down either. I have argued against that before and I still would: the blank page does not get started, and asked cold what your rules are you get a shrug, which is not evasion but the ordinary fact that a decision made often enough stops having steps. Instead, put the work in front of something that is allowed to stop, and record what it stopped on and what the person decided. Deciding in advance where it must stop is the whole of the governance you need at this stage, and the same decision is what generates the record. Keep that up and the implicit rules arrive in the order they actually matter, by how often they come up, rather than by how weighty they felt in a meeting.

Then the sequence has a middle, and nothing waits on a documentation project that was never going to happen. Explicit rules get automated now. Implicit ones get watched until they are explicit, and then they get automated too.

There is outside support for changing the work rather than the software, though it comes from a scale that is not yours. McKinsey’s State of AI, fielded across May and June and published in August, reports that nearly three-quarters of the organisations it classes as high performers say they have fundamentally redesigned workflows because of their AI use, against about a quarter of everybody else. It draws the distinction as redesigning the work rather than inserting AI into the work as it stands. Two caveats belong with that. Those are large enterprises with resources you do not have, and “high performer” there means a business that says it has seen significant value and attributes five per cent or more of its operating profit to AI. Self-reported, not audited. What survives both caveats is only the direction: among the ones reporting value, redesign is the norm; among everybody else, it is the exception.

I should say where my own system sits, since I have been arguing about other people’s. I have mentioned before that my weekly agent logs every decision that escalates to me and never records why it was decided. What it does record is why it came to me at all, in four categories: habit, policy, preference, or nobody else qualified. I built those to find bottlenecks. Read against this argument they are something more useful than that. Policy means an explicit rule exists and I could hand the decision over tomorrow. Habit and preference mean there is no rule, only me. Which means I have been keeping an inventory of my own unwritten decisions and filing it as a workload report.

Three things I am not claiming

Not all judgement can be made explicit, and the part that cannot is the owner’s actual job. The aim is to shrink what only one person can answer, not to get to zero, and anyone promising the second thing is selling something.

Changing how work is organised is neither free nor quick. It costs management attention, which in a business your size is the scarcest thing in the building. Scarcer than the money for the agent.

And the agents are not the wrong purchase. They are the second purchase.

Automate the steps of a job whose rules were never written down and you get a faster business that still cannot decide anything without you. That is not a failed AI project. It is worse than one, because it looks like a successful one.

Of the things you want a system to take on: which could somebody write on a card and which only work because you are there.

If you would rather not work out where that line falls on your own, 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