Before you automate anything, ask whether it should exist

We saw all of this once before, with robotic process automation ten years ago. The lesson most of us paid for: often, simplifying the process worked better than automating it. It is very easy to make the same mistake again at a hundred times the scale — to point something brilliant at a process that should not exist, and congratulate yourself on how fast it now runs.

There is a distinction worth stealing here, made sharply in the exponential-organisations world: automation makes what you already do cheaper. Transformation changes what you are. Most enterprise AI programmes are automation wearing transformation vocabulary, and you can tell which one you have in an afternoon: pull your ten largest AI initiatives and mark where the value is booked. If it is labour or cost removed from an existing process, it is automation, whatever the title slide says. Automation savings are real — but they are competed away within a few years, and they quietly convert into a supervision cost nobody priced.

The precedent is a century old and exact. When factories first replaced steam engines with electric motors, they bolted the motors onto the same floor plans — single drive shafts, fixed layouts, centralised power. Productivity barely moved for a generation. It arrived when factories were redesigned around what electricity actually made possible: distributed power, flexible space, machines that ran independently. The economic historian Paul David documented it in 1990, and AT&T's chief data and AI officer said the modern version plainly: you cannot force agentic AI into the existing human process — you have to reimagine the process to be agentic. The process redesign is the work.

The most dangerous number in AI is a big efficiency percentage on a process that should not exist. You have not transformed anything. You have optimised your way into keeping it.

The sorting tool is the fourth question — the one I have asked in every discovery session for twenty years. For each step: what happens five minutes before, and five minutes after? It exposes the constraint, the ecosystem, and the steps that exist only because a person once needed time. Asked honestly, it sorts any workflow into three piles:

The pile	What to do	What to expect
Steps that should not exist	Delete them. No AI required. Approvals that ration a cost that has collapsed, handovers that broker information a system now holds, reports nobody decides from.	The cheapest value in the whole programme, and the least glamorous. This pile is why process mapping comes before anything clever.
Repeatable work	Automate it, under the gates, measured against a baseline.	This is where the return lives. On programmes I have measured, repeatable work improved several times more than judgement work, on the same technology — and the published studies show the same shape (see the evidence room).
Judgement work	Augment it, and log the decisions — the record is what eventually makes more possible.	Expect a fraction of the repeatable return, for now. The instinct in a professional firm is to point AI here, because this is where the pride is. It is where the return is worst.

Then hold the line on what the gain actually is. Efficiency, productivity and value are three different things, and one does not automatically become the next. What AI hands you is time. Something that took a day now takes an afternoon — real, but not money. It becomes money on exactly two paths: the person does more work that earns something, or the cost comes out. Most programmes never make that choice, because it is uncomfortable and nobody clearly owns it. So the saving sits on a slide as a percentage, the headcount stays where it was, and a year later the CFO asks where the money went.

The discipline that fixes it is blunt: any efficiency gain above a threshold is paired, in the same quarter, with a named initiative that consumes the freed capacity — or it is not counted. Freed time with no destination gets reabsorbed into producing three decks an hour instead of one, which is activity, not value. This is the third gate from Part six restated as a standing rule, and it is the single most transferable practice in AWS's entire study of 154 executives: the organisations that broke through decided where the freed capacity was going before the gains arrived.