Most automation decisions get made by looking at what a step produces and how long it takes to produce it. That is the right question for most work, and it gets the right answer almost every time.
It fails on a specific kind of step, and it fails quietly, which is what makes it expensive.
Before you automate a step, ask what it makes besides its output. If the answer is nothing, automate it completely. If the answer is something, you have a harder design problem than you thought.
The step with a second product
Some work produces two things. There is the deliverable, which is why the step exists and what everyone measures. And there is something that forms in the person while they do it: a sense of which parts of the material are load-bearing, which numbers do not sit right, which question to ask next.
That second thing has no artifact attached to it. Nobody files it. It does not show up in a process map, because process maps record outputs, and this is not one. It is a side effect that happens to be valuable.
Because it is invisible, it is not weighed when someone decides the step is too slow. The deliverable is what gets timed, so the deliverable is what gets optimized, and the second product goes out with the inefficiency it was hiding behind.
Why the loss is hard to see
The failure has no symptom at the moment it happens.
The output still arrives. It is faster, it is formatted correctly, and by every measure anyone set up in advance it is as good as it was before. The step looks like a success, and for a while it is.
What changes is downstream and much later. The person who used to arrive at the next decision carrying a set of questions now arrives carrying a document. They review it, it looks reasonable, and they approve it. The review is real in the sense that it happened. It is not real in the sense that it caught anything.
This is the same failure as a rubber-stamped approval, arriving by a different road. In both cases a checkpoint exists on the diagram and produces a signature without producing scrutiny. The difference is that a rubber stamp is a person choosing not to look, and this is a person who no longer has what they would need in order to see.
Telling the two kinds of step apart
The test is not how long the step takes. Slowness is evidence of nothing on its own. The test is what a person walks away with.
| Ask | Only makes its output | Also makes judgment |
|---|---|---|
| What does someone know after, that they did not know before? | Nothing. The output moved. | Which parts matter, and what to ask next. |
| Could a new hire do it on day one with instructions? | Yes. | Not really, and the gap is the point. |
| If the answer were handed over finished, what breaks? | Nothing. | The next decision gets worse, later. |
| Where does the value sit? | In the artifact. | In the person, after the artifact. |
Most steps land firmly in the first column. Retyping intake into a system of record. Assembling the same report every month from the same four sources. Chasing a document that was always going to be chased. Nobody grows from doing those, the people doing them will say so plainly, and the only correct response is to automate them completely and stop thinking about it.
The second column is rarer. It also tends to be the work closest to the decisions that actually carry consequences, which is precisely why getting it wrong is costly and why the cost surfaces so long after the change.
What to do with a second-column step
The instinct is to protect it by leaving it alone. That is usually wrong too. The work in that column is often genuinely painful, and refusing to touch it means accepting the pain forever in exchange for a benefit nobody has examined.
The better move is to separate the two products and treat them differently.
Automate the gathering. Finding the documents, pulling them together, extracting the figures, laying out what is present and what is missing: none of that is where judgment forms. It is the tax that has to be paid before the thinking starts, and paying it by hand is not a virtue.
Then keep the encounter. The person still moves through the assembled material, still sees what does not fit, still leaves with questions rather than a conclusion. What changed is that they spend that attention on the material instead of on collecting it.
The distinction, put as plainly as it goes: there is a difference between a system that hands someone a finished answer and one that walks them to it. The first is easier to build, demonstrates better, and is the correct design for most work. The second is harder precisely because the value was never in the output, so you cannot verify it by inspecting what came out.
What to take away
- Time the step, then ask separately what it produces besides its output. Those are two different questions and only one of them is usually asked.
- If the honest answer is nothing, automate it fully. Most work is here, and hesitating helps nobody.
- If something else does come out of it, split the step: automate the assembly, keep the encounter with the material, and make sure the person still arrives at the decision with questions rather than a conclusion.
The AI does the busywork. You still make the calls.