The response to the forward-deployed-executive post that stuck with me came from someone who'd lived the argument from the other side. Fine, he said, say you get the right person, real domain standing, everyone defers to them. How can you confirm that the recommendations were implemented? That's a fair question that the last post didn't answer.
Every AI project I've reviewed reports success the same way: a system got built, it runs, people use it, some metrics on accuracy and uptime. Almost none of them report what process steps were eliminated as a result and in there lies the problem. A tool that adds a faster way of performing an existing step and a tool that eliminates a step can be sold with an identical demo but deliver a different return, and the demo won't tell you which one you're looking at.
The transformations that actually generated returns, in the processes I've looked at closely, didn't add a faster version of the old step. They removed the old step and routed around it. Take, for example, double-entry bookkeeping, which didn't speed up reconciliation, it made a whole category of reconciliation error visible on sight. A modern ERP's value was never that it displayed the numbers faster than the spreadsheet it replaced, it was that it eliminated three roles dedicated to reconciling five spreadsheets that disagreed with each other. The pattern holds often enough to trust: the money is in subtraction, not acceleration.
Most AI projects do the opposite, and they do it effectively enough that nobody notices until its budget review season. The intake form stays. The compliance review stays. The supervisor sign-off stays. The data quality check against three source systems stays, except now a model does the scanning instead of a person. The final review by someone senior before it goes out stays too, because nobody wants to be the one who removed it when the time invariably comes that something goes wrong. What you get is AI plus everything else: the same five steps, one of them faster, all five still on the org chart and the cost center.
Take a version of this. Five steps: intake, compliance review, supervisor sign-off, a manual data quality check run against three source systems, and a final senior read before release. The AI automation lands on step four. The check that used to take a person forty minutes, cross-referencing three screens by hand, now takes six, because a model flags the likely mismatches and a person confirms them. Everyone is pleased. The dashboard says the step got 85% faster.
Nobody asked the two questions sitting next to that number. Why does reconciling three source systems require a manual step at all, when two of the three could feed the model's input pipeline directly and never produce a mismatch to reconcile? And why is a supervisor still signing off on work that the data quality check has cleared at a better than 98% rate for two years running? Answer both, and the five steps become three: intake, an automated reconciliation that only surfaces genuine exceptions, and the senior read. Not a faster version of the old process. A shorter one. Same outcome, reviewed by fewer people, in roughly a third of the elapsed time the old five-step version took — and that number came from redesigning the process, not from the model getting better.
That's two different projects wearing the same demo. One bought a faster middle step. The other deleted two of the five and left the model doing real work in what remained. From the outside, in a steering committee slide, they can look identical for months. The difference only shows up in headcount, cycle time, and whether anyone can explain why a step still exists.
The reason the first version wins by default has nothing to do with which one is harder to build. It's which one is easier to sell and easier to sit through. A faster step demos beautifully: same screens, visibly quicker, no one's job title changes, no one in the room has to explain to their own boss why a control they used to own is gone. Deleting a step demos as an absence. Nobody has ever put an absence in a slide, and the org chart doesn't reward the person who made three boxes disappear the way it rewards the person who shipped a new system.
This is where the standing argument from the last post actually gets spent. A forward-deployed executive earns the right to say a step shouldn't exist. Saying it once, in a workshop, costs nothing. Spending that standing means it shows up on a chart with two fewer rows than it had six months ago, on a project where somebody senior put their name against removing a control instead of adding a feature. If nobody's standing got spent that way, the engagement produced digitisation, whatever the invoice calls it.
There's a one-question test for telling which project you're funding, and it costs nothing to ask. What stopped happening. Not what got faster — what stopped. If the honest answer is a form, a review, a handoff, or a meeting, you've bought a redesign. If the honest answer is nothing, you've bought AI plus everything else, and the accuracy number on the dashboard was never going to tell you that.
That's the seventh pattern.