I Turned Down Work Because They Only Wanted the AI
Earlier this year I walked away from a client, and it was real money, the kind that would have made a quarter look tidy. I've thought about it more than I expected to.
They came in well prepared, which is part of why it stuck with me. They had a list. Actual processes, named, with rough time estimates attached, all of them genuinely painful and all of them things I could have built against. On paper it was the easiest engagement I'd been offered all year.
The problem showed up when I started asking the other questions. What would be different about how the leadership team spends its week. Which of these steps exists because the work needs it, and which exists because a person used to be required and nobody took the scaffolding down afterward. What happens to this role if the retyping part disappears. I got answers, they were polite answers, and what they added up to, in several different phrasings, was that nothing about how they worked needed to change. They wanted the automation laid on top of the company exactly as it currently runs.
So I passed. And I want to be careful here, because that's an easy thing to say in a blog post and a less comfortable thing to do on a Thursday when you're looking at your own pipeline.
A year earlier I'd have taken it and been thrilled
This is the part I have to own. For a long stretch, that engagement was exactly what I thought my job was. Clean, well-scoped agents. Something that watches a folder, reads the document, pulls the fields, writes the summary, files it where it belongs. Real work, real hours saved, and I got decently good at building them.
Clients were pleasant about it. That's the tell I missed for longer than I'd like. Pleasant is what people are when a thing is fine and nothing in their week has actually moved. The agent ran, it did what it said, somebody glanced at the output on Tuesdays, and the company operated in precisely the shape it had operated in the month before, with one fewer irritating task inside it.
What I'd been doing was automating the edges of a shape nobody had questioned. The shape is the whole thing. Workflows, where the data sits and who can reach it, who reports to whom and what they're accountable for: all of that was designed for a world where a human had to touch every step. Hanging an agent off the side of that design doesn't change the design. It makes one step cheaper while every assumption underneath stays exactly where it was.
Where it actually goes somewhere
The engagements where something real happened share one feature, and it isn't the technology. A senior person started using the tools themselves. Not approved a pilot. Used them, personally, in their own work, badly at first.
Somebody who hasn't written code since college writing a query to answer a question they'd been routing through an analyst for two years. Somebody building a small database for themselves after realizing that nobody else was going to build it and, as of about six months ago, they no longer need anyone to. The output usually isn't impressive. That's fine, because the output isn't the product. The product is that a person with actual authority has now done something they'd filed under impossible-for-me, and once that happens they stop asking whether AI can help and start asking what their team should even look like.
That's the value. Not the agent. The executive who now has enough direct contact with the thing to rethink the org around it.
Which is why the company I turned down was a bad fit rather than a bad company. They were smart and their pain was legitimate. But they were buying a layer, and a layer would have run fine and changed nothing, and eighteen months later somebody would have reasonably concluded that this AI stuff doesn't do much. I'd rather not be the reason for that conclusion.
Why so many get stuck at the first stop
Most organizations I talk to are still writing emails with it. Nearly four years in. The technology went through several genuinely distinct eras in that window, and plenty of competent companies are still on the first one, using a system that can work unattended for an hour to make a paragraph sound more polished.
I don't think that's stupidity and I don't think it's fear of the tools. The first stage is the only one that asks nobody to change anything. Faster emails slot neatly into the existing shape of the job. Everything past that point starts raising questions with political weight: why does this approval exist, why does this data live in four systems, why does this role spend most of its week moving information between them, what is this team for if the moving part goes away.
Nobody has a calendar slot labeled “question the premise of my department.” So the safe version gets done, it gets called an AI initiative, and there it sits.
The job I actually do
I put fractional Chief AI Officer on the website, which is accurate and a little misleading. Most weeks the work is closer to coaching. I'm not there to install something. I'm there to keep asking the annoying question until the leadership team answers it themselves, and then to build against the answer.
There's a smaller version of the same problem in my own week, for what it's worth. Something lands that I could finish in eleven minutes. Handing it off properly, with enough context to come back right, takes maybe fifteen. I lose that decision constantly and do it myself, feeling efficient, having chosen the worse option again. The eleven-minute path leaves nothing behind. The fifteen-minute path leaves a written description of how the task actually works, which I did not have before and can use forever. I say this to clients and then fail at it on a Tuesday. It's a habit, and habits don't respond to demos.
Do it ten times and something shifts. You stop asking what AI can do and start noticing how much of your week is structure that outlived its reason. That's the real starting line, and most companies haven't found it yet, which is either a problem or the best news you'll hear this quarter, depending on where you're standing.
