Everyone Is AI-First. Almost Nobody Is.
I sit in a few hundred enterprise meetings a year. The gap between what companies say about AI and what they've actually deployed is the widest I've seen for any technology cycle.
Every company I meet with is AI-first now. It’s in the earnings call, it’s on the careers page, it’s the first slide of the deck the CIO’s team built for the board.
Then you ask what’s in production and the room gets quiet.
I’m not being cynical here. I’m describing an actual measurement problem. There’s a real transformation happening and there’s an enormous amount of theater layered on top of it, and from the outside they look identical. Both produce press releases. Only one produces a P&L change.
The tell is always the same
The question I ask now is boring on purpose: who owns the budget line, and what were they doing with that money last year?
If AI spend came out of the innovation budget — the discretionary pool that exists to generate announcements — it’s theater. That budget has always existed. It funded blockchain pilots in 2018 and metaverse storefronts in 2022 and it will fund something else in 2028. The dollars rotate, the outcomes never arrive.
If AI spend came out of an operating budget — a support org’s headcount plan, a marketing agency line, a claims-processing vendor contract — somebody has decided this thing replaces a cost they were already paying. That’s real. That’s someone putting their number on the line.
The first kind of project gets a press release. The second kind gets a renewal.
Why the gap is so wide this time
Every hype cycle has a say-do gap. This one is unusually large for three reasons.
The demo is free and the deployment isn’t. Anyone can put a model in front of a customer service transcript and get something that looks like magic in an afternoon. Getting that same thing to be correct 99.5% of the time, on your actual data, with auditability, in a regulated workflow, is a two-year program with a real budget. Executives saw the afternoon and priced the two-year program accordingly.
The pilot is the deliverable. In most large companies, the person who launches the AI initiative is not the person who has to run it in year three. They get promoted on the launch. This is the incentive structure of an org that produces a lot of pilots and very few systems, and it’s nobody’s fault individually — it’s just what happens when you reward the announcement.
Nobody wants to be the company that says it’s going slow. There is enormous pressure to have an answer on the earnings call. So you get language engineered to be technically true and directionally misleading. “We’ve deployed AI across the enterprise” can mean nine thousand seats of a chat assistant that people use to rewrite emails. That’s fine! It’s just not a transformation.
It’s the same thing as a team announcing a new offensive philosophy in June. Cool. Let’s see it on third and long in November.
What actually working looks like
The companies I’ve seen genuinely get value share a pattern, and it’s less impressive than the keynote version.
They picked one workflow that was already measured. Not “customer experience” — first-contact resolution rate in one support queue. Not “developer productivity” — time-to-merge on one repo. Something with a number that existed before the project started, so there’s no argument later about whether it moved.
They kept a human in the loop far longer than the vendor recommended, and they were fine with that, because the goal was throughput, not headcount elimination.
And they treated the data plumbing as the actual project. That’s the unglamorous truth of this whole cycle: most AI programs are data-integration programs wearing a better outfit. The model is commoditizing fast. Your permissions model, your document retention mess, and the fact that four systems disagree about who your customer is — that’s the work, and it’s the same work it was ten years ago.
The model isn’t your moat. The three years you spent making your data coherent is the moat, and the model is what finally makes it pay off.
Where I could be wrong
Two ways.
One: the tooling gets good enough fast enough that the integration work I’m calling the moat gets commoditized too. If agents can navigate a messy enterprise the way a competent new hire does — badly at first, then fine — then the companies with clean data lose their advantage and everybody starts from roughly the same line. I don’t think that’s the next twelve months. I’m not confident about the next thirty-six.
Two: I’m sampling on the wrong population. I talk to companies that are actively buying infrastructure, which skews toward organizations that have decided to do something. The genuinely native-AI companies may be building without ever taking a meeting like mine, and I’d never see them.
What would change my mind: a wave of large enterprises reporting AI-attributable margin expansion in their actual filings — not in the keynote, in the 10-K, with the line item traceable. That’s the number I’m watching. When operating margin moves and the CFO is willing to attribute it on the record, the theater is over.
Until then, I’d ask a simpler question in every one of these meetings: what did you stop paying for?