The consultant who never ran payroll

Here’s the pattern. A consultant arrives with a deck. The deck has a vision: reinvent the customer journey, transform the operating model, become an “AI-native company.” A year later there’s an impressive system nobody uses, an invoice that could have bought a truck, and — in the worst cases — an AI making decisions in a business its builder never understood. Running amok is the technical term.

What most AI consultants are missing

It’s not intelligence and it’s not technical skill. Plenty of consultants have both. What they don’t have is scar tissue. They’ve never personally covered payroll in a slow month, never had a customer emergency at 2am, never been the last line of defense when a system failed. So they literally cannot know what the 3am list feels like — and the 3am list is where all the best AI projects live.

There’s a second craft they’re missing, and it’s the harder one: managing change. The code was never the hard part — getting real people in a real business to trust and adopt a new way of working is. And knowing how to manage change isn’t found in a YouTube video, or even a year of classes at a university. It’s top-shelf education, updated continually, honed in the furnace of real-world experience. I’ve written about how we actually do it — gently — starting with change without the growing pains.

When you don’t know where the pain is, you guess. And when you guess, you guess big, because big is what justifies the engagement. The incentives all point the same way: a grand redesign bills more than a watchdog, demos better than an invoice matcher, and sounds better in the case study. The only party ill-served is the business.

The tell

You can spot the problem in the first meeting. Count how many minutes pass before you’re asked what keeps you up at night — not as small talk, but as the actual agenda. A builder who’s been an owner opens there, because that’s where the value is. A builder who hasn’t opens with the technology, because that’s all they’ve got.

The industry data tells you exactly how this movie ends. RAND interviewed 65 veteran data scientists and concluded that by some estimates more than 80% of AI projects fail — twice the failure rate of ordinary IT projects — and the number-one root cause is stakeholders misunderstanding, or miscommunicating, what problem needs to be solved. Read that again. The leading killer of AI projects isn’t the technology; it’s not understanding the business. MIT’s 2025 State of AI in Business found the same from the enterprise side: 95% of pilots returned nothing, largely because the tools never learned the real workflow.

Now look at who builds the failures. Usually it’s people who know computers and AI cold — often genuinely brilliant — who have never run a business. Never set a price, never made payroll, never lost sleep over a customer they might be losing. They can make the technology do anything; they just don’t know what it should do, because they’ve never lived inside an operation. That 80% isn’t a technology gap. It’s an experience gap — and it’s why the same tools that run amok in inexperienced hands quietly make money in experienced ones.

And breadth matters as much as depth. I’ve done this work at every altitude — five-person shops where the owner is also the repairman, and the upper echelons of a Fortune 500, presenting to executives and the board. Here’s what that range teaches you: the failure patterns are identical at both ends, and so are the fixes. Someone who has only ever seen the technology side hasn’t seen either end of the business — and it shows in what they build. I broke down the habits of the projects that do pay in an earlier note.

What to demand instead

Whoever you hire — me or anyone — hold them to three things.

They start with your pain, not their platform. The first deliverable should address something on your 3am list or your time-tax audit. If the first deliverable is a strategy document, you’ve bought a very expensive bookmark.

They ship something small in weeks. Working, in production, on your real data. You should be judging results, not slideware, within a month.

They stay accountable for it running. Monitoring, alerts, a human who answers. An automation abandoned after launch is a liability with your name on it.

None of this is anti-ambition — the grand ideas have their turn, and part five is about exactly when. It’s anti-starting with them. The world doesn’t need reinventing. Your Tuesday does.