AI consulting got crowded the same way crypto consulting did: a gold rush pulls in a hundred guides for every prospector. A lot of people calling themselves AI consultants right now learned the term last quarter, and what they sell is a ChatGPT subscription with a slide deck wrapped around it. Some of it's even useful. But you're paying builder prices for tour-guide work, and you can't tell the difference until the project's already underway.

So here's how to tell, from someone on the building side of the line.

The tell: ask what they've actually shipped that runs

Not "what have you advised on." What have you built that's running right now, that you'd put your name on, that someone uses.

A real AI builder has a trail of working systems and can show you how they work and where they broke. I can point at a model I fine-tuned for construction estimating, a screen-memory engine I open-sourced, a four-model setup I run daily, and the formal-verification rig I built to check my own work. Those aren't credentials. They're receipts. The point isn't the list — it's that a builder has a list, and a reseller changes the subject to "strategy" when you ask for one.

If the answer to "what have you built that runs" is a case study about someone else's tool, you've found a tour guide. Fine for a tour. Wrong for a build.

Why it matters: advice you can't build is just a wish

The thing nobody tells you about AI projects is that the gap between "here's what you should do" and "here's it working" is the entire job. Anyone can recommend "fine-tune a model on your data." Doing it means knowing how to assemble the data, pick the base model, run the training without lighting money on fire, evaluate whether it actually got better, and wire it into something your team uses. A consultant who's never crossed that gap can't tell you what's on the other side, so they keep you safe on the advice side, where the bill runs and nothing ships.

I learned to build before I learned to consult, in the order that makes the advice worth anything. Seven months ago I was a construction worker who taught himself to code by shipping real things, and I still build every week — that's not a side effect of the consulting, it's the qualification for it. I won't recommend a thing I couldn't build myself, which rules out most of what gets recommended.

What to ask before you hire anyone for AI

Three questions separate builders from resellers fast:

  • "Show me something you built that's running." Receipts, not slides. If they pivot to strategy, you have your answer.
  • "When is custom AI the wrong call for me?" A builder will happily tell you to just use ChatGPT when that's right — they're not threatened by it. A reseller can't, because then there's nothing to sell.
  • "Will you build it, or hand me a plan?" The plan is the easy 10%. Make sure someone's on the hook for the 90% that's actual engineering.

None of this means consulting is worthless — good advice from someone who's built is the most valuable thing in this field. It means the credential that matters isn't a certificate or a vocabulary. It's a working thing with your name on it.

I'm a builder first. If you want AI in your business, I'll tell you honestly whether you need a custom model or just a better prompt — and if you need the model, I'll build it, not hand you a deck. Tell me what you're trying to do and you'll get a straight answer either way.