AI Passed the Professional Licensing Exam. So What's the Moat Now?

In 2026, Thomson Reuters put one of its AI tools through the U.S. Customs Broker License Exam – the actual licensing test for a regulated profession – and it passed. Not once, and not on a friendly subset. The tool passed all six publicly available exams administered between April 2023 and October 2025, across eighteen total runs, at a mean score above 84%.

Take a second with that. A credential that human professionals study for, sit for, and sometimes fail, got cleared by software at a comfortable margin, repeatedly, across years of changing rules.

The reflex reaction is the wrong one. The reflex is to read that sentence as "the AI can do the job now," feel a cold drop in your stomach, and start updating your résumé or your pricing. I'd push back on that, and not to be comforting. The exam was never the job. It was the door.

What the exam actually tested

A licensing exam measures one thing well: whether you know the material. It's a closed-book proof that the body of knowledge is in your head and you can retrieve it under pressure. That's genuinely useful – it's why we license people instead of trusting vibes – but it's a test of recall, and recall is exactly the thing machines have gotten frighteningly good at.

Here's what the exam doesn't test. It doesn't test what you do when a client's situation is specific and strange and doesn't map cleanly to any answer on the test. It doesn't test which of three technically-correct options is the right one for this client, this quarter, this risk tolerance. It doesn't test whether you'll put your name on the recommendation and stand behind it when someone asks why.

That's the job. The exam was the cost of admission to it.

Recall was never your moat

I work with two kinds of people who are both feeling this right now, and the shape is identical for both.

Boutique advisors – the independent consultant, the senior specialist running a small practice – have spent careers being the person who's read everything, tracked every rule change, absorbed the whole filing. And somewhere along the way it became tempting to believe that the reading was the value. It wasn't. The reading was overhead. What clients actually paid for was the judgment on the other side of it: which parts matter, what to ignore, what to do. The synthesis is leverage. The recommendation is the product. When I help an advisor set up managed intelligence, that's the line we protect on purpose – the machine and the back office handle the extraction and the synthesis, and the advisor keeps the judgment and the call. Not because the advisor can't do the synthesis. Because the synthesis was never the scarce part.

Software and product leaders are in the same spot from a different door. AI can now generate the code, draft the analysis, produce the first version of nearly anything. The instinct is to feel replaced. The reality is that "what should we build" and "is this actually right to ship" never lived in the generation step. They live in judgment, and judgment is still yours to carry.

Same pattern, both rooms. The thing the machine just got great at was never the thing you were actually selling.

The part that doesn't commoditize

There's a word doing quiet, heavy work in all of this, and it isn't "knowledge." It's accountability.

A license is not only a certificate that you know the material. It's a name on a line. It's a person a regulator, a client, or a court can hold responsible when the call goes wrong. When you sign off on a recommendation, you're not just transmitting an answer – you're accepting that the answer is yours, with consequences attached to your name.

An AI passing the exam transfers none of that. The model can produce the answer at 84%. It cannot be accountable for it. You can't escalate a bad outcome to a language model, and you can't ask one to care that it was wrong. The accountability stays with a human because that's the only place it's ever made sense to put it. That's not a temporary gap that the next model closes. It's a different category of thing.

So the honest read of "the AI passed the exam" isn't that your expertise just evaporated. It's that the commodity half of your expertise – the recall – finally got automated, the way arithmetic did, the way spell-check did, the way a dozen things you used to be paid for did. What's left is the half that was always the actual work.

What to do with this

Stop competing with AI on recall. You'll lose that race, you should lose it, and losing it frees you from a competition you never should have been in. The machine can hold the whole body of knowledge now. Let it.

Compete on the call. Get sharper about the judgment – the part where you look at a synthesized, AI-assisted, technically-correct stack of options and say "this one, here's why, and I'll own it." Price for that. Position for that. Build your practice or your team so that the recall is cheap and fast and the judgment is where your name shows up.

An AI can pass the exam. It can't sign its name to the answer and mean it. That signature – the judgment behind it, the accountability under it – is the thing you've actually been selling all along. The machine passing the test just made it obvious.

If you're rethinking where your real value sits as the recall gets automated out from under you, that's a conversation I have a lot. I'm at matthew@fieldway.org.