The boring part is the whole business

Everyone’s building the wrong agent.

The exciting demos, agents that write poems, agents that trade, agents that pretend to be your friend, are not where the money or the difficulty is. The real opportunity is the most boring sentence in tech: agents that do compliance-gated paperwork in regulated industries.

I want to walk through why I believe this, because it’s not obvious and it cuts against where most of the excitement is pointed. It comes down to economics first, and then to a deeper point about where the actual difficulty in these industries has always lived.

The economics force it

Start with a brutal fact people gloss over: a capable model running continuously is expensive. If you want it genuinely good, always-on, and working on real problems, you’re looking at real money per agent per month, not pennies. That number matters enormously, because it determines who can possibly be the customer.

Consumer subscriptions can’t cover it. The big labs already give you near-magic for twenty or thirty dollars a month, subsidized, at a scale and price nobody can undercut. If your business model is “a slightly better consumer assistant,” you are competing on price against companies that can run inference cheaper than you’ll ever manage and are willing to lose money to own the relationship. That’s a knife fight in a market with no margin.

So who can pay enough to make the unit economics work? The customers who genuinely can’t use an off-the-shelf consumer tool. And that turns out to be a very specific, very valuable set: finance, healthcare, insurance, the industries where compliance is the blocker, not capability. Where the work is high-stakes, judgment-laden, and legally fraught, and where a human currently has to sit in the loop because the rules don’t quite fit the case and someone has to be accountable.

That’s not a coincidence, and it’s not a consolation prize. That’s the market. The expensive-to-run nature of capable agents pushes you, almost deterministically, toward the customers for whom the value is high enough and the alternative is unavailable. The economics don’t permit the consumer dream. They do permit the enterprise-compliance one.

Compliance is the product, not the hygiene

Here’s the deeper part, the part people miss even when they’ve accepted the economics. In these industries, the hard 80% was never the technology. It was the paperwork, the review, the audit trail, the regulator who needs to be satisfied, the liability that has to land somewhere. The reason these jobs resisted automation isn’t that they’re intellectually hard. It’s that they’re fuzzy and high-consequence, the exact combination that rules-based software has always choked on.

Most software companies treat compliance as a tax, the annoying thing you bolt on at the end so the lawyers stop emailing you. SOC 2, HIPAA, the transfer-agent dance, the audit logging, all of it filed under “cost of doing business,” a grudging overhead.

For an agent company serving regulated industries, the compliance isn’t the cost of doing business. It is the business. The audit trail, the bounded authority, the provable-safety story, that’s the product. The moat is the paperwork.

Flip the framing and everything changes. The thing that makes an agent valuable to a hospital or a bank or an insurer is not that it’s clever. Clever is cheap now. It’s that it can act inside the constraints, with bounded authority, a perfect record of every decision, and a clean escalation path when it genuinely should escalate. An agent that can do that is worth far more than a thousand chatbots, because it’s a new kind of worker for the exact jobs that were too regulated to automate, and there’s nobody undercutting you on price, because the consumer labs can’t touch this market without becoming a different kind of company entirely.

Why this is my lane

I’ve spent years in the back office of capital formation, and I’ve crossed paths with the same fundamental problem from the finance side, tokenization, real-world assets, the grunt work of moving money under rules that don’t bend. I know, in my hands, why this layer is hard and why it stayed manual for so long.

And I now build by directing agents every day, so I have a current, visceral sense of what these things can actually do versus what the hype claims. That combination, deep familiarity with a regulated back office and daily practice at the frontier of agentic building, is rarer than it should be. Most people who understand compliance don’t build with agents. Most people building with agents have never sat with a regulator. The bridge between “agents that can act” and “industries that couldn’t automate” runs right through the exact spot where I’ve spent my career.

It’s the least glamorous frontier in AI. No viral demo, no friendly companion, no poem. Just the unsexy, enormously valuable work of letting regulated industries finally automate the part that’s always needed a careful human. I think it’s the most valuable frontier there is, precisely because it’s unglamorous, the glamorous stuff is crowded and margin-free, and the boring stuff is wide open and load-bearing.

I’d rather build the unsexy thing that’s true than the sexy thing that isn’t. That’s the whole thesis, and it’s where I’m putting my time.