Every major AI company right now is racing to build a better copilot.
Better writing assistance. Better code suggestions. Better research summaries. The pitch is always the same: we make you faster at your job. Buy a seat. Subscribe per user. Ship features that win the benchmark.
I think this is the wrong race. Not because copilots aren't valuable — they are. But because there's a different, more durable, more defensible business sitting right in front of the industry. And almost nobody is building it.
The Question Nobody Is Asking
Here's the question most AI founders aren't asking: who actually wants the work done, versus who wants a better way to do it themselves?
These are different buyers with different problems and radically different willingness to pay.
The professional — the accountant, the insurance adjuster, the revenue cycle specialist — wants a faster workflow. They buy copilots. They pay seat prices.
The business that hired the professional doesn't care about the workflow. They care about the outcome. The claim processed. The books closed. The invoice collected. They're already paying for that outcome — to a vendor, to a staffing firm, to an outsourced service provider.
That's a completely different market. And it's bigger.
Intelligence vs. Judgement
Before you can understand why this matters, you need a clean distinction.
Knowledge work runs on two things that look similar but aren't.
Intelligence is rules-based and repeatable. Processing a standard insurance claim. Extracting line items from a contract. Classifying a diagnosis code. Drafting a routine legal filing. The input-output relationship is learnable. Given enough examples, a model can match or beat human performance — and do it at a fraction of the cost.
Judgement is something different. It's the senior underwriter who reads a risk submission and senses something is off before they can articulate why. The litigator who knows when to settle versus fight. The consultant who understands the real politics behind a client's stated problem. Judgement is accumulated pattern recognition that hasn't been formalized yet — and may never fully be.
Most knowledge work sits closer to the intelligence end than professionals like to admit.
The percentage of a claims adjuster's day that requires genuine judgement — novel situations, ethical edge cases, genuine ambiguity — is smaller than their job title implies. The rest is intelligence work. Repeatable. Learnable. Automatable.
That's the wedge.
Copilot vs. Autopilot
Here's where I think the industry has made a fundamental strategic error.
A copilot sells the tool to the professional. The professional does the work; AI makes them faster. You capture a fraction of the efficiency gain as subscription revenue. And here's the uncomfortable dynamic: every time the foundation model gets better, you have a slightly improved product — but so does your competitor, and so does the next startup that launches tomorrow with the same underlying model.
Model improvements commoditize copilots. They're a treadmill.
An autopilot sells the outcome to the buyer. Not a faster accountant — just closed books. Not a better claims adjuster — just processed claims. The buyer doesn't care what's inside the box. They care that the output comes out correctly and on time.
The economics are inverted. Every time the foundation model gets better, an autopilot company's margins improve — because the same outcome costs less to produce. Model progress becomes a structural tailwind instead of a competitive threat.
That asymmetry is enormous. It's the difference between building a business that fights the tide and building one that rides it.
Why Outsourced Work Is the Right Starting Point
The smartest wedge into this market isn't to disrupt a company's internal team. It's to replace an external vendor.
Outsourced, intelligence-heavy service categories are ideal: insurance brokerage, accounting and bookkeeping, healthcare revenue cycle management, claims adjusting, tax preparation, legal operations, IT managed services, recruitment, back-office consulting. In every one of these, the same structure holds.
The buyer already has a budget line for this outcome. They already don't care who delivers it — it's outsourced precisely because they'd rather not think about it. The current vendor is often a services firm running on expensive human labour, with thin margins and high turnover. Switching is a vendor decision, not a reorganisation.
You're not asking a company to change how they work. You're asking them to sign a different contract for the same deliverable at a lower price with better consistency.
That is a much easier conversation.
The verticals where this plays out first are the ones with high volume, standardised inputs, and measurable outputs. Healthcare revenue cycle — where billing accuracy directly affects cash collections — is already seeing this. Accounting close processes. Insurance claims with binary outcomes. Tax return preparation for standard situations.
These aren't glamorous. They're enormous.
The Trap Facing Today's Copilot Companies
Here's what keeps me up about the current landscape.
Most of the well-funded AI companies today have built their distribution through the professional. The accountant adopted the tool. The lawyer uses it daily. The adjuster's workflow depends on it.
To shift to autopilot — to sell the outcome directly to the CFO and bypass the professional entirely — these companies would need to disintermediate the exact customer who made them successful.
That's the innovator's dilemma in slow motion. The product that got you here is incompatible with the product you need to build next. The early customers who championed you internally are now the ones blocking your evolution.
New entrants don't carry this baggage. They can go straight to the buyer, never touch the professional, and compete on a fundamentally different value proposition. They can price against outsourced headcount — which is a much higher ceiling than per-seat SaaS — and improve margins automatically as models improve.
What This Means
The next genuinely transformative AI company won't announce a new model. It won't win a benchmark. It might not even look like a technology company from the outside.
It will look like a services firm. It will quote you a price per claim, per filing, per close cycle. It will sign outcome-based contracts and carry the performance risk. It will run on a fraction of the headcount of the firm it replaced, with margins that compound as the underlying intelligence gets cheaper.
On the inside, it will run like a software company. Every process encoded. Every edge case documented. Every improvement shared across all clients simultaneously. The leverage that software always promised — but applied to work, not workflows.
This is what software eats next. Not the tool. The job itself.
The question worth asking, if you're building something in AI right now: are you selling access to intelligence, or are you selling the outcome that intelligence produces?
One of those is a subscription. The other is a business.
The insight here draws significantly on Sequoia Capital's framing in "Services: The New Software" by Julien Bek (March 2026) — one of the clearest articulations I've read of where the real AI opportunity sits.