01 / Vertical Intelligence
AI that runs the whole function.
An autonomous worker for one regulated job. It does not sell software to the team that does the work. It does the work, carries the accountability, and is priced against what the work used to cost.
What it is
One regulated function, end to end.
Medical coding and clinical documentation. Claims adjudication. Insurance underwriting. Know-your-customer and anti-money-laundering review. Accounts payable, reconciliation and audit. Contract review and legal intake. Tax preparation. Loan servicing and collections. Each is a function with a large payroll, a rulebook, a record that someone must keep, and an outcome that someone must be accountable for. Each is now runnable end to end by a system that reads, decides, acts and documents.
The regulation is not an obstacle to the thesis. It is the thesis. Rules make the function legible enough to automate and consequential enough that the customer will pay for an accountable system rather than a tool. They also create the durable assets: the audit trail, the certification, the approval that took years to earn and that a better model does not hand to a competitor.
It is not software for the claims team. It is the claims team, with a system of record and an audit trail, priced per claim.
Why now
The whole task, at the quality bar, below the human cost.
Three things had to be true at once, and they are now true. Models can do the complete task, including the reading, the judgement and the write-up, at or above the accuracy of the person doing it today. The marginal cost of a decision is below the fully loaded cost of the human decision, with room to fall on every release. And buyers who own the cost of a function accept outcome pricing when the system is accountable for the outcome, because that is how they already buy from outsourcers.
The interesting part is the last few percent of cases: the exceptions, the appeals, the edge cases the rulebook did not anticipate. That is where a co-pilot hands back to a human, and where an autonomous system builds its moat, because handling them well requires the history, the controls and the domain knowledge that only accumulate by running the function.
Where value accrues
The record, the outcome data, the standing.
The system of record for the function
Once decisions are made and documented in the company's system, it is the thing the auditor, the regulator and the counterparty look at. It is replaced once a decade, if ever.
Outcome data with feedback
Every decision produces a result: the claim paid or denied, the appeal won or lost, the audit passed. That feedback loop is proprietary by construction. A competitor with the same model does not have it, and cannot buy it.
Regulatory standing
Audits passed, certifications held, approvals granted, and in some functions a license to operate. These are earned over years and they are the reason a buyer can hand the function over at all.
Distribution through the channel the buyer already trusts
Carriers, networks, associations and platforms that already sit between the function and its customers. Being the system those channels recommend is worth more than any feature.
The fork
Stronger, or erased.
Owns the outcome and the record. Closes most cases with no human, and gets the exceptions right because it has seen them before. Priced per decision against payroll. A better model widens its margin and raises its autonomy rate on the same day.
A co-pilot inside the incumbent's tool, a prompt over the existing system, a horizontal agent without the function's controls. Priced per seat against the software budget. A better model ships its feature for free, and the incumbent adds the rest.
What we ask
The questions the memo has to answer.
What fraction of cases close with no human?
Today, measured, and after the next model. The trend matters more than the level.
Who is accountable when it is wrong?
And what does the audit trail look like when the regulator asks. If the answer is the customer, it is a tool.
What does the buyer pay per decision, and what did the human cost?
The gap between the two is the company's margin and the customer's reason to switch.
What happens to the pricing conversation when the next model ships?
If the customer's first question is why the price is not lower, the company is selling capability, not outcomes.
Does the data improve the system, or just pass through it?
A flywheel accumulates. A pipe does not.
Which regulation is a moat, and which is a ceiling?
Some rules protect the incumbent system. Some require a human at the point of decision. We price both.
The bar by stage
What good looks like from first cheque to Series A.
Pre-seed
A function chosen for the size of its payroll and the density of its rules. A team that knows the function from the inside and can ship at the model frontier. A design partner who owns the P&L of the function and will let the system run real cases.
Seed
Real cases closing autonomously, with accuracy measured against the human baseline and the exceptions handled inside the product. Expansion across sites, lines or entities at existing customers. Unit economics that hold on today's model prices.
Series A
A repeatable sale to the owner of the function, annual recurring revenue in the low millions and growing, net revenue retention above roughly 110 to 120 percent, and an autonomy rate that widened since the seed.