Thesis
The work doesn't wait on a hire.
In the industries we build in, volume arrives whether or not anyone was hired to meet it. Claims come in. Faxes come in. Filings come due. Every one of them costs a unit of human attention, and the only response the industry has ever had is to find another person or let the backlog grow. We form companies on a closed-loop agentic platform and point it at that work.
The number that makes this a business
In American healthcare, roughly $740 billion a year goes to administrative services against about $63 billion to information technology. Eleven and a half dollars of people for every dollar of software.
That ratio is the thesis. It is not that these industries lack software — most of them bought some. It is that the software never touched the work. The work stayed with people, and the cost of the work stayed tied to how many people you could hire.
The same shape shows up everywhere we look. A specialist practice pays $15.99 to chase a claim status by hand and $5.40 to do it electronically. A mortgage servicer pays $176 a year on a loan that behaves and $1,573on one that doesn't — an eight-fold jump driven entirely by exceptions, not by loan size.
That gap between the routine case and the exception is where the whole opportunity sits.
Sources: Menlo Ventures, April 2026 · CAQH Index · MBA Servicing Operations Study, 2024 data.
The filter, applied
The scorecard
Each company, against the published tests. We have not adjusted the grid to make it look full.
| Company | Regulated | Document-heavy | Fragmented | Cost scales with volume | Market | The honest caveat |
|---|---|---|---|---|---|---|
| Currently | ●●●●● | ●●●●● | ●●●●● | ●●●●● | ●●●●● | Sells into a federal deadline it cannot slip, against incumbents who already hold payer contracts. |
| CNBS | ●●●●● | ●●●● | ●●●●● | ●●●● | ●●● | The licensed operator base is shrinking — seven consecutive quarterly declines. Fewer buyers each year, each with more to comply with. |
| Kora | ●●● | ●●●●● | ●●●●● | ●●●●● | ●● | Africa is about 0.6% of global collections. This only works as a global-catalogue play for African rightsholders, not an Africa-domestic one. |
| ZoomProp | ●● | ●●●● | ●●●●● | ●●●● | ●●●● | Regulation here is local and uneven rather than federal. No single deadline forces a purchase. |
| ClearInfo | ●●● | ●●●●● | ●●● | ●●●● | ●●●● | The least fragmented market we work in, which means the most direct competition. |
| Colony | ●● | ●●●● | ●●●● | ●●●● | ●●●● | The first seat is ZoomProp sales. The product has to prove it transfers to another company before it is a studio-wide layer. |
| Multiplying | ●● | ●●●●● | ●●●●● | ●●●●● | ●●●● | The clone is live. It is not yet a company with its own operator and raise. |
The argument against us
What would have to be true for this to be wrong.
These are the strongest objections we know, in the words the people making them use. We would rather put them here.
The moat isn't real
A Stanford Law analysis this June put it plainly: the model is not the moat, because the frontier models are interchangeable. Workflows can be rebuilt. Buyers run bake-offs. Our answer is that defensibility comes from operating inside a specific regulated corpus long enough to learn what it does when it goes wrong — which is exactly why we hold the runtime across companies instead of rebuilding it each time. It is also why we are careful not to claim that moat on day one of a new company.
Everyone will have agents by 2028
Gartner projects a third of enterprise applications will ship with agentic AI within two years, up from under one percent in 2024. If that is right, "agent-native" describes an industry norm rather than a thesis. We agree. The differentiation has to be which industries and what proprietary operating data — not the architecture.
Efficiency gains track layoffs, not AI
An analysis of ten companies' audited filings this August found that revenue-per-employee gains map to headcount cuts rather than to AI adoption. Salesforce is the case in point: agents handle roughly half of support interactions and headcount still rose. It is a fair challenge and it is the reason we measure per workflow rather than per company, and publish the definition of what we counted.
Most of this fails in deployment
Gartner expects over 40% of agentic projects to be cancelled by the end of 2027. MIT's NANDA study found 95% of organisations getting no return. We take both seriously — and note that NANDA's own data says back-office work returned more than front-office work while receiving half the budget. Which is where we build.