Owning the outcome: Bessemer's AI-Native Services evaluation framework

1. Is this a market that's structurally ready to be taken — fragmented supply, incumbents who can't respond, essential work, and demand that expands rather than shrinks when AI collapses the price?
2. Can AI actually do the work at software-like margins — and can you keep the surplus rather than competing it away?
3. Once you've won the work, does anything stop it from leaving — recurring revenue, compounding data, or a regulatory moat that scales with you rather than against you?

Why this wave of services is different

The cloud era produced many tech enabled services companies that made a real dent in their industries, but these companies largely focused on delivering exceptional customer experiences. LegalZoom digitized the front door of consumer legal services, but the work of law is still priced and delivered the way it always was. Lemonade won over customers but never proved it had fixed the carrier cost structure. Compass became the largest residential brokerage in the country and is still, at its core, a brokerage. Ultimately the cloud era never changed the underlying delivery economics of services, despite its digitization of the customer experience. Long horizon agents flip that equation, increasingly completing hours-long tasks with the only variable cost being that of inference. This has far reached effects beyond gross margin, which we explore in our framework below.

1. The first is multimodality, on both the action side and the input side. Browser agents can now operate portals and legacy systems they were never trained on, with real-desktop benchmark success rising from roughly 12% to within a few points of human performance. Frontier models are now pre-trained and fine-tuned to map pixels directly to precise UI coordinates, replacing the brittle DOM selectors and screen-scraping heuristics of RPA and the error-prone click estimation of 2024-era agents. Browser agents are also now trained end-to-end on verified task completion rather than next-token imitation, which instills self-correction if a click misfires or a page changes. On the voice side, speech-to-speech models are gaining frontier-level reasoning at conversational latency, skipping the transcribe-then-synthesize step entirely. Amperos, an AI biller for healthcare providers, runs on both tailwinds, navigating payer portals and phoning insurers to work denials no clinic can afford to staff. The same progress applies to inputs, since services work arrives as scanned faxes and hundred-page PDFs rather than clean API payloads. EvenUp, an AI demand package writer for plaintiff firms, ingests thousands of pages of scanned medical records and billing tables per case—work that improves in cost and accuracy with every gain in frontier vision models. Unlimited Industries, an AI-native civil engineering firm acting as engineer of record, benefits from the same multimodal input progress, turning the document-heavy front end of site design into stamped design packs.
2. The second force is that the post-training and eval stack has matured into products sold off the shelf, letting small teams tune systems on their own production data. Much of the frontier's recent gain comes from reinforcement learning on verifiable rewards rather than bigger pretraining, and the scarce input has become the RL environment, a simulated workplace with a checkable reward that labs pay big bucks to acquire. Services firms hold a structural advantage here, because their output is verifiable (i.e., a claim pays or it doesn't), so every delivered unit of work doubles as a training signal, where the firm gets its environment for free (and can keep as alpha) as a byproduct of doing the work. Open frameworks (Trinity-RFT, Thinking Machines' Tinker) extend this to companies that want weights and IP in-house, wiring internal datasets and simulators into their own RL environments. Strala, an AI-native claims administrator, captures every adjuster correction as labeled data and tunes its system until entire classes of misses stop recurring. Crosby, an AI-native law firm for commercial contracts, has its own lawyers design the test sets and ship releases until the models beat their redlines.
3. An emerging shift worth watching is how open-weight models now trail the closed frontier by roughly one model generation at a fraction of the cost. Few AI services firms have made this part of their story yet, as it’s still early days, but we expect the best firms to defend gross margin by routing routine volume to tuned open models and reserving frontier models for the hardest cases. 

Bessemer's framework for evaluating AI-services opportunities

AI will impact services work unevenly. These markets are changing quickly, and the old guard of TAM and market growth doesn’t sufficiently capture the dynamic nature of AI services. Some markets are structurally advantaged to incorporate AI through a high automation; others possess characteristics of Jevon’s Paradox, where the abundance of the cheap services will massively increase the market size; and others have incumbents whose innovator’s dilemma make it a more compelling market to win share. The best markets spike on all of these. 

Agent availability doesn’t necessarily determine the market opportunity

In the cloud era, the technology of “tech enabled services” sat in the workflow or experience layer, while the delivery layer stayed human, but in this wave the technology layer impacts the delivery mechanism itself. But the availability of agents is not the same thing as market opportunity. Some markets expand dramatically when the price of the service collapses, because enormous latent demand was always sitting there unserved at human price points.