Services sold as software
Software sold seats to people. The next decade sells the finished work, and that budget is labor: more than ten times what the world spends on software.
Neil Gaikwad
The size of the prize
What the consensus misses
The consensus story is that agents replace software seats. I think that undersells it. Seats were never the prize. The prize is the work itself: the claim processed, the return filed, the ticket resolved. Buyers already pay for that work, through salaries and outsourcing contracts, at more than ten times what they spend on software.
The second thing the market gets wrong is where the money settles. Horizontal agents are being priced toward zero by the platforms that own the customer: Salesforce lists agent work at $2 a conversation or about ten cents an action. The durable margin sits with whoever takes responsibility for an outcome in a regulated domain, where being wrong costs real money and someone has to sign for it.
And the most interesting companies may not look like software companies at all. General Catalyst's Long Lake has bought 40 services businesses since 2024. Crete, backed by Thrive, runs more than $300M of accounting revenue through firms it acquired. Buying the firm and running it with agents gets around the slowest part of this shift, which is small firms that do not adopt new tools on their own.
Why now
Agents keep getting longer leashes. METR finds the length of task an agent can finish has doubled roughly every seven months for six years, and the best models now hold up for over two hours of work.
The cost of a given level of capability falls by roughly 40 times a year on Epoch's measure. Work that was too expensive to automate last year is cheap enough this year.
Pricing caught up with the product. Intercom charges $0.99 per resolved outcome, Sierra prices on completed work, and buyers can now compare an agent's price directly with the person or vendor it replaces.
Where the value pools
| Layer | My call | Why |
|---|---|---|
| Frontier models | Concentrates | A handful of labs with the capital to train at the frontier. Not a seed-stage game. |
| Horizontal agents | Commoditizes | Customer service and generic assistants are bundled by the platforms that own the customer, at cents per action. |
| Vertical agents that own an outcome | Most value | Legal, accounting, claims and health admin, where accuracy carries liability and the buyer pays per result. |
| AI-native services firms and roll-ups | Durable value | They own the customer relationship and the labor budget, and capture the margin agents create. |
| Agent infrastructure | Underpriced | Evals, identity, payments and audit for agents are early, with small volumes and no clear leaders yet. |
Who is winning so far
| Company | What it does | Stage and last round |
|---|---|---|
| Cognition | Autonomous software engineering agent; owns the Windsurf editor | Over $2B Series E at $48B, led by a16z and Accel Sep 2026 · cognition.com |
| Sierra | Customer experience agents priced on outcomes; $150M+ ARR | $950M Series E at $15B+, led by Tiger Global and GV May 2026 · dealroom.co |
| Harvey | Legal agents; used by 80% of the Am Law 100 | $550M at $15.5B, co-led by Diffusion and Lightspeed Sep 2026 · harvey.ai |
| Decagon | Customer service agents for airlines, banks and telcos | $250M Series D at $4.5B, led by Coatue and Index Jan 2026 · texxr.com |
| EvenUp | Agents for personal injury firms; 2,000+ firms, 200,000+ cases | $150M Series E at $2B+, led by Bessemer Oct 2025 · fortune.com |
| Basis | Agents for accounting firms; about 30% of the Top 25 firms | $100M Series B at $1.15B, led by Accel Feb 2026 · pulse2.com |
| Tennr | Automates patient referral intake for providers | $101M Series C at $605M, led by IVP Jun 2025 · multiples.vc |
| Crete | AI accounting roll-up; $300M+ revenue across 20+ firms | Thrive-backed; plans $500M+ of acquisitions Jun 2025 · finance.yahoo.com |
Where I would write a seed check
Claims, compliance filings, billing appeals and permits still have no category leader. EvenUp and Basis show the path from seed.
Early examples: EvenUp, Basis, TennrThe smallest firms adopt slowest. Buying them and running them on agents moves the margin to the owner of the model.
Early examples: Long Lake, CreteGartner expects over 40% of agent projects to be cancelled by 2027, much of it for weak risk controls. Evals, certification and agent identity are the fix buyers will pay for.
Early examples: AIUC-1 certificationOpen standards exist but volumes are small, about $24M in 30 days on x402. Billing and settlement for agent labor is unclaimed ground.
Early examples: x402 FoundationHow it plays out
- 2027
The shakeout. Gartner's predicted cancellation wave clears out thin agent startups; outcome-priced vendors with references survive.
- 2028
Roll-ups publish real margins. If AI-run firms show audited margin expansion, capital floods into buying small services firms.
- 2031
My expectation: agents handle multi-week professional tasks end to end, and pricing per outcome is the default in legal, accounting and support.
- 2036
My expectation: the largest professional services firms by revenue include several that did not exist in 2024.
As of Oct 2026
Risks
- MIT NANDA reported 95% of organizations seeing no business return on enterprise gen AI spending in 2025. virtualizationreview.com
- Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027, and estimates only about 130 of thousands of agent vendors are real. outlookbusiness.com
- Incumbents cap prices: Agentforce lists at $2 per conversation, with a free starter tier. salesforce.com
- Valuations price in years of growth: Sierra at $15B+ on $150M of ARR. dealroom.co
- Labor backlash is building: employment for 22 to 25 year olds in AI-exposed jobs is 19% below trend. digitaleconomy.stanford.edu
Sources
- General Catalyst, The future of services (Aug 2025)
- General Catalyst, Long Lake (Sep 2026)
- Foundation Capital, Service as software
- FRED, Value added by professional and business services
- Fortune on McKinsey, Agents, robots and us (Nov 2025)
- METR, Measuring AI ability to complete long tasks
- Epoch AI, LLM inference price trends
- Intercom Fin pricing
- Salesforce Agentforce pricing
- Cognition, Series E
- US Census, AI use by businesses (May 2026)
- Stanford Digital Economy Lab, Canaries in the coal mine
Nothing here is investment advice: these are my own notes on public information, written to sharpen judgment, not to recommend buying or selling anything.