In May 2026, Cognition AI raised a billion dollars at a $26 billion valuation. Eighteen months earlier, the company was worth $2 billion. That is a 13x jump in under two years, for a company that sells software engineering work, not software engineering tools.
Cognition's product, Devin, is an AI agent that takes an engineering ticket, works on it, and returns a pull request. Customers pay for the completed work, not for a tool to help them do it themselves. Sequoia Capital calls this "Services: The New Software," and it is the single most aggressive investment thesis in venture capital right now.
The thesis inverts the SaaS model. SaaS means you rent software and do your own work. Service as Software means the software does the work and you pay for the outcome. The framing comes from three sources: Sarah Tavel at Benchmark named it "sell work, not software" in August 2023. Julien Bek at Sequoia gave it the structural framework in March 2026. Foundation Capital sized the market at $4.6 trillion.
Why $4.6 trillion? Because the global services economy (salaries, outsourcing, professional fees, administrative work) is roughly twelve times the size of the SaaS market. Bek's line is blunt: for every dollar spent on software, six are spent on services. A company that delivers services through software competes for the operations budget, not the software budget. The ceiling is ten times higher.
The proof is in the valuations
You can tell a thesis has moved from whiteboard to deployment when the numbers get violent. Six service-as-software companies have hit unicorn status or better in the last eighteen months, and the valuation escalation is steep:
- Harvey (legal AI): $3B to $11B, $300M ARR, 290% growth, 142,000 lawyers on the platform.
- Sierra (customer experience AI): $4.5B to $15.8B, $200M ARR, 400% growth, 40% of the Fortune 50 as clients.
- Cognition/Devin (AI software engineering): $2B to $26B, $492M annualised revenue.
- Basis (AI accounting): seed to $1.15B in three years, 30% of the Top 25 accounting firms as customers.
- Crosby (AI law firm): did not exist in September 2024, $400M valuation by March 2026.
- EvenUp (AI personal injury claims): $1B to $2B+, 200,000 cases resolved, $10B in damages secured for victims.
These aren't SaaS multiples. SaaS companies trade at around 7x ARR. Euclid Ventures points out that service businesses trade at 3x revenue. But the valuations above are being paid for companies that structurally resemble service businesses: pricing by outcome, bearing delivery cost, employing humans for oversight. The market is pricing them like software companies anyway, because the growth rates look like software and the TAM looks like nothing venture has seen before.
VCs are now buying companies to do this
The thesis has hardened into a deployment strategy. In June 2026, CNBC reported that venture capital firms are buying legacy services companies outright and rebuilding them around AI. General Catalyst has co-created a dozen rollup vehicles since 2023. Thrive Capital runs Thrive Holdings with over a billion dollars in capital. Long Lake Management, three years old, has acquired 30+ businesses across HOA management, construction, and corporate travel, and plans to hold them permanently: a Berkshire Hathaway model run on AI.
This puts traditional private equity on the defensive. PE firms buy services companies to optimise margins through cost reduction and consolidation. VC-backed rollups buy the same companies to rebuild their operations around AI agents, which is a different value proposition: the cost structure compounds downward as models improve, not upward as labour costs rise.
PitchBook's data confirms the trend. Service as Software is now the dominant framing for new enterprise investment. Healthtech funding more than doubled in Q4 2025. Cybersecurity hit a record high. EY and Info-Tech Research both name it as a top 2026 trend. When the consultancies and the research houses agree with the VCs, the thesis has crossed into consensus.
The margin objection, handled honestly
The sceptics have a real case. Euclid Ventures made it most sharply in November 2024: service-as-software companies run at 50-60% gross margins, not the 80-90% that SaaS companies enjoy. AI inference has real compute costs. Human oversight for edge cases adds labour. Forward-deployed engineers add headcount. You can't deliver work without absorbing the cost of doing the work. Euclid's warning is worth quoting directly: "Taking on services just because SaaS is hard and LLMs exist will be a fast track to a low-margin, low-moat, poorly valued enterprise."
This is a fair criticism. The margin gap is real and structural. But the absolute-gross-profit argument cuts the other way. A service-as-software company capturing 100% of a £10M accounting firm's spend at 50% margin generates £5M in gross profit. A SaaS company capturing 10% of that same spend at 90% margin generates £900K. The lower margin on the larger budget wins by 5.5x. The TAM difference is doing the work, not the multiple.
A second angle on margins gets less attention. Delivered work builds relationships in a way that software licences don't. When a company closes your books, handles your claims, or drafts your contracts, switching costs are structural. You're replacing a provider, not a tool. That's a retention moat that SaaS companies have to fight for every renewal cycle.
The operator’s problem: trust
The thesis is clean on a slide. The hard part is deployment. When an AI agent delivers legal work, claims processing, or financial analysis, the client needs to know what the agent did, where a human stepped in, and how quality was verified. Foundation Capital calls this the "quality spec problem": in traditional software, quality is measurable (uptime, latency, bug rates). In service as software, the output is knowledge work where "good" is subjective and context-dependent.
This is where the operator perspective diverges from the investor perspective. Investors debate margins and TAM. Operators have to answer a simpler question: will the client trust the output enough to pay for it?
The answer is operational. It requires mapping the workflow before you touch the model, designing human-in-the-loop checkpoints where judgement matters, and showing the client exactly what the agent did versus what a person handled. These are operations problems, not technology problems. They're the reason most AI deployments stall, not because the models aren't good enough, but because nobody mapped the workflow first.
Crosby is an interesting test case. It is a registered law firm, not a software vendor. Thirty lawyers, eight AI agents, pricing by contract not by hour. The "neofirm" model, where the AI does the drafting and the lawyers do the judgement, is one way to solve the trust problem. The client gets a law firm's letterhead and a law firm's accountability, with an AI engine underneath. The trust attaches to the firm, not the software.
Why the UK, and why real estate
Sequoia's opportunity map covers ten verticals: recruitment, supply chain, insurance brokerage, IT managed services, accounting, healthcare revenue cycle, claims adjusting, tax advisory, legal. Real estate is not on the map. That's a gap, and for a UK company it's a useful one.
UK real estate is process-heavy, document-intensive, and historically slow to adopt technology. Conveyancing alone involves dozens of manual steps, weeks of delay, and reams of compliance paperwork. It's exactly the kind of back-office work that AI agents can absorb. The front office, where client relationships and negotiation and judgement live, stays human. The back office goes to the agents.
At AIP, we started with UK real estate because the back office is large, manual, and underserved. The model is straightforward: agents take the back office, people keep the front. Expansion into financial services, healthcare, legal, and ESG follows from the same operational pattern: map the workflow, deploy agents where the work is repetitive, keep humans where the work is relational.
The US discourse on service as software is written by investors, for investors. There's very little from the operator's seat, meaning what it's actually like to deploy AI agents in a real services business, manage the human-agent interface, and deliver work a client trusts. That's the gap AIP is positioned to fill, and it's the reason we're writing about this.