The last thirty years of B2B software had a clear playbook. Build a product, sell seats, grow monthly recurring revenue. Salesforce did it. Workday did it. ServiceNow did it. The model worked because software has near-zero marginal cost and scales infinitely.

AI is breaking that model — not by building better software, but by making the software category irrelevant.

The $1:$6 ratio

Sequoia Capital recently put a number on something the AI world has been circling around without naming directly.

For every dollar companies spend on software, they spend six on services. Consultants, analysts, contractors, agencies, outsourced operations. That’s not a niche. That’s the entire budget.

Software has always competed for the $1 bucket. Every SaaS company ever built is premised on: “pay us instead of building it yourself.” But the six dollars? Those have been off the table. You couldn’t automate a consulting engagement. You couldn’t replace a claims adjuster with a subscription fee.

Until now.

Copilots vs. autopilots

Sequoia draws a clean line between two strategic bets in AI:

Copilots sell tools to professionals. Harvey sells to law firms. Rogo sells to finance teams. The buyer is an expert; the AI makes them faster. Revenue comes from software budgets.

Autopilots sell outcomes to buyers who don’t want to think about how the work gets done. WithCoverage sells insurance review to CFOs who need a certificate, not software. Anterior processes medical claims — it replaces a workflow, not a seat. Revenue comes from services budgets.

The copilot market is large. The autopilot market is six times larger.

As Sequoia frames it: “The next $1T company will be a software company masquerading as a services firm.”

Why now: the intelligence/judgement divide

Here’s the mechanism that makes this possible in 2026 and wasn’t in 2020.

All knowledge work sits on a spectrum. At one end: intelligence work — following rules, processing information, applying known patterns. At the other end: judgement work — navigating ambiguity, weighing tradeoffs, reading rooms.

AI has crossed a threshold in intelligence work. Software engineering hit first and hardest because it is mostly intelligence work. Writing code, reviewing pull requests, translating requirements into implementations: these are rule-following tasks that AI handles autonomously at scale today.

But the frontier isn’t fixed. Today’s judgement becomes tomorrow’s intelligence. As AI systems accumulate proprietary domain data and compound on experience, work that currently requires human judgement gradually crosses into territory AI can handle on its own. The line moves.

This is why outsourced work is the natural wedge. Start where work already flows outside the company — insurance brokerage, claims adjusting, revenue cycle management, IT managed services. The buyer already purchases outcomes. The budget line already exists. AI just has to be good enough to do the work.

Then expand inward toward higher-judgement territory as the models improve.

The VC consensus

What makes this thesis notable isn’t that one firm believes it. It’s that the entire industry is converging on the same conclusion from different angles.

Andreessen Horowitz has rebuilt their investment thesis around the premise that AI is eating software itself. Their 2026 Request for Startups includes a category called “Software for Agents” — with the explicit framing that the next trillion users of software aren’t people, they’re AI agents acting on behalf of people. The category they’re betting on isn’t SaaS. It’s whatever replaces it.

Y Combinator’s recent batches have trended toward roughly half AI agent companies. YC’s framing is direct: trillions of dollars are spent on knowledge work. That’s the market. The companies that capture it won’t look like traditional software businesses — they’ll look like staffing firms, but with no staff and software-level margins.

Bessemer Venture Partners published an AI services roadmap projecting that India’s IT services sector alone reaches $400B by 2030 under AI disruption. Their framework identifies three categories: pure AI automation that handles work end-to-end, AI-enabled services that blend automation with human oversight, and infrastructure services that support both. The common thread across all three: outcome-based pricing replaces seat-based pricing.

Three different firms. Three different entry points. Same destination.

For the rest of us: what this actually means

If you’re not tracking VC thesis documents, here’s the plain version.

Every company buys two things: tools and work. Tools are software licenses, subscriptions, platforms. Work is consultants, agencies, contractors, employees handling operations. The tools category has been carved up by tech giants for thirty years. The work category is mostly untouched.

AI is going after the work category. Not by building better tools for the humans doing the work — by doing the work itself, and billing for it the way a contractor does: per outcome, per result, per deliverable.

The companies that figure out how to sell work at software margins will be extraordinarily valuable. Not because they’re competing for a slice of the software budget. Because they’re competing for a budget that’s six times larger.

The early forms already exist

This isn’t speculative. The pattern is live.

Medical coding companies are replacing human coders with AI that processes claims at scale and bills per claim processed. Revenue cycle management firms that once employed thousands are rebuilding as AI-native operations. Insurance brokerage startups are approaching CFOs with “we’ll handle your certificate of insurance review” — not “buy our software to handle it yourself.”

Recruitment platforms charge per successful hire rather than per seat. IT managed services companies bill per incident resolved rather than per user covered.

Every one of these looks more like an agency than a software company. That’s the point. They’ve crossed the budget line from tools to work — and they’ve done it at margins that no traditional services firm could match, because there’s no human labor cost underneath the delivery.

What changes when the model flips

The metrics change. You don’t measure daily active users or seat expansion. You measure work completed, outcomes delivered, error rates.

The sales motion changes. You’re not selling to a VP of Engineering with a software budget. You’re selling to a COO with an operations budget who wants the work off their plate.

The moat changes. It’s not switching costs or network effects. It’s proprietary domain data — the closed loop of AI doing work, receiving feedback, improving, doing more work. Every autopilot that processes enough claims, enough contracts, enough invoices builds an asset that a copilot never accumulates. The data flywheel is the defensibility.

And the ceiling changes. The software TAM is large. The services TAM is the entire economy.

References