Agents change the unit of economic exchange: software stops being sold as access (seats, API calls) and starts being sold as completed outcomes. Watch tasks flow from buyer to agents — gold coins are money. Drag to orbit.
Classic software pricing sells access: a seat license, a subscription, a metered API call. The buyer pays whether or not anything useful happens. Agents invert this — when software can complete a task end-to-end, you can sell "this task, completed to this spec" as the unit. That is outcome pricing, and it is closer to hiring a contractor than licensing a tool.
Under seats and usage, execution risk sits with the buyer: a failed task still consumed a seat-month or API tokens. Under outcomes, risk moves to the vendor — they eat the compute cost of retries and failures, and only bill when the spec is met. Flip the model toggle at a 40% success rate and watch vendor margin collapse: that is why outcome vendors obsess over reliability, and why only high-success-rate tasks get outcome-priced first.
vendor margin = successes × price − attempts × compute cost, where attempts ≈ volume × min(1/p, 3)
"Completed to this spec" hides two hard problems. Spec ambiguity: most business tasks aren't precisely specified ("write a good landing page"). Verification cost: someone must check the work — if checking costs as much as doing, the market collapses. Outcome markets therefore start where success is machine-checkable: a test suite passes, a booking is confirmed, a document parses, a refund is issued.
Once tasks are the unit, you need infrastructure that subscriptions never needed: pricing (what is this task worth? auctions and bids beat fixed rate cards when task difficulty varies), matching (route each task to the agent with the best success-rate × cost profile — reputation systems become load-bearing), and arbitration (a dispute path when buyer and vendor disagree on whether the spec was met — escrow, judges, appeal). This is why outcome-based agent platforms look less like SaaS and more like labor markets or ad exchanges.
Outcome pricing isn't hypothetical — it already exists at the edges of software. Intercom's Fin charges ~$0.99 per AI-resolved support ticket (resolution = customer doesn't reopen). Performance ad markets price CPA (cost-per-acquisition), not impressions — with fraud detection playing the arbitration role. Freelance platforms (Upwork escrow, 99designs contests) solved matching + dispute resolution for human task markets years ago; agent markets inherit those designs. Bug bounties pay only for verified vulnerabilities — a pure outcome market with a security team as judge. The common thread: each invented a verification mechanism BEFORE the pricing model could work.
Predictability: seats best for vendor, outcomes best for buyer. Incentive alignment: seats = none (shelfware is profitable), usage = perverse (retries earn money), outcomes = aligned (vendor paid for success only). Vendor capital needs: outcomes require balance-sheet cushion to absorb failed-attempt compute — expect outcome vendors to look more like insurers, pricing risk pools rather than features. Buyer trust required: lowest for outcomes — you can try one task for $8 instead of committing to an annual contract. That collapse in adoption friction is the strongest argument that agent-native software drifts toward outcome pricing wherever verification is cheap.
Seats: $99/seat, 1 seat per 250 tasks/mo (min 2). Usage: 12 API calls per task attempt at $0.08/call, failures billed. Outcome: vendor retries up to 3× (expected attempts ≈ min(1/p, 3)), compute cost $0.42/attempt, only delivered outcomes billed. Toy numbers — but the risk geometry they reveal is real.