Outcome-based pricing, where an AI vendor charges only when a defined piece of work gets done, has moved from a startup experiment to mainstream accounting. Sierra, which sells strictly on outcomes, passed $400 million in annual recurring revenue in October 2026 having added over $150 million in Q3 by itself.
The model is spreading, but unevenly, and the way vendors define success deserves scrutiny.
From selling tools to selling finished work
Classic software as a service sold access to tools, and the customer still had to finish the job. Now AI vendors sell the finished job. One closed support ticket runs $0.99. Legal research memos show up within minutes. A literature review that used to need a consultant now costs about a lunch.
Consulting is being repriced as well. McKinsey told reporters in November 2025 that roughly 25% of its fees worldwide were tied to results. Two months on, its CEO put the firm’s AI agents at 25,000, next to 40,000 people. Two regulated law firms sell AI-produced work at fixed prices too. Garfield AI, cleared by the regulator in England and Wales in 2025, charges £2 per demand letter and £50 per court claim. Crosby, a New York firm, has sent 13,000 contracts through AI agents, each reviewed by a lawyer, at $250 to $1,000 apiece.

Who bills what
- Fin, formerly Intercom: charges $0.99 when it resolves a chat and nothing when it doesn’t. ARR topped $100 million in April 2026, the company took the Fin name in May, and Salesforce bought it in September for $3.6 billion. Fin claims it resolves 76% of chats on its own.
- Zendesk: roughly $2.00 per automated resolution on pay-as-you-go, or $1.50 with committed volume, billed only if no person handles the ticket.
- HubSpot: in April 2026 its Breeze agent went from $1.00 for each conversation it handled to $0.50 for each one it resolved, which moves the performance risk onto HubSpot.
- Salesforce: Agentforce crossed $1.5 billion in ARR during July 2026. Its pricing spans $125 and up per user each month, about $0.10 an action, $2 a conversation and, since mid-2026, $2 a resolution.
- Harvey: above $400 million in ARR from over 3,000 organizations, and a $15.5 billion valuation set in September 2026. Most of its billing is still per lawyer seat.
The market is slower than the headlines
A 2025 survey of 240 software companies by Kyle Poyar found that over 12 months the share of companies on seat-based pricing dropped from 21% to 15%, and hybrid pricing grew from 27% to 41%. Just 5% called outcome-based pricing their primary model, though a quarter expect to by 2028.
The strongest sign of permanence came from accountants. Deloitte issued revenue-recognition guidance for agentic AI products on June 4, 2026. One example has a vendor taking $5 million in advance for 250,000 refund cases; it can record each $20 only when that case is actually resolved.
What counts as resolved?
The three published meters can all treat a customer’s silence as a win. Fin bills an assumed resolution once a customer has said nothing for a day after the agent’s final reply. HubSpot calls a conversation resolved when its agent shared an article or acted and no human was requested within three days. Zendesk waits from 2 to 72 hours depending on channel, then asks an AI model to read the transcript, a check it added in May 2026. Fin cancels the charge if the customer comes back to the same conversation.
Silence can mean the customer gave up and left, yet the meter counts it. Critics have labeled this practice “outcomewashing.”
Why support came first
Outcome meters need three conditions:
- A system of record already registers the result, such as a ticket closed, a loan funded or a claim paid.
- The work is dependable enough that the vendor can eat the failures. In spring 2026, METR estimated that the best model it tested had even odds on software tasks that take expert humans 16 to 20 hours, yet needed tasks of just 3 to 4 hours to hit 80% reliability.
- The buyer can define “done” before the work starts.
Support questions take minutes, so they fit inside the reliable zone. Strategy memos and UX redesigns often do not, because people discover what they want while drafting. The danger is receiving precisely what you requested, plus a valid invoice, and only then seeing that the request was wrong.
What to watch
Expect more vendors to copy the meter. Buyers should ask how each one defines a resolution, how it handles silence, and who audits the count. Outcome-based pricing shifts the cost of failed work to the seller, but the consequences of failure, and any uncounted “phantom resolutions,” stay with the buyer.
Source: UX Tigers