AI economics
Outcome pricing needs a budget for the human queue
- An AI agent priced per outcome bills only for the conversations it resolves. Every other conversation still reaches a person, and that queue needs its own line in the budget.
- Fin's guide for CFOs sizes the vendor's bill from four numbers. The same numbers size the human queue, which in month one holds a third to more than half of the conversations the agent joins.
- Finish the arithmetic on a worked example from Fin, and the conversations left for people cost more than five times the vendor's bill.
- A hard cap on resolutions sends work to people on the busiest days. Fund the overflow and a review sample of billed outcomes before you set the cap.
We think a budget for an outcome-priced AI agent is half written if it covers only the vendor's bill. Conversations the agent does not resolve still go to people, and if that queue has no budget, a spending cap pushes the hardest work onto an unfunded team. In our view, the first task that team drops is checking what the agent billed as resolved.
Outcome pricing is reaching bigger contracts
Salesforce completed its purchase of Fin on 10 September. On 22 September, Fin published a guide to talking to your CFO about outcome-based pricing, written by Eoin Gilligan Martyn, a senior data scientist there. In this model, the customer pays only when the agent delivers a defined result.
Salesforce is offering the model in larger deals. Its fiscal second-quarter results on 26 August put Agentforce annual recurring revenue above $1.5 billion, a figure that now also counts Slackbot and Headless 360. On the earnings call, Marc Benioff said Salesforce is "still trapped in some ways in old per user pricing models." Miguel Milano, its chief operating officer, described offering a $40 million customer a choice between an outcome-based deal and an unlimited agreement.
The vendor's formula sizes one line
Fin's guide builds the forecast from four numbers multiplied together: annual conversation volume, the involvement rate, the resolution rate and the price per resolution. The involvement rate is the share of conversations the agent takes part in, and the resolution rate is the share of those it resolves. The guide asks for conservative, baseline and optimistic cases.
Those four numbers size the vendor's line. They also size a second line that the vendor's invoice does not show. Multiply involvement by resolution to get the share of all conversations the agent resolves. Everything else goes to people: conversations the agent never joins, conversations it hands over, and every conversation after a cap is reached.
The three cases are built for the vendor's line. In the guide, the conservative case has volume falling and the optimistic case has resolution rising. For the human queue, the costly case is volume rising while resolution stays flat, and none of the three describes it.
- Conversation arrives. Leads to Agent takes part, Agent not involved.
- Agent takes part: involvement rate. Leads to Resolved by the agent (resolution rate), Escalated to a person, Usage limit reached.
- Agent not involved: scope, channel or policy. Leads to Human queue.
- Resolved by the agent: billed per outcome. Leads to Review sample (sample).
- Escalated to a person: user asked, frustrated, out of scope. Leads to Human queue.
- Usage limit reached: the agent stops helping. Leads to Human queue.
- Review sample: a person checks billed outcomes.
- Human queue: outside the vendor's forecast.
Month one leaves much of the work with people
Fin's guide publishes median first-month resolution rates for 15 industries, drawn from its customer data. Turned around, they show the share of conversations an agent typically left for people in its first month.
One in nine companies using Fin have the agent take part in every conversation. For the rest, the human share of all conversations is larger. Resolution also improves gradually, with a median rise of 16 percentage points over the first 18 months, by Fin's count. Plan for a human queue that starts large and shrinks over quarters.
People cost more per conversation than the agent
Fin's guide to resolution rates, from June, prices a resolution at $0.99 against $6 to $12 for a conversation a person handles. Those are Fin's figures. The guide works an example: 50,000 conversations a month at $8 per human interaction, with 60 percent moved to the agent at $0.99, saves about $2.5 million a year.
Finish the arithmetic on the same example. The vendor's bill comes to $356,400 a year. The 40 percent left with people is 240,000 conversations a year, which cost $1.92 million at the same $8, more than five times the vendor's bill. The saving is real, and the larger line still needs a budget.
That 40 percent is not an average slice either. The CFO guide notes that informational questions are the easiest for an agent out of the box, while personal and action-taking ones need more setup. Fin's escalation reasons run from frustrated users to out-of-scope requests. The conversations left for people skew to the harder ones, so we would treat the old average cost per conversation as a floor for the queue.
Human review is already one of the costs in the cost per successful outcome formula. A per-outcome invoice shows only the vendor's part of that cost.
A cap moves the work to people
Fin's guide is candid about caps. It recommends asking finance to approve the optimistic case, says hard resolution limits keep spending inside the approved budget, and warns that if the cap is hit, the agent cannot help and the load falls suddenly on human support reps. Fin's escalation reporting lists "usage limit reached" among the reasons a conversation goes to a person.
A cap is most likely to be reached when volume peaks. The guide's own example of a spike is Kalshi, a regulated prediction market that saw a record number of support conversations on Super Bowl Sunday. On a day like that the human queue is already long, and a reached cap adds every later conversation to it.
This is how a cap starves review. When a queue fills up, we expect people to answer waiting customers first and leave the sample of billed outcomes unchecked. That sample is the buyer's only independent check on what the invoice counts as resolved, and it needs a written definition of resolved, which is why we argue for keeping your own test set.
Customers care about the human side too. In Fin's AI Sentiment Report, a survey of more than 1,000 end users, one of their top two requests before trusting an agent was "easy escalation to a human when needed."
Budget three lines and a cap rule
What to put in the budget
Line 1Size the vendor line with the four numbers
Spread the forecast by month, as Fin's guide suggests, so seasonal peaks show in the budget.
Line 2Size the human queue from the same numbers
Count the conversations the agent will not resolve and price them at your own cost per human-handled conversation. Use the lowest resolution rate you expect and last year's busiest week, with a margin for harder cases.
Line 3Fund a weekly review sample of billed outcomes
A person checks a fixed number of resolved conversations against your written definition of resolved. Keep the sample when queues are long.
CapWrite the cap rule before you set the cap
Name the share of the cap that triggers a decision, who makes it, and whether the answer is a higher cap or more people. Approve the cap and the staffing plan together.
ContractAsk whether a handoff can be billed
Fin's March billing change says outcomes reflect transfers as well as task completion. Get the answer for your contract in writing.
MonthlyTrack escalations by reason
Any conversation escalated because a usage limit was reached means the cap rule fired. Review the budget that month.
Running an agent and owning its number, the argument in Anyone can build a digital worker, includes owning the human share of the work. Our human-in-the-loop approach designs that queue on purpose: what routes to people and why, a review sample with its own rubric, and automation rate, error rate and cost on one dashboard.
Before you sign, take last year's busiest week, run it through Fin's four numbers, and price what is left at your own cost per human-handled conversation. Put that figure in the budget next to the cap.
Sources
- Fin, Talking to your CFO about outcome-based pricing for AI Agents, 22 September 2026
- Fin, AI Resolution Rate: What It Is, 2026 Benchmarks, and How to Improve It, 17 June 2026; Escalation Reporting, 3 August 2026; Fin's Pricing Metric is Evolving, 12 March 2026; Salesforce completes acquisition of Fin, 10 September 2026
- The Intercom Blog, Announcing the 2026 AI Sentiment Report, 9 September 2026
- Salesforce, fiscal 2027 second-quarter results and earnings call transcript, 26 August 2026