Outcome-based pricing for AI agents: what changes

Salesforce started billing its customer service agent only for the cases it closes on its own. When a customer asks for a human or leaves without a resolution, the company does not charge. This is not a cosmetic price change but a shift of risk onto the vendor, and a signal of a wider trend: when software handles a task end to end, paying by the number of seats no longer matches the value. For boards and managers this means a new way to buy, budget, and control AI, where the billing unit becomes the result, not access to the tool.

What happened

On 25 June 2026 Salesforce announced Agentforce Help Agent, an autonomous customer service agent with guided setup that, according to the company, deploys in minutes and works across every channel. What stands out is its billing mechanic. In the pay-per-resolution model, a charge applies only when the agent resolves a case autonomously, from start to finish. If a case goes to a human or the customer leaves without a resolution, Salesforce does not charge. The company priced its earlier conversational model at 2 USD per conversation, and the new variant moves the point of billing from the conversation to the result.

A few weeks later, on 14 July 2026 on CNBC, Sierra co-founder Clay Bavor argued that AI agents could change how software is billed. His point is simple: as agents move out of demos into real service, sales, and support workflows and close whole tasks, the natural billing unit becomes the outcome achieved, not the number of people with access.

These are not isolated moves. In its 2026 TMT Predictions, Deloitte expects a shift away from per-seat licensing and classic subscriptions toward hybrid models that blend usage- and outcome-based billing, and estimates that up to 75% of companies may invest in agentic AI. A Gartner forecast cited in the same material says that by 2030 at least 40% of enterprise SaaS spend will move toward usage-, agent-, or outcome-based pricing. Behind it sits a simple economics that Gartner described back in August 2025: the share of enterprise applications with task-specific agents is set to rise to 40% in 2026 from less than 5% a year earlier.

Why it matters for business

The billing model stops being a minor clause and becomes part of the AI buying strategy. When you pay for seats, the cost is predictable but detached from value. When you pay for outcomes, the cost rises with the benefit, but only if you define that benefit well. For a company buying a customer service agent, the difference is tangible: in an outcome model you do not pay for a thousand conversations in which the agent settled nothing.

The distribution of risk also changes. In an outcome model, the vendor stakes its own revenue on whether the agent actually completes the task. For the buyer that signals the vendor believes in its product, but it is also a commitment to define precisely what counts as success. Without that definition disputes are easy, because formally closing a case and genuinely solving the customer's problem are not always the same thing.

There is another side. McKinsey points out that outcome-based pricing sounds attractive but is hard to execute, and customers do not always choose it. In an example cited by McKinsey, at an AI-native company that offered an outcome-based option, 90% of customers still reached for the simpler activity-based model. The reason is practical: outcomes are hard to measure and contract, while usage-based billing is understandable and easier to control. For a company this means the new model is worth treating as an option to assess deliberately, not as a default better choice.

Business use cases

Customer service billed per resolved case

Business problem: the cost of frontline support grows with ticket volume, and a large share of cases are repetitive questions about order status, returns, or basic setup.

AI solution: a support agent on a pay-per-resolution model, such as Agentforce Help Agent, that closes repetitive tickets on its own across chat and email channels and hands complex cases to a human without a charge.

Data and processes to put in place: a knowledge base with current procedures, a history of resolved tickets as reference material, clear escalation rules, and integration with the ticketing system.

Potential effect: cost moves from the number of seats to the number of genuinely resolved cases, and the company pays only for closed tickets, not for failed attempts.

Risk and limitation: the agent can close a case formally without solving the real problem, so outcome billing has to be tied to a measure of satisfaction and the rate of repeat contacts on the same case.

Lead qualification in sales

Business problem: the sales team wastes time on contacts that never mature into a sales conversation, and tool cost grows regardless of how many customers actually enter the funnel.

AI solution: an agent that runs the initial conversation, gathers qualifying information, and passes to a rep only leads that meet agreed criteria, billed per booked meeting or qualified contact.

Data and processes to put in place: a definition of a qualified lead, a history of won and lost opportunities as material to train the rules, and integration with the CRM and the reps' calendars.

Potential effect: spend tracks the number of real sales opportunities rather than the number of conversations held, which makes the return easier to calculate.

Risk and limitation: if the unit of outcome is the meeting alone, the agent may optimize the number of meetings at the expense of their quality, so it is better to bill only for a meeting confirmed by the rep.

Finance operations and receivables recovery

Business problem: the finance team spends a lot of time on payment reminders and reconciling transactions, and these processes scale linearly with the number of invoices.

AI solution: an agent that runs polite, repeated reminder communication about overdue payments and reconciles simple discrepancies, billed per recovered payment or reconciled transaction.

Data and processes to put in place: receivables data, payment terms, a history of customer contact, and rules for when a case must go to a human or the legal team.

Potential effect: tool cost grows only with real cash recovery, which directly ties spend to the company's liquidity.

Risk and limitation: communication about money is sensitive to reputation, so tone rules and escalation boundaries must be defined, and irreversible actions left to a human.

IT support and the internal help desk

Business problem: the internal help desk handles repetitive tickets about password resets, application access, or device setup, which burdens the IT team at the cost of development projects.

AI solution: an internal agent that resolves typical employee tickets end to end, billed per closed ticket, and routes unusual cases to a specialist.

Data and processes to put in place: an IT knowledge base, a service catalog, system permissions scoped to allowed actions, and clear limits on what the agent may do without human confirmation.

Potential effect: the IT team recovers time for projects, and the agent's cost tracks the number of genuinely closed tickets, not the number of employees in the company.

Risk and limitation: an agent with access to internal systems increases the attack surface, so its permissions should be limited to reversible actions and every completed task should be logged.

What companies can do now

Measure the baseline before you compare offers. Without knowing the current cost of handling one case, the share of cases closed without escalation, and the time the process takes, you cannot judge whether outcome billing is cheaper. That measurement matters more than the rate on the price list.

Define the outcome precisely in the contract. Set out exactly what counts as a resolved case, a qualified lead, or a recovered payment, and confirm in writing that escalations to a human and unresolved cases are not billed. A vague definition of the outcome is the most common source of invoice disputes.

Set a spend cap and monitor usage. Outcome billing means a variable cost that grows with volume. Set a monthly ceiling and an alert as you approach the limit, so a strong sales season does not turn into an uncontrolled bill.

Assign a process owner and quality control. Name a person responsible for overseeing the agent, reviewing a sample of closed cases, and measuring real customer satisfaction. Without a process owner and quality control, outcome billing turns into paying for formal rather than real resolutions.

Take care of data and security. The agent runs on customer data and company systems, so the scope of its access, data-handoff rules, and the limits on irreversible actions must be set before launch, not after the first incident.

Start with one process and count the result. Pick one well-measured process, run a bounded pilot, and compare cost and effectiveness against the baseline. The pilot result, not the promise in a deck, should decide whether to expand the rollout.

Risks

A vague definition of the outcome. If the contract has no precise clause on what counts as success, vendor and customer will count the invoice differently. Formally closing a case is not the same as solving the customer's problem.

Unpredictable cost. Outcome billing is a variable cost. During a volume spike, for example a season or a promotion, the bill can grow faster than the budget if there is no cap and no monitoring.

Tollgating. Deloitte describes the risk where an agentic SaaS vendor places paid gates on successive agent actions or outcome stages, which makes cost hard to predict and control on the buyer side. It is worth checking how many gates are billed within a single task.

Optimizing for the metric. An agent billed by the number of meetings or closed tickets may maximize that number at the expense of quality. The outcome metric has to be tied to a measure of satisfaction and the rate of repeat contacts.

Shadow AI. When individual teams buy outcome-billed agents without the knowledge of IT and finance, the company loses control over cost, data, and security. Buying AI tools should go through a defined process.

No quality control over results. Without reviewing a sample of cases, no one knows whether closed tickets actually helped the customer. Outcome billing does not remove responsibility for quality, it moves it onto the definition and measurement of that outcome.

Key takeaways

Salesforce Agentforce Help Agent, billed only for cases resolved without a human, is a concrete example of a wider move from per-seat licensing to outcome billing. Clay Bavor's remarks from Sierra on CNBC and the forecasts from Deloitte and Gartner show the direction is durable, not a passing moment.

For a company, what matters most is not the rate itself but the definition of the outcome, the transparency of billing, and quality control. These decide whether the new model actually ties cost to value.

Outcome billing is not better by default. McKinsey shows that customers often choose the simpler usage-based model, because outcomes are hard to measure and contract. The decision is best grounded in your own baseline and a pilot.

Before you sign, measure the current process, define the outcome in writing, set a spend cap, and assign a process owner. Without those four steps, the new price list shifts risk onto the company instead of taking it away.