Outcome-Based Pricing for AI Agents in Back-Office Work

Outcome-based pricing for AI agents means you pay when the agent finishes a defined piece of work, not for seats or raw usage.

It sounds fair. It is fair only if you write down what “finished” means. In support that is easy. In back-office work it is not.

This guide covers how to define an outcome, who carries which risk, where the model breaks, and what to ask a vendor.

What outcome-based pricing for AI agents is

Under outcome pricing, the vendor bills per result. A resolved ticket. A posted invoice. A completed reconciliation.

Bain put the shift this way in its 2025 technology report: stop charging for access and start charging for work done. When an agent does a task that a person used to do, a per-login fee stops tracking value.

There are four common ways to bill an agent:

Model	You pay for	Who carries volume risk	Who carries quality risk
Per seat	Each licensed user	Buyer	Buyer
Per usage	Calls, tokens, actions	Buyer	Buyer
Per outcome	A result that meets a test	Shared	Mostly vendor
Fixed fee	A scope over a term	Vendor	Shared

Most real contracts blend two of these. More on that below.

Gartner predicts that by 2028 over half of enterprises will stop paying for assistive AI like copilots and favor platforms that commit to workflow results. That is a forecast, not a fact. But it explains why buyers keep seeing outcome terms in proposals.

Published examples and their limits

Most public examples come from customer service. They are useful, but they do not carry over cleanly.

Zendesk bills per automated resolution. Its outcome pricing post, published in April 2026, lists $1.50 per automated resolution. It defines a resolution as an issue the AI agent solved with no human, confirmed when the ticket stays closed through a 72-hour quiet period. Its help center adds that escalated conversations and ones that fail an automated check are not billed as resolutions.

Sierra published its view in December 2024. Its post on outcome-based pricing names outcomes such as a resolved conversation, a saved cancellation or an upsell. It says escalations are generally not charged and that criteria are agreed up front. It also admits outcome pricing may not suit every job. Routing or greeter-style work may fit a blended consumption model better.

Salesforce shows how fast units move. Agentforce was sold at $2 per conversation. In May 2025 the company added Flex Credits, where each action costs $0.10. Those are not strictly outcome prices. They are a reminder that even large vendors change the billing unit.

Three limits show up in all of these:

1. The unit is tied to a conversation. A support chat has a start, an end and a customer who can confirm. A back-office task often has none of that.
2. The success test is a proxy. “Customer did not come back for 72 hours” is a stand-in for “problem solved.” It works at volume. It can miss slow failures.
3. The vendor often measures. If the vendor runs the counter, you need audit rights. I’d never accept a number I cannot reproduce.

Defining a unit of outcome in back-office work

Back-office work does not arrive as a chat. It arrives as invoices, claims, forms, exceptions and reconciliations. Each has a messier path from start to “done.”

A usable outcome definition has seven parts. Write them in the contract.

1. The unit. What is counted? One invoice, one claim, one account, one exception.
2. The success test. The exact conditions that make it count. Be binary where you can.
3. The evidence. The system record that proves it, such as an ERP document ID or a workflow log.
4. The window. How long it must stay done. A transaction reversed in week two should not count.
5. The exclusions. Cases that never count, such as bad inputs, manual overrides or duplicates of the same item.
6. Who measures. The system of record, not the vendor’s dashboard alone.
7. The dispute path. Who decides when you disagree, and how fast.

Deloitte’s accounting team published a useful illustration in June 2026. In its Technology Spotlight on outcome-based pricing, a hypothetical finance-operations platform charges a $600,000 annual access fee, plus $1.25 per invoice processed straight through with no human touch, plus $1 per duplicate payment prevented. It is an accounting example, not a market price. But the success test is the part to copy. An invoice counts only if it posts correctly to the ERP, approvals follow the delegated authority matrix, controls pass, and it is not reversed for cause within an agreed window.

That is a good model. The unit is countable. The evidence sits in your own systems. There is a reversal window. And note the structure: a fixed fee plus per-outcome fees, not a pure per-outcome deal.

For more on where invoices break in practice, see our post on AI agents for accounts payable exceptions.

Who carries which risk

Outcome pricing moves some risk to the vendor. It does not remove yours.

Risk	Fixed fee	Per outcome	What to do
Volume swings	Vendor	Buyer pays more at high volume	Cap or tier the rate
Agent fails to finish	Buyer	Vendor (no bill)	Define “finished” tightly
Agent finishes wrongly	Buyer	Shared, unless the test catches it	Add quality checks and a reversal window
Bad inputs from you	Buyer	Vendor, so expect exclusions	Agree what counts as bad input
Disputes about counts	Rare	Frequent	Audit rights and a named referee
Budget predictability	High	Lower	Spend caps and a base fee

BCG’s 2025 pricing report puts numbers on the friction. It reports that 47% of buyers struggle to define clear, measurable outcomes and 36% worry about cost predictability. It also notes that financial-outcome pricing, where the vendor is paid on a business result like revenue, is the riskiest form because vendors do not control the full result. BCG’s page does not state the sample behind these figures, so treat them as indicative.

My view: the fewer outside factors that affect the outcome, the better the model works. Straight-through invoice posting is mostly in the agent’s hands. “Reduce days sales outstanding” is not.

Where it breaks

Here are the failure modes I would plan for.

Low volume. If you only have a few hundred events a month, the vendor cannot spread risk, and one dispute is a big share of the bill. Expect a minimum fee.

Ambiguous quality. If “good” is a judgment call, you will argue about every count. Use a binary test or a sampled review.

Cherry-picking. The vendor earns only on easy cases, so the agent may pass the hard ones to your staff. You pay for the easy work and keep the hard work. Watch the share of items the agent attempts, not just the ones it completes.

Gaming the test. If success means “not reopened in 72 hours,” a vendor can close tickets in ways that discourage reopening. Pair the test with a quality sample.

Slow failures. Some errors show up in month two. Use a reversal window and a clawback.

Vendor-controlled counting. If the vendor owns the meter, you cannot verify the bill. Insist on logs you can reconcile. Our post on what to log for an AI agent audit trail lists the fields.

Price decay. Unit prices in a young market move. Salesforce added a new billing unit to Agentforce in May 2025. Lock a rate review into any multi-year deal.

The trap that surprises buyers

Per-outcome pricing protects your spend. It does not protect your results.

Say a hypothetical team processes 20,000 invoices a month at $1.50 per straight-through invoice. At a 70 percent straight-through rate, the bill is $21,000. If the agent slips to 50 percent, the bill falls to $15,000. That looks like a $6,000 saving.

But 4,000 more invoices now go to people. At 8 minutes each and $40 an hour loaded cost, that is about 533 hours, or roughly $21,300 in extra labor. The vendor bill dropped by $6,000 and your total cost went up by about $15,000.

So an outcome contract needs an operating target too: a minimum straight-through rate, with a remedy if the agent misses it. Without that, a falling bill can hide a failing agent. Our guide to the AI agent SLA covers the service terms to pair with the price.

Hybrid models

Most deals I expect to see are hybrids. BCG expects vendors to widely adopt hybrid approaches that blend subscriptions with agentic pricing in the near term. That fits the Deloitte example above.

Common blends:

• Base plus outcome. A fixed platform fee covers hosting, monitoring and support. A per-outcome fee covers the work done.
• Tiered outcome. The rate falls as volume rises.
• Capped outcome. A monthly ceiling keeps budgets safe.
• Shared savings. The fee is a share of measured savings. This is the hardest to audit.

Outcome pricing versus fixed price

Neither is cheaper by default. It depends on volume and success rate.

Say a hypothetical vendor offers two options: a fixed $15,000 a month, or $1.50 per straight-through invoice.

• Break-even is 10,000 straight-through invoices a month.
• At 8,000, outcome pricing costs $12,000, which beats the fixed fee.
• At 14,000, it costs $21,000, which is $6,000 more.

Situation	Better fit
Volume is low or uncertain	Outcome pricing
Volume is high and steady	Fixed or tiered
The success test is crisp	Outcome pricing
The success test is fuzzy	Fixed fee plus a quality target
You are in a pilot	Fixed fee with exit rights
You want the vendor to share the downside	Outcome pricing with a minimum success rate

For a pilot, I’d choose a fixed fee, a short term and a clear exit. Move to outcome terms once you have real counts. Our guide to AI agent pricing models walks through reading each type of quote, and the CFO’s guide to AI agent investment covers how to put the numbers in front of finance. You will also want realistic run costs. See what an AI agent costs to run per month.

Questions for the vendor

Take this list to the next pricing call.

• What exactly is one billable outcome? Show me the success test in writing.
• What system record proves each outcome? Can I export it?
• Who runs the counter, and can I audit it?
• What is the reversal window, and is the fee refunded on reversal?
• What happens to cases the agent attempts but cannot finish?
• Are there exclusions I would not expect?
• Is there a platform or minimum fee? What does it cover?
• How does the rate change with volume?
• What is the monthly cap, and what happens at the cap?
• What is the minimum success rate, and what is the remedy if you miss it?
• Can the unit or price change mid-term?
• How do we settle disputes, and how fast?

Bottom Line

Outcome pricing works when the outcome is countable, checkable in your own systems, and mostly in the agent’s control.

Write the success test, the evidence and the reversal window into the contract.

Pair the price with an operating target, because a lower bill can hide a failing agent.

Start pilots on a fixed fee. Move to outcome terms when you have real counts.

FAQ

What is outcome-based AI pricing? It is a billing model where you pay when an AI agent completes a defined result, such as a resolved ticket or a posted invoice, rather than paying per seat or per call. The contract must define what counts as a completed result and how it is measured.

Who carries the risk? The vendor carries the risk that the agent fails to finish, since unfinished work is not billed. You still carry volume risk, quality risk if the test is weak, and the cost of any work the agent passes to your staff. A written success test and a minimum success rate shift more of that risk to the vendor.

How do I define an outcome outside support? Pick a countable unit, such as one invoice. Write the success test, the system record that proves it, a reversal window, exclusions, who measures, and a dispute path. The test should be binary where possible, and checked in your own system of record.

What can go wrong? Common problems include low volume, disputes over counts, the vendor handling only easy cases, gaming the success test, slow failures that appear after billing, and vendor-controlled meters. A falling bill can also hide a failing agent that pushes work back to your team.

How does it compare to fixed price? Neither is cheaper by default. Outcome pricing tends to win at low or uncertain volume and with a crisp success test. Fixed or tiered pricing tends to win at high, steady volume. Many deals combine a base fee with per-outcome fees.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect