AI Agent Pricing Models: How to Read a Vendor Quote

Most AI agent quotes use one of four pricing models. Seat, usage, outcome, or project plus retainer. Most real quotes blend two of them.

The model matters more than the price. It decides who pays when volume spikes, when the agent gets worse, or when adoption stalls.

Here is how I read an AI agent quote, and how I model the 12-month bill before anyone signs.

The four AI agent pricing models

1. Seat pricing. A fixed fee per user, per month. Microsoft priced Microsoft 365 Copilot at $30 per user per month. Easy to budget. Easy to explain to finance.

2. Usage pricing. You pay for what the agent consumes. That can be a message, an action, a conversation, or tokens. Microsoft’s Copilot Studio launched pay-as-you-go at $0.01 per message, or a 25,000-message pack for $200 a month. Salesforce prices Agentforce at $0.10 per action, after earlier charging $2 per conversation.

3. Outcome pricing. You pay when the agent finishes a defined result. Intercom charges $0.99 per outcome for Fin, with a maximum of one charge per conversation. I wrote a longer explanation in the outcome-based pricing glossary entry.

4. Project plus retainer. A fixed fee to build, then a monthly fee to run and maintain. This is how most custom agents are sold. There is no public price list for it, so what follows in this section is my operator view, not survey data.

Most quotes mix them. Bessemer’s AI pricing playbook (February 2026) calls a hybrid of a base subscription plus usage or outcome tiers the effective middle ground. Its example is a $12K annual platform fee that includes 100 resolutions, plus $5K per additional 100.

What each model rewards

Each model pays the vendor for something different. That shapes how the vendor behaves.

Model	The vendor earns more when	What that can reward
Seat	You add users	Selling seats, not results
Usage	The agent does more work	More steps, more calls, longer runs
Outcome	Tasks end in a billable result	Counting easy wins, loose definitions of “done”
Project + retainer	Scope grows or the contract renews	Scope creep, slow handoff

None of these is evil. They are incentives. Read the quote and ask: what behavior does this price push the vendor toward?

Seat pricing fits poorly when an agent does the work. One agent can handle hundreds of cases while no one adds a seat. That is why vendors cap it with limits. Gartner’s Jan Cook told The Register that seat deals often come with “fair use limits”, and named examples like SAP’s AI units and ServiceNow’s credit top-ups. A seat price with a hidden usage ceiling is a usage price in disguise.

Usage pricing rewards the vendor for activity. If the agent takes six steps where three would do, you pay double. Ask whether the meter counts steps, tokens, or finished tasks.

Outcome pricing rewards the vendor for being right. It also pays for being generous with the word “outcome.” Intercom, for example, does not bill when Fin escalates to a human or when the customer abandons a conversation after Fin asks a question. That kind of exclusion list is what you want to see in writing.

Which AI pricing model is cheapest?

It depends on volume. No model wins at every volume.

• At low volume, usage pricing is usually cheapest. You pay for little.
• At high, steady volume, seats or a fixed retainer usually win.
• Outcome pricing is rarely the cheapest per task. You pay a premium because the vendor carries the risk of failure.

The premium is the point. If the agent fails, you do not pay. If you do not trust the vendor to deliver, that is worth money.

I would not pick a model on the lowest sticker price. I would pick the model whose cost curve matches how my volume will behave, then check what happens at the edges.

How to model a 12-month AI agent bill

Do this before you compare quotes. It takes an hour in a spreadsheet.

1. Pick the unit. Tasks per month is the cleanest. Convert each quote into dollars per task.
2. Set three volume cases. Flat, moderate growth, high growth. Use month-one volume you can measure today.
3. Add the success rate. For outcome pricing, only completed tasks bill. Use a conservative completion rate.
4. Add the fixed costs. Platform fees, minimums, onboarding, and the build.
5. Add the costs the quote leaves out. Data platform credits, extra seats, the person who reviews exceptions.
6. Add a spike month. One month at 3x volume. See what the bill does.

Here is an example. The numbers are hypothetical. They are not from any vendor. Say you start at 3,000 tasks a month. Each case below shows the year-one total.

• Seat: 25 users at $60 a month. Flat at $18,000.
• Usage: $0.30 per task.
• Outcome: $0.99 per completed task, with an 80 percent completion rate.
• Project plus retainer: $40,000 to build, then $3,000 a month. $76,000 in year one.

Volume growth per month	Seat	Usage	Outcome	Project + retainer
0 percent (flat)	$18,000	$10,800	$28,512	$76,000
8 percent	$18,000	$17,079	$45,090	$76,000
15 percent	$18,000	$26,102	$68,908	$76,000

What the table shows:

• Usage is cheapest if volume stays flat. It passes the seat price somewhere between 8 and 15 percent monthly growth.
• Outcome costs the most in every case here. The premium buys you not paying for failures, so it only makes sense if you doubt the agent will finish the work.
• The project fee dominates year one. It looks different in year two, when only the retainer is left.
• The seat price looks safest. But only if the fair use limit does not bite at high growth.

Run the same sheet with your numbers. For the full cost picture beyond the quote, see what an AI agent costs to build.

Who carries the risk under each model?

This is the question I would ask first. Price is easy to compare. Risk is not.

Risk	Seat	Usage	Outcome	Project + retainer
Volume spikes	Vendor (until limits hit)	You	Shared	Vendor (until scope changes)
Agent is wrong or fails	You	You	Vendor	Depends on acceptance terms
Low adoption	You	Vendor	Vendor	You
Vendor’s model costs rise	Vendor	Shared	Vendor	Vendor
Bill is hard to forecast	Low	High	Medium	Low

Two of these deserve a note.

Usage pricing puts forecast risk on you. Gartner’s Jan Cook said customers stay cautious on generative AI because pricing is unpredictable. That is part of why Salesforce moved toward seat-based licensing for agents in late 2025, after trying consumption and per-conversation models. Bessemer’s playbook makes a related point: CIOs often ask vendors to push overage invoices into next year’s budget. Surprise bills are a real buyer problem.

Outcome pricing puts delivery risk on the vendor. Bessemer notes that outcome pricing gives the strongest value alignment, but the vendor absorbs the cost variance. Expect them to price that in. Expect them to define “outcome” narrowly in their favor, too.

Hidden fees and minimums

Read the order form, not the pricing page. These are the items I look for.

• Minimums. Intercom’s standalone Fin has a 50-outcome monthly minimum. Many vendors have a floor, and you pay it even in a slow month.
• Expiring credits. Prepaid packs often reset monthly. The Copilot Studio capacity pack above refreshes monthly, and unused messages do not roll over.
• Meter changes. Salesforce changed its agent meter from $2 per conversation to $0.10 per action in 2025. That worked in the buyer’s favor in its own example, where a conversation that cost $2 would cost $0.30 to $0.60 in actions. A meter can also move the other way at renewal.
• Seats on top of usage. Some agent products still require platform seats on top of the per-outcome fee.
• Adjacent platform charges. The agent needs data, search, and a place to run. Ask what bills separately.
• Overage rates. Overage can run well above the contracted rate if it is billed at list price.
• Human review cost. Not on the quote, but real. Someone checks exceptions. Count their time.
• Model cost pass-through. If the vendor passes model costs through, ask what happens when they change.

Negotiation levers

Most of these cost the vendor little. They cost you a lot if you skip them.

1. Cap the spend. A hard monthly limit or an alert at 80 percent. Intercom, for example, offers spend limits that stop the agent and route work to people.
2. Fix the overage rate. Ask for the same unit price above your commitment. Know the difference between a true-up and a true-forward: a true-up bills the overage retroactively, often at list price, while a true-forward raises your future commitment at the contracted rate. Ask for the true-forward.
3. Ask for rollover. Even partial rollover on unused credits helps.
4. Define the outcome in the contract. What counts, what does not, and what happens if a customer reopens a case.
5. Tie fees to acceptance. For project work, pay on tested milestones, not calendar dates.
6. Pilot first at a small commitment. A short, scoped pilot gives you real volume and real completion data. Then negotiate the year.
7. Own your exit. Ask about data export and termination.

Before you sit down, run through the questions to ask an AI agent vendor. If your finance lead needs the investment case, the CFO guide to AI agent investment covers it. And if the quote is for something past a pilot, read how to get from experiment to production first.

Bottom Line

The pricing model decides who carries the risk. The price just decides how much.

Before you sign, build the 12-month sheet with three volume cases and a spike month. Get the outcome definition, the overage rate, and the spend cap in writing.

I’d pick the model that matches how my volume will behave. For a first agent, that is often a small fixed pilot fee, then usage or a retainer once real numbers exist.

FAQ

What are the main AI pricing models? Four: seat (per user), usage (per message, action, or token), outcome (per completed result), and project plus retainer. Many vendors blend them, such as a platform fee with usage tiers. Bessemer calls this hybrid a common middle ground.

Which AI pricing model is cheapest? It depends on volume. Usage is usually cheapest at low or flat volume. Seats or a retainer win when volume is high and steady. Outcome pricing costs more per task because the vendor takes on the failure risk.

How do I model a 12-month AI agent bill? Convert each quote to a cost per task. Run three volume cases (flat, moderate growth, high growth) plus one spike month. Add fixed fees, minimums, and the costs the quote leaves out. Compare the totals, not the unit prices.

Who carries the risk under each model? Under usage pricing, you carry the forecast risk. Under outcome pricing, the vendor carries most delivery risk. Under seat pricing, you carry adoption risk. Under project plus retainer, risk depends on the acceptance terms in the contract.

What hidden fees show up in AI agent contracts? Monthly minimums, expiring prepaid credits, overage billed at list price, seats required on top of usage, and adjacent platform charges. Also count the human time spent reviewing exceptions.

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