AI Agent Pricing Models: Where the Overshoot Comes From
Gartner predicted on 26 October 2026 that agents on uncapped token pricing will overrun cost expectations by more than 50% through 2027. Five days earlier, Forrester predicted rising AI costs will force B2C leaders to cut a third of AI initiatives.
TL;DR: Two analyst houses now name uncapped consumption pricing as the budget risk; at fixed volume, our worked example shows work-metered models overshoot 60% while capped, per-outcome and per-conversation models stay bounded.
By Rob Jiang, Chief AI Engineer · Published 31 August 2026 · Last updated 26 October 2026 · Rates checked 26 Oct 2026
Methodology & sources
- Analyst figures: Gartner, 26 Oct 2026 (via ITPro coverage); Forrester, 21 Oct 2026; Gartner, 26 Jan 2026.
- Every vendor rate read from that vendor's own pricing page or documentation on 26 October 2026. "Not published" means the page carries no dollar figure.
- The overshoot scenario uses stated assumptions, not benchmarks. Aissist figures: 83% average resolution, 4.8/5.0 CSAT.
- Disclosure: this is Aissist's blog. Aissist sells AI agents on a capped per-interaction rate.
What did Gartner and Forrester actually predict?
Gartner predicts that through 2027, enterprises running AI agents in IT operations on uncapped, token-linked pricing will exceed initial cost expectations by more than 50%, compared with those on predictable subscription or bundled pricing. The release sits alongside a survey of 272 heads of infrastructure and operations, run April–May 2026. Read it precisely: the survey measured priorities, and the 50% figure is Gartner's forecast. It is also about IT operations, not customer service. The mechanism is what transfers.
Forrester's 2027 predictions reach the same place from the B2C side. Forrester says usage-based martech pricing turns every agent interaction into a billable event, and predicts leaders will scale back a third of AI initiatives as a result. The same release predicts CX staffing will rise 20% after what it calls overzealous cuts.
"The challenge for leaders in 2027 will be gaining greater control without limiting growth." — Sharyn Leaver, Chief Research Officer, Forrester
Two houses, four days apart, independently named the meter as the thing that kills AI programmes. Neither named a rate. That is the useful part: the problem is the shape of the bill, not its size.
What are the five AI agent pricing models in 2026?
The five AI agent pricing models are per seat, per resolution, per interaction, per outcome and credit-based, and they differ in the event that triggers a charge. That trigger decides what makes your bill move, which matters more than the headline rate.
| Model | Charged when | Vendors and published rate (checked 26 Oct 2026) | What moves the bill |
|---|---|---|---|
| Per seat | A human holds a licence | Agentforce add-ons $125/user/month, employee usage unmetered; Freshdesk $19–$89/agent/month | Headcount |
| Per resolution | AI closes a case without a human | HubSpot Customer Agent $0.50 (50 credits), Pro or Enterprise; Zendesk not published; Gorgias set by tier, not published | Resolved volume |
| Per interaction | A turn, session or conversation occurs | Aissist $0.09/interaction, capped at $0.60/resolution; Agentforce $2/conversation; Freddy AI $49 per 100 sessions after 500 | Turns per case |
| Per outcome | A defined result is reached | Fin $0.99/outcome, 50/month minimum, plus $19+/seat on Intercom; Sierra criteria agreed per customer, no public rate | Outcomes, including handoffs |
| Credit-based | The agent spends credits on steps | Salesforce Flex Credits $500 per 100,000, about 20 per action; HubSpot credits $0.010 each, 3,000/month on Professional | Steps per case |
Two of these meter volume: conversations or outcomes arriving. Two meter work: turns and steps inside each case. Per seat meters neither. Work-metered models start cheapest and are the ones Gartner is warning about, because work per case is the number nobody forecasts well. Zendesk's action credits add a credit meter on top of a resolution meter, which is two forecasting problems for the price of one.
"Full automation will be prohibitively expensive for most organizations." — Patrick Quinlan, Senior Director Analyst, Gartner
What does an overshoot look like at fixed volume?
At a fixed 10,000 conversations a month, a rise from five to eight steps per case lifts credit-based and uncapped per-interaction bills by 60%, while volume-metered and capped models move by 0–33%. Volume never changes in this example. Only the work inside each case does: messier orders, more lookups, an extra policy check.
Assumptions, stated plainly: 10,000 conversations a month, 70% resolved by the AI under every model, five actions or interactions per conversation in the budget, eight in production. Rates are each vendor's published figure from the table above.
| Model (published rate) | Budgeted | Actual | Variance |
|---|---|---|---|
| Credit-based (Flex Credits, 20 per action at $0.005) | $5,000 | $8,000 | +60% |
| Per interaction, no cap ($0.09) | $3,150 | $5,040 | +60% |
| Per interaction, capped at $0.60 (Aissist) | $3,150 | $4,200 | +33%, ceiling known |
| Per outcome (Fin, $0.99, before seats) | $6,930 | $6,930 | 0% |
| Per conversation (Agentforce, $2) | $20,000 | $20,000 | 0% |
Credit and uncapped interaction bills grow one-for-one with work, so a 60% rise in steps is a 60% rise in cost. The capped line rises until each resolution hits $0.60, then stops. Budget at the cap ($4,200) and the overshoot is zero by construction.
Volume-metered models do not overshoot on complexity, because a harder conversation is still one conversation. They pay for that stability up front: the per-conversation line costs nearly five times the capped one in this scenario, and it bills failures at full price. For more volumes, see our cost benchmark of AI service or run your own numbers in the price comparison tool.
Why does a price cap exist, and is a spend limit the same thing?
A price cap bounds what each resolution can cost while the agent keeps working; a spend limit stops the agent when a budget runs out. Both control the bill. Only one keeps serving customers on the busiest day of the month.
We re-read eight pricing pages on 26 October to see which controls each vendor offers. Four let you stop the agent at a limit: HubSpot lets you decide how much each feature can spend, Zendesk can pause its AI agent, Fin offers configurable limits and Gorgias lets you adjust usage limits. Salesforce offers alerts, and states there is no overage penalty. Freshdesk states no limit. One, Aissist, caps the price per resolution. Sierra publishes no rate to cap.
A spend cap stops the agent; a price cap stops the bill. If your agent switches off at 3pm on the last Friday of the quarter, your customers will notice before finance does.
"Our success is fully aligned with your outcomes." — Elliot Greenwald, Sierra
The honest framing for Aissist's own model: the cap does not make it cheapest in every scenario. It makes the worst month knowable, which is the property Gartner's subscription-and-bundle comparison group already has. Escalated cases are not charged, per the pricing page. For how the cap changes cost per ticket by region, see customer service cost per ticket.
Does a cap protect you if AI prices rise anyway?
A cap protects you within a contract term, not across renewals. Gartner's January 2026 prediction is that GenAI cost per resolution will exceed $3 by 2030, above many B2C offshore human agents. It names rising data centre costs and vendors pivoting from subsidised growth to profitability.
That is the strongest counter-argument to this article, so it stays on the page. If unit prices rise for every model, choosing a capped or per-outcome model changes your variance, not your trend. A $0.60 cap today can be a higher cap next year.
Two things still follow. First, variance compounds trend: a programme that overshoots 50% in year one rarely survives to argue about year three, which is the Forrester mechanism. Second, the price of a resolution and the work inside it are separate risks. Contracts can fix the first for a term; only the pricing model fixes the second. Ask for both in writing: the rate, and what happens to the bill when each case takes more steps than planned. Our AI agent pricing benchmark tracks those terms across 18 vendors.
Bounded pricing beats cheap pricing when work per case is unknown
The pricing model, not the rate, decides your budget variance. Work-metered models (credits, uncapped interactions, tokens) are cheapest on day one and least predictable by month three. Volume-metered models are predictable and expensive. A cap keeps work-metering's entry price with a ceiling you can budget against. Before signing, model your volume at two step counts, five and eight, and see which line moves. If a vendor cannot tell you the worst month, assume Gartner's 50%.
Frequently asked questions
Why do AI agent bills come in over budget?
Because most AI agents bill for work, not arrivals, and work per case grows as agents take on harder tasks. Gartner predicts uncapped token-linked pricing will exceed expectations by more than 50% through 2027.
Which AI agent pricing model is most predictable?
Per seat and per conversation are the most predictable, because neither moves with the work inside a case. Capped per-interaction pricing is predictable up to its ceiling, which you can budget against.
Is outcome-based pricing the same as per-resolution pricing?
Not quite. Fin bills per outcome, and its outcomes include procedure handoffs to humans as well as resolutions, so check which events count before comparing rates.
How do credit-based AI agent prices work?
You buy credits and each agent step spends some. Salesforce lists Flex Credits at $500 per 100,000, with a standard action using about 20, so a ten-step case costs ten times a one-step case.
Will AI customer service get more expensive?
Gartner thinks so: it predicts GenAI cost per resolution will exceed $3 by 2030. A cap or fixed rate limits variance within a contract term but does not stop prices rising at renewal.
What should I ask a vendor before signing?
Ask what triggers a charge, whether the bill rises when cases take more steps, whether there is a price cap or only a spend limit, and what sits underneath: seats, platform fees or minimums.
