Seats, credits or outcomes: how to evaluate AI agent pricing as a buyer

In brief

Seat, credit, per-conversation and per-outcome prices each push a different risk onto buyers. This guide covers what each exposes you to and what to ask for.

Steve Elliott

Steve Elliott · · 7 min read


On this page
  1. What are you exposed to under each pricing model?
  2. How do seat prices with agent allowances behave?
  3. How do you forecast a credit or per-action bill?
  4. What does outcome pricing actually charge for?
  5. What does the same workload cost under each model?
  6. Which contract terms should a buyer ask for?
  7. Sources

Each AI agent pricing model moves a different risk onto the buyer. Seats fix the bill and leave you paying for people who don't use the agent; credits and per-action meters follow usage but make you forecast a volume you can't yet observe; per-outcome prices tie spend to results, but the vendor's definition of a result decides what you pay.

In our view, the question is not which model is cheapest, but which risk you are best placed to carry and which terms limit the rest. This post compares the models on vendors' published prices and terms as of November 2026, works one illustrative workload through four of them, and lists what we think a buyer should ask for. For definitions, see AI agent pricing.

What are you exposed to under each pricing model?

ModelYou pay forThe bill moves withMain buyer exposurePublished examples, November 2026
Per seatNamed people, monthly or yearlyHeadcount licensedSeats nobody uses; overage once pooled allowances run outMicrosoft 365 Copilot, $30 per person per month paid yearly; Agentforce add-ons, $125; GitHub Copilot Business, $19
Credits or consumptionMetered units from a vendor rate cardVolume times complexity times rateForecast error, expiry of unused credits, agents blocked at a capCopilot Studio credits; Salesforce Flex Credits; GitHub AI credits
Per conversationEach interactionVolumePaying for conversations that failAgentforce Conversations, $2 each
Per outcomeEach result the vendor counts as successVolume times success rateWho defines and verifies the outcomeIntercom Fin, $0.99 per resolution; Zendesk verified resolutions
BundledThe base product licenseRenewal priceAI cost moved into a base price you can't separateServiceNow, AI included in every product since April 2026

Some offers mix rows: GitHub Copilot seats fund a pool of metered credits, and Salesforce's Agentforce 1 Editions, from $550 per person per month, include 2.5 million Flex Credits per org per year.

How do seat prices with agent allowances behave?

We think seats are the easiest line to budget and to waste. Microsoft lists Microsoft 365 Copilot at $30 per person per month on a yearly subscription that auto-renews, on top of a qualifying Microsoft 365 plan. Agent usage inside Microsoft 365 is listed as "Included" for licensed staff, where Copilot Chat alone is "Metered." Microsoft's billing rates page limits this to employee-facing agents acting under the licensed person's identity, within fair usage limits, excluding computer-using agents.

Salesforce offers a direct choice between seat and meter. Agentforce add-ons for Sales, Service and Field Service cost $125 per person per month with unmetered employee use. The alternative is a $5 per-person license that requires Flex Credits, listed at $500 per 100,000; Salesforce's pricing examples use 20 credits per action, or $0.10. On price alone, the add-on pays off above (125 − 5) ÷ 0.10 = 1,200 actions per person per month, about 55 a working day assuming 22 working days. The $5 license covers only limited CRM objects, but the arithmetic shows how heavy use must be before a seat wins.

GitHub turns the seat into funding for a meter. Since June 1, 2026, GitHub Copilot Business and Enterprise have billed by usage: each seat adds a monthly allowance of AI credits, worth $0.01 each, to a pool shared across the organization, and usage is priced per token by model. Extra usage beyond the pool is allowed by default; an administrator must turn the policy off to stop overage charges. Whether each seat earns its keep is covered in our copilot seat renewal review.

How do you forecast a credit or per-action bill?

A credit-based bill is volume times the mix of actions times the rate card.

Microsoft's Copilot Studio rate card (October 31, 2026) charges 1 credit for a classic answer, 2 for a generative answer, 5 for an agent action and 10 for tenant graph grounding. Capacity packs cost $200 a month for 25,000 credits, or $0.008 a credit. Microsoft says prepaid commit units save up to 20%.

In our view, two terms matter more than the rates:

  • What happens at the cap. Microsoft enforces at 125% of prepaid capacity. Agents are then blocked, including in conversations already under way, unless overage bills to a pay-as-you-go plan. Salesforce takes the opposite approach for Flex Credits: no overage penalty, with usage above the entitlement billed monthly in arrears at the contracted rate.
  • Expiry. Unused Flex Credits don't roll over into the next term. GitHub's pooled credits reset monthly.

We think the forecast should be built bottom-up from a pilot, with a separate line for expensive actions such as grounding. Treat it as a driver-based forecast, as in our post on forecasting volatile cloud spend, and re-run it whenever the rate card changes.

What does outcome pricing actually charge for?

Outcome-based pricing pays only when the agent succeeds, so the definition of success is the price.

Intercom charges $0.99 per outcome for Fin over chat and email: a resolution, a procedure handoff, a disqualification or self-serve routing. A sales qualification is $9.99. A resolution counts when the customer confirms the answer was satisfactory, or is assumed when the customer leaves without asking for more help. It charges at most once per conversation, not for failed attempts such as a customer asking for a human, and deducts the charge if the customer returns to the same conversation.

Zendesk introduced resolution tiers on May 18, 2026. Only a "verified resolution" is charged against the allowance: one where a large language model reviews the conversation text and confirms the issue was resolved. Zendesk's help center articles give no per-resolution price and direct pricing questions to sales.

Per-conversation pricing sits in between. Agentforce Conversations cost $2 each, which Salesforce describes as flat pricing for customer-facing agents, and the same org can't run Conversations and Flex Credits together.

In our view, outcome pricing fits best when your own systems can see the outcome. An assumed resolution is a customer who left: a success, or someone who gave up and phoned instead. Check whether your repeat-contact data can test the vendor's count.

What does the same workload cost under each model?

Take an illustrative customer service agent handling 10,000 conversations a month. Assume each conversation uses three generative answers and one agent action, that each of those four steps runs as one Agentforce action, and that the agent resolves 60% of conversations. The products aren't equivalent: Copilot Studio is a kit for building agents, while Fin and Agentforce are finished agents with their own prerequisites. The point is how each bill responds to the same traffic.

PricingArithmeticMonthly billWhat makes it rise
Copilot Studio credits10,000 × (3 × 2 + 5) = 110,000 credits; 5 packs at $200$1,000More actions per conversation
Agentforce Flex Credits10,000 × 4 actions × $0.10$4,000More actions per conversation
Agentforce Conversations10,000 × $2$20,000More conversations, resolved or not
Intercom Fin10,000 × 60% × $0.99$5,940A higher resolution rate

Now suppose each conversation needs a second agent action. Copilot Studio credits rise to 10,000 × (3 × 2 + 2 × 5) = 160,000. Five packs give 125,000 credits and enforcement starts at 156,250, so without pay-as-you-go overage the agent would stop answering around day 29. The Flex Credits bill rises to $5,000. Under Fin, if the resolution rate climbs from 60% to 75%, the bill rises to $7,425. You pay more because the agent does more. Under Conversations, it doesn't move.

Licenses are rarely the whole cost; our post on AI initiative costs covers why labor usually outweighs tokens.

Which contract terms should a buyer ask for?

We think these are worth raising in any agent negotiation. This is not legal advice: confirm contract positions with your counsel and procurement advisers.

  1. A written definition of the billable unit. What counts as an action, a conversation, a resolution or a credit, and who determines it. For outcome pricing, ask for the evidence behind each billed outcome.
  2. Rate card protection. A fixed rate card for the term, or notice before credit rates per feature change.
  3. Overage behavior you choose. Block, alert or bill, set by you, with the overage rate stated. Know the defaults: GitHub and Zendesk allow overage unless turned off, and Microsoft blocks agents at 125% of prepaid capacity.
  4. Rollover or true-down. Carry unused credits forward, or reduce the commitment at the anniversary if adoption is lower than planned.
  5. Usage data you can export. Per agent, team and day, so you can allocate cost and test the vendor's counts. Our post on who owns SaaS spend covers who should hold that data.
  6. A switching path. Salesforce requires all Conversation SKUs to be replaced to move to Flex Credits. Ask what switching models costs before you need to.
  7. Renewal price caps. Promotional seat prices can expire after year one. Microsoft 365 Copilot Business is $18 per person per month paid yearly, down from $21, under an offer that runs to December 31, 2026; the promotional price applies to the first year only.

Bundling, as ServiceNow announced on April 9, 2026, adds one question. Its release says AI is included in every product and is "not a separate purchase," with a tiered offer model and no published prices. In our view a buyer should then ask how much of the renewal increase is AI, so the business case can still be tested.

Whichever model you choose, the agent's cost only means something next to the work it does and the budget owner who answers for it, which is why we think agent spend belongs in the same operating model as the services and outcomes it supports.

Sources

Frequently asked questions

Is per-seat or usage-based AI agent pricing cheaper?

It depends on how heavily each person uses the agent. On Salesforce's published prices, the $125 Agentforce add-on beats a $5 license plus Flex Credits, at 20 credits or $0.10 per action in Salesforce's examples, only above about 1,200 actions per person per month. Seats suit heavy, predictable use; meters suit light or uneven use. Measure use in a pilot first.

What happens when prepaid AI agent credits run out?

It varies by vendor, so check before you sign. Microsoft blocks Copilot Studio agents when a tenant reaches 125% of prepaid capacity, unless overage bills to pay-as-you-go. Salesforce bills Flex Credit usage above entitlement monthly in arrears, with no penalty. GitHub allows extra usage at $0.01 a credit by default unless an administrator turns it off.

How do vendors decide that an AI agent resolved a ticket?

Each vendor uses its own rule. Intercom counts a resolution when the customer confirms the answer helped or leaves without asking for more help. Zendesk charges only for verified resolutions: the customer stops asking for help, then a large language model reviews the conversation text and confirms the issue was resolved. Ask for that definition in writing and test it against your own repeat-contact data.

How should I budget for consumption-priced AI agents?

Build the forecast from measured pilot data: conversations per month, actions per conversation, and a separate line for expensive actions such as data grounding. Multiply by the vendor's current rate card, add headroom for growth, and decide in advance what happens at the cap. Re-run the forecast whenever the vendor changes its rates or your traffic mix changes.

Free playbook

Run the enterprise from any AI.

Read The Headless Enterprise
Steve Elliott

Steve Elliott

Published November 22, 2026

More on AI Economics

Tokens, inference cost, cost per task.