Per seat, per use or per result: how AI is changing what software vendors charge for

A software invoice used to be easy to check. Count the people who log in, multiply by the price, done. If the number looked wrong you could walk round the office and count again.

That is going away, and not only at the large vendors. When the work inside a product is done by a model, the number of people stops being a good measure of anything, so vendors charge for something else: a credit, an action, a conversation, a resolution. Each of those words is defined somewhere in a help article, and the definition, not the price beside it, decides what you pay.

We wrote earlier about the market side of this in SaaS in 2026: seats are shrinking and pricing is moving. This piece is the practical half. What four vendors' definitions actually say, what happened when they changed them, how to read one as a buyer, and how to choose one if you are the vendor.

Three things a vendor can count

People. The seat. Predictable for both sides, and blind to how much work gets done.

Activity. Tokens, credits, actions, minutes, messages. The meter runs whenever the product does something, whether or not it helped.

Results. A support question answered, a lead qualified, an invoice matched. You pay when it worked, and somebody has to decide what "worked" means.

Most of the argument of the last two years has been about the third, and customer support is where it has gone furthest, because a support conversation has an ending you can roughly observe.

What the definitions say

Intercom: a resolution, confirmed or assumed

Intercom charges $0.99 each time its Fin agent resolves a conversation. Its help centre gives two ways that can happen. Either the customer "confirms the answer was satisfactory", or the customer "exits the conversation without requesting further assistance". The second is called an assumed resolution, and it is the one to think about: silence counts as success.

The same article sets limits that matter just as much. A resolution needs "an actual answer to a customer's query", so a greeting does not count. "You will only be charged at most once per conversation." And if the customer comes back unhappy, "that resolution will be deducted and not charged", even when the return falls in a later billing period (Intercom).

Zendesk: three tiers, and only one is counted

Zendesk also bills by the automated resolution, and its definition has moved. Under the older scheme a conversation that was not handed to a person was counted after a period of silence: two hours for chat, 72 for email. A language model then read the transcript to check.

On 18 May 2026 Zendesk split the outcome into three. An assisted escalation is where the AI "contributed to the interaction before a human agent completed the resolution". A contained resolution is where it "handled the interaction to completion without the customer requesting further assistance". A verified resolution is where a second model, reading the conversation, confirms "that the customer's request was satisfactorily resolved". Only the verified kind is taken from the customer's allowance (Zendesk).

Put the two vendors side by side. A customer who reads the bot's answer and closes the window is a billable event at one and, unless a checking model is persuaded, a free one at the other. The headline price per resolution tells you nothing until you know which of those you are buying. Zendesk's help pages do not print a price at all; the figures in circulation, $1.50 on a commitment and $2 without, come from third-party guides and should be confirmed in a quote.

Salesforce: from the conversation to the action

Salesforce launched Agentforce at $2 per conversation. In May 2025 it added a second meter. Flex Credits are sold at $500 per 100,000, an action uses twenty of them, and so an action costs ten cents. Its own explanation of the change is candid: a flat $2 was charged for a question about opening hours and for a long troubleshooting session alike, and "it doesn't scale for customers who want multiple use cases". An action is defined as one of "the functions your agent executes on the platform to get information or perform tasks", and "if your interaction with Agentforce doesn't require an action, there's no charge" (Salesforce).

Its example is a sales agent that answers a question about two products and books a meeting: three to six actions, so $0.30 to $0.60 instead of $2. Notice what moved. The customer now pays less for a simple exchange and carries the risk of a complicated one, and pays for the actions whether or not the meeting was worth having.

The pricing page today shows all three kinds of unit together: credits, the $2 conversation, and add-ons at $125 per user a month (Salesforce). That is not indecision. Different buyers can budget for different things, and a vendor that size can afford to run three meters.

HubSpot: half the price, a narrower unit

In April 2026 HubSpot moved its customer agent from $1 per conversation to $0.50 per resolved conversation, and its prospecting agent from a monthly fee to $1 per lead recommended for outreach. Its chief executive, Yamini Rangan, put the reason in one sentence: "there is a huge gap between AI output and AI outcomes" (Constellation Research).

Cursor: changing the unit is harder than changing the price

The cautionary tale is not from support. In June 2025 the code editor Cursor replaced a plan of 500 requests a month with "$20 of frontier model usage per month at API pricing", plus unlimited use of models it chose automatically. Many customers read "unlimited" more widely than intended and were charged for the difference. On 4 July the company wrote: "We were not clear that 'unlimited usage' was only for Auto and not all other models", and "We recognize that we didn't handle this pricing rollout well, and we're sorry." It refunded the unexpected charges (Cursor).

The price had not gone up. The unit had changed, from something a customer could count on their fingers to something they could not, and that was enough.

If you are buying: eight questions for the definition

Find the help article that defines the unit, not the pricing page, and read it with these in hand.

1. Who decides that it happened? The customer clicking a button, the vendor's rule, or a model reading the transcript. If it is the vendor's model, can you see its verdict for each case?

2. Does silence count? If yes, how long a silence, and can you change the window?

3. What happens when the customer comes back? Is the charge reversed, and for how long afterwards?

4. Once per conversation, or per step? A resolution is counted once. Actions and credits are counted every time.

5. What is never counted? Tests, sandbox use, hand-overs to staff, greetings. Zendesk, for one, excludes testing and sandbox environments explicitly.

6. Can I set a ceiling? A monthly cap, an alert at half, a hard stop. If the only control is the invoice, it is not a control.

7. What happens to what I don't use? Committed volume and prepaid credits may expire. Ask when.

8. Can I export the list? Every counted unit with a date and a link to the conversation. Without it you cannot audit the bill, and the old trick of walking round the office has no equivalent.

Then run your own month through it

Take a support desk with 4,000 conversations a month and ten agents, and suppose the bot settles half of them.

The prices are the list figures quoted above; four actions per conversation is our assumption, inside the three to six of Salesforce's own example. The point is the shape. Under result pricing the bill rises as the product gets better, which is fair, provided each result is worth more to you than it costs. Under activity pricing the bill is the same whether the bot helped or not. Under seat pricing nothing moves at all. None is wrong. They put the risk in different places, and you should know which place before you sign.

The other half of the sum is what a resolved ticket costs you today with people, counting the ones that bounce back. We built a calculator for exactly that.

If you are selling: how to pick a unit

A small vendor reading the cases above may conclude it should charge per result because that is where the big names are going. Be careful. They can absorb a bad quarter of margins. You may not.

Know what one unit costs you, including the failures

With seats, another user cost you almost nothing. With a model in the product, every attempt costs money, and under result pricing you are paid only for the attempts that succeed. Suppose each attempt costs you $0.12 in model and hosting fees and you charge $0.99 per result.

The $0.12 is an illustration; put in your own figure. What the table shows is that your margin now depends on something your customer controls. A customer with tidy help pages gives you 70%. A customer with none gives you 20% and costs you three and a half times as much per dollar earned, on the same price list. Bessemer's 2025 review of AI startups found the fastest-growing group running at about 25% gross margin, against about 60% for the steadier group (Bessemer). A unit chosen without this sum is one way to end up in the first number.

Four tests for a unit

The customer can count it without you. Documents converted, messages sent, reports run. If they need your dashboard to believe the bill, expect arguments.

It is clearly yours. A closed support ticket is plainly the bot's work. "Revenue influenced" is not plainly anyone's.

It moves with your cost. If one unit can take three model calls or three hundred, either narrow the unit or put a limit inside it.

Neither side can game it. A unit the customer can split in two to pay less, or you can pad to charge more, will be.

Where most small vendors should land

A base fee that covers your fixed cost of serving the account, an allowance of units inside it, and a plain price for more, with a cap the customer sets. Sell extra units as a prepaid balance, because card fees eat very small charges. It is less exciting than "pay only for results", and it keeps you solvent while you learn what a unit really costs.

Keep pure result pricing for a result that is unambiguous, logged, and reversible when it turns out to be wrong. Write the definition as carefully as the vendors above have, publish it, and give the customer the list.

And if you already have customers on one unit, remember Cursor. Announce the change before it bites, show each customer their last three months restated in the new unit, and let existing plans run to renewal.

One more buyer is arriving for whom the seat means nothing at all: software acting for a person, with a spending limit and no login. We looked at what that needs from a price list in agents with wallets.

Sources

Intercom Help, Fin AI Agent resolutions; Zendesk Help, About automated resolution tiers and About automated resolutions for AI agents (the scheme before 18 May 2026); Salesforce, Flex Credits (23 May 2025) and Agentforce pricing; Constellation Research, HubSpot to price Breeze Customer Agent, Breeze Prospecting Agent on outcomes (2 April 2026); Cursor, Clarifying our pricing (June and July 2025); Bessemer Venture Partners, The State of AI 2025 (13 August 2025). All pages read on 12 October 2026. The two worked tables are our own arithmetic; the support desk, the four actions per conversation and the $0.12 cost per attempt are assumptions, stated where used.