AI Sales Agent Pricing Models Explained: Per Resolution, Per Conversation, Per Seat, Flat

TL;DR: AI sales agents are priced on four main units: a fee each time the AI produces an outcome, a fee for each conversation, a fee for each person on your team, or a flat monthly price with a set number of conversations included. Many real price lists mix two of them. The unit matters more than the headline number, because the unit decides what the vendor earns more from, and so what the product is quietly built to do. Below, each model is explained in plain terms, priced against the same worked month, and tested against one question: can an agent priced this way afford to tell a poor-fit visitor no? We argue flat pricing answers it most cleanly: with a flat price we earn nothing extra from a yes, so we have no reason to tune the agent to say yes. We use it, and we say where it has edges too.

AI sales agent pricing pages are hard to compare for a simple reason. They do not just charge different amounts. They charge for different things.

One vendor bills you when its AI decides a visitor is qualified. Another bills you for every chat that starts. A third bills for every person on your team who logs in. A fourth bills the same amount every month whatever happens. Put those four side by side and the numbers on the page tell you almost nothing, because they are not measuring the same thing.

We have written about what AI sales agents cost across the best-known vendors, and about the success tax built into outcome pricing. This post is the explainer that sits underneath both: the four units you can be billed in, what each one rewards, and how to work out your own month under each before you speak to anyone's sales team. If you are new to AI sales agents, start with what an AI sales agent is and come back.

Why the unit matters more than the number

A price is two things: an amount and a unit. "Fifteen dollars" means nothing until you know fifteen dollars per what.

The unit matters for a reason beyond arithmetic. Whatever a vendor bills for is the thing it earns more from when it goes up. Product teams are not villains, but they do build towards what pays. Over time, a product drifts towards producing more of whatever its price list counts.

So when you read a pricing page, ask two questions in this order:

1. What is being counted? Outcomes, conversations, people, or nothing.
2. Who decides when the count goes up? You, your visitors, or the vendor's own AI.

The second question is the one many buyers never ask, and it is the one that separates the four models most sharply.

The month we will price

To keep this fair, every model below gets priced against the same business in the same month. The figures are an illustration, not a prediction, and you should swap in your own.

• Website conversations this month: 200. A conversation is one visitor chatting with the agent.
• Visitors the AI judges qualified: 30 of the 200.
• Other conversations where the AI answers the question or reaches some other outcome: 100.
• Conversations that end without an outcome: the remaining 70.
• People on your team who handle the leads: two.
• What you want at the end: a booked video meeting in your calendar, with the details in your CRM.

Then, for each model, a second month: the one where your marketing works and everything doubles. The best month is where pricing models show what they are made of.

Model one: per resolution, or per outcome

What it counts. Each time the AI produces a defined result: a support question resolved, a lead qualified, a meeting handed off. Different results often carry different fees. Usually one outcome is charged per conversation, and a conversation with no outcome costs nothing.

Where it came from. Customer support, where it makes real sense. In support, a resolved ticket is the goal, and paying per resolution means paying only when the work is done. Several AI support products price this way, and for support it is a defensible model.

What happens when it moves to sales. In sales, the most valuable outcome is a qualified lead, so that becomes the most expensive line on the price list. And here is the catch: you may set the criteria, but the AI that decides whether a visitor meets them is the vendor's own AI, and the vendor bills on that verdict.

The worked month. The best-known per-outcome price list is Fin for Sales, checked on its pricing page in October 2026. A seat is from US$19 a month billed annually, or US$29 monthly. Each conversation carries one charge: US$9.99 for a qualified lead and US$0.99 for any other outcome. Its booking step needs a scheduling tool you connect: HubSpot Meetings is free, Calendly needs a paid plan, or Chili Piper. Competitor prices here are in US dollars, as published; ours are in Australian dollars. Our comparison page carries those figures with their source.

Against our month:

Line	Calculation	Amount
One seat	billed annually	US$19
Qualified leads	30 at US$9.99	about US$300
Other outcomes	100 at US$0.99	just under US$100
Scheduling tool	HubSpot Meetings, free; Calendly needs a paid plan	from US$0
Month total		about US$418

The best month. Double the traffic: 400 conversations, 60 qualified, 200 other outcomes. The total climbs to roughly US$816. Nothing about the product changed. It worked, so it cost more.

What it rewards. A higher count of outcomes, and especially of the expensive one. That does not mean any vendor is inflating its verdicts. It means nothing in the price list pushes the other way.

Ask: what exactly counts as each outcome, who decides, and can you see and dispute the count?

Model two: per conversation

What it counts. Every conversation that starts, whatever happens in it. A visitor who books, a visitor who asks one question and leaves, and a visitor who gets a polite no all cost the same.

The worked month. 200 conversations at the vendor's rate. Some vendors publish a rate (Salesforce Agentforce, for example, is US$2 per conversation). Either way, the shape is simple: the bill is your traffic multiplied by a number.

The best month. 400 conversations, twice the bill. The bill moves with traffic, not with how well the agent did.

What it rewards. More conversations, which is mostly in your interest anyway. Crucially, it is neutral on the verdict. A yes and a no cost the vendor and you the same. That is a real improvement on per-outcome pricing.

Where it bites. Two places. First, the definition: does a conversation start when the chat window opens, when the visitor types, or when the agent replies? Does a bot or a spam message count? Second, predictability: a pure per-conversation bill has no ceiling and no floor, so every month is a forecast.

Ask: what starts a conversation, what ends one, whether spam and bots are counted, and what happens to your bill in a month with an unusual spike.

Model three: per seat

What it counts. People on your side who have a login: the reps, managers and admins who see the leads, transcripts and reports.

The worked month. Two seats at the vendor's rate. Your website traffic does not appear anywhere in the calculation.

The best month. The same as any other month, unless your team grows.

What it rewards. More people inside the product. That sounds harmless, and often is, but notice the mismatch. An AI sales agent's whole job is to do the work without people: answer at nine at night, qualify, book. Its value grows with your visitors, not your headcount. A seat price measures the one thing the agent is designed to need less of.

Where it bites. Seat creep, where reports, transcripts or approvals are only visible to paid seats, so the product slowly teaches you to buy more. And in our reading, pure seat pricing is uncommon for AI agents. More often, seats are the base layer and a per-outcome or per-conversation meter sits on top, which is the hybrid you saw in model one.

Ask: what can someone without a seat see, and is there a usage meter on top of the seats?

Model four: flat, with conversations included

What it counts. Nothing per event. You pay a fixed monthly price for a tier, and each tier includes a set number of conversations. If you go past it, there is an overage rate per conversation, which should be printed on the pricing page.

The worked month. This is the model we use, so here is ours, exactly as published on our pricing page. Starter is $99 a month with 50 conversations included, and Growth is $199 with 200. Our 200-conversation month fits inside Growth. The bill is $199. The qualified leads, the other answers, the booked meetings and the conversations that went nowhere all cost the same: nothing extra.

The best month. 400 conversations on Growth means 200 past the allowance. Each one is $0.50. On a paid plan the agent keeps answering through all of them: it does not go quiet when you pass the allowance, because a buyer who arrives on the 29th of the month should not meet a dead chat window. You get an email as you reach half, three quarters and nine tenths of your allowance, so the extra never arrives as a surprise, and any overage is added to your next invoice. If that becomes your normal month, Pro is $449 with 500 conversations included.

What it rewards. Nothing that happens inside a conversation. The vendor earns the same from a booking and a polite decline. What it does reward is keeping you as a customer next month, which is the incentive you actually want a vendor to have.

Where it has edges. We would rather say these ourselves. A flat tier is not "unlimited", and anyone who tells you theirs is should be asked what happens at the limit. Overage is still per conversation, so a flat plan has a small per-conversation meter at its edge. And if your volume swings wildly, from 50 one month to 900 the next, a tier sized for one of those months will be wrong for the other. The honest test is whether the tiers are sized so that a normal month for a business your size fits inside one, and whether the overage rate is printed where you can see it before you sign up.

Ask: what is included, what is the overage rate, what happens to the agent at the limit, and is there a contract.

The four side by side

Model	Counts	Best month	A "no" costs the vendor
Per resolution	Outcomes, priced by the vendor's AI verdict	Rises with success	A smaller fee: about a tenth of a yes, in the best-known case
Per conversation	Every chat started	Rises with traffic	Nothing
Per seat	Your team's logins	Flat unless the team grows	Nothing, but the seats price the wrong thing
Flat, conversations included	Nothing per event, overage past the allowance	Flat, then a printed rate	Nothing

The one under which an agent can afford to say no

This is the question we care most about, because it decides what kind of agent you end up with.

A good sales agent turns some people away. Not rudely, and not because they asked too many questions. Some visitors are simply not a fit: wrong size, wrong country, wrong need, or a budget that does not match what you sell. Booking them wastes your team's time and theirs. We wrote about why we built an AI that says no, and about how an agent should judge who is worth a meeting by facts rather than behaviour.

Now look at each model from the vendor's side.

Per resolution: every visitor judged qualified is the most valuable event on the price list. A visitor turned away earns the vendor a smaller fee: about a tenth of a yes, in the best-known case. The product may still decline people. The price list simply pays more for the yes.

Per conversation: the decline costs nothing extra, because the conversation was already counted. The agent can say no freely. The bill still moves with every chat, including the junk ones.

Per seat: the decline costs nothing, but the model prices the people reading the leads, not the work the agent does.

Flat: the decline costs nothing, the booking costs nothing, and the bill does not move inside the allowance. The agent's job and your interest point the same way: book the right people, warmly turn away the wrong ones, and leave the conversation saved either way.

That is why we chose flat. On Growth and above, SureGate lets you write that door policy yourself, in your own rules, with three outcomes: booked, passed to a person for review, or a warm no with the transcript saved. With a flat price we earn nothing extra from a yes, so we have no reason to tune the agent to say yes.

How to price your own month in twenty minutes

Before any demo, collect five numbers.

1. Your monthly conversations. If you already have a chat window, count last month's chats. If you do not, start from your monthly website visitors and assume a small share will start a conversation. Be conservative and then check it in a trial.
2. The share that would qualify. Your sales team knows roughly what fraction of inbound enquiries are worth a meeting. Use that.
3. The people who need access. Who reads transcripts, who takes the meetings, who runs reports.
4. Your booking setup. Does the agent book the meeting itself, or hand off to a scheduling tool you pay for separately?
5. Your best month. Double the first number. That is the month that tells you whether the price list has a ceiling.

Then run all five through each vendor's model. Per-outcome vendors need numbers 1, 2 and 4. Per-conversation vendors need 1. Per-seat vendors need 3, plus any meter. Flat vendors need 1 and their printed overage rate.

If a vendor cannot turn those five numbers into a monthly figure in writing, that is useful information too.

The bottom line

Four units, four sets of incentives. Per resolution counts the vendor's own verdicts, so it rises with success and charges for every yes. Per conversation is neutral on the verdict but moves with every chat. Per seat prices the people the agent is meant to save. Flat with conversations included makes a booking and a polite no cost the same, and moves only when your volume genuinely changes, at a rate printed in advance.

None of this tells you which product is best. It tells you what each price list is quietly asking the product to do. Read that first, and the numbers make sense.

If you want to see the flat model on your own site:

• Read the pricing page: every tier, every allowance and the overage rate are printed on it.
• See the comparison table, with sources and dates, including the rows where others lead.
• Start a 14-day free trial on Starter, no credit card, and count your real conversations before you pay for any of them.