AI agent pricing models explained: per action, per conversation, per seat and flat rate

Short answer: AI agents are sold on five incompatible pricing models, and the model matters more than the number. Per action bills every step the agent takes. Per conversation bills each thread regardless of effort. Per seat bills people rather than work. Credit systems bundle several meters into one balance. Flat rate bills nothing else at all. Two tools advertising similar monthly figures can produce invoices that differ by a factor of ten, and the difference is almost always the unit, not the rate.

Last updated August 2026. Written for US small business owners and operators. Vendor figures were checked at each vendor's own published pricing on August 4, 2026.

What are the AI agent pricing models?

There are five in common use. Per action charges for each discrete step an agent takes, so a single request can bill many times. Per conversation charges a flat amount per thread no matter how much work it takes. Per seat charges per user per month like traditional software. Credit or token systems convert everything into an internal currency you buy in packs. Flat rate charges one subscription and meters nothing.

The reason this matters more than the headline price is that the models make different things expensive. Per action punishes complexity. Per conversation punishes volume. Per seat punishes headcount and is indifferent to usage. Credits punish whatever the vendor decided to weight heavily, which you cannot see until you read the rate card. Flat rate punishes nobody and is priced accordingly at the top end of the simple tiers.

Per action pricing, and the arithmetic nobody publishes

Per action is the model enterprise vendors have converged on, and Salesforce Agentforce is the clearest published example. Its Flex Credits sell at $500 per 100,000 credits, which makes a credit worth half a cent, and a standard agent action costs 20 credits. So one action is about ten cents. A voice action is 30 credits, about fifteen cents. Those figures come from Salesforce's own pricing page and its Flex Credits Rate Card dated April 21, 2026.

The trap is the word action. An action is not an answer to a customer. Resolving one support case might take three actions or thirty depending on how many lookups, record updates and escalations it involves, and you genuinely cannot know your own number until agents are live on your data. That is why per-action vendors publish worked examples rather than a price per outcome, and why budgeting from a competitor's example is close to meaningless. The vendors that do sell by outcome run into the mirror-image problem of having to define one, which we work through using the two clearest examples in the market on Sierra AI vs Intercom Fin pricing.

Where a vendor offers both per action and per conversation, there is a clean breakeven you can calculate in one line. Agentforce also sells conversations at a flat $2 each. At ten cents an action, $2 buys exactly 20 actions. So if your agents typically finish a conversation in fewer than 20 actions, credits cost less; if they routinely take more, the flat conversation price is cheaper and caps your exposure on the messy cases. We work the full stack through on the Agentforce pricing page, including the per-user licenses and the platform underneath, which is the part most comparisons leave out.

Per conversation pricing

Per conversation is the easiest model to explain and the easiest to underestimate. You pay a fixed amount each time someone starts a thread with your agent, and the vendor absorbs the risk of that thread being complicated. For customer-facing support agents this is often the honest choice, because it aligns what you pay with something you can actually count in advance.

The catch is volume, and specifically the cheap conversations. If a third of your threads are someone asking for your opening hours, you are paying full freight for a one-line answer. The model rewards you when conversations are hard and penalizes you when they are trivial, which is the opposite of most people's intuition about it. Before choosing it, look at your existing ticket or chat log and count how many threads resolve in a single exchange.

Per seat pricing, and why it is coming back

Per seat is the model everyone already understands: a price per user per month. It went out of fashion in AI briefly, on the theory that agents do work rather than occupy chairs, and it has come back because finance teams can approve it without a spreadsheet. Salesforce sells Agentforce add-ons at $125 per user per month for Sales, Service and Field Service, and $150 for Industries, with employee-facing usage unmetered inside that.

Watch for the hybrid, which is where per seat gets misleading. A low per-user price sometimes buys access rather than usage. Salesforce's Agentforce User License is $5 per user per month and its own pricing page states plainly that it requires Flex Credits, so the seat fee is the door and the meter is inside. A seat price that does not say what it includes is not a price, it is an entry ticket.

Credit and token systems

Credits are not really a pricing model so much as a layer that hides one. The vendor invents a currency, sells it in packs, and publishes a rate card that converts every operation into it. This is genuinely useful when one balance funds several different products, and genuinely opaque when the multipliers are buried in a PDF.

Three questions cut through it. First, what is a credit worth in dollars? Divide the pack price by the credits in it and every rate on the card becomes readable. Second, do unused credits roll over? Salesforce's rate card is explicit that Flex Credits must be used before the order end date and that no rollover is permitted, which means pre-purchasing more than you need is simply a loss. Third, does testing consume them? In Agentforce, sandboxes and scratch orgs bill at 16 credits an action rather than 20, a discount rather than an exemption, so building costs money before anything goes live.

The other thing to check is whether there are several meters rather than one. Some products bill AI usage, telephony, and data processing on separate scales that all draw from the same balance, so forecasting next month means predicting three unrelated behaviors at once. We took the same look at GoHighLevel AI Employee pricing, where the AI meter stops on the tier called Unlimited and the phone and SMS meters keep running.

Flat rate pricing

Flat rate is one price for the account with nothing metered behind it. Its advantage is not that it is always cheapest, because it is not. Its advantage is that the invoice is the same in a busy month as a quiet one, so nobody rations work to protect a balance. That behavioral effect is larger than most buyers expect. Teams on metered plans routinely underuse them, which is the most expensive outcome available: you pay for capability and then avoid using it.

The fair criticism of flat rate is that heavy users are subsidized by light ones, and vendors offering it have to price for that. Ours is $149 a month, published, with no credit meter, and we would rather defend one number than explain a rate card. The honest framing is that flat rate suits businesses whose usage is uneven and unpredictable, and metered pricing suits organizations with a stable baseline and someone to monitor it.

How to compare two tools on different models

Convert everything to a monthly total for your own usage, not the vendor's example. That means four numbers. Start with the platform: does this tool require another subscription underneath it? Agentforce requires a Salesforce cloud license, which runs $25 to $350 per user per month for Sales Cloud, and that line is often larger than the AI itself. Then add licensing, then estimate consumption honestly, then add anything separately metered like telephony or data processing.

Then apply a sanity check that has nothing to do with arithmetic. Ask what happens in your busiest month, when the work you bought the tool for actually shows up. If the answer is a bill you would have to explain to someone, the model is wrong for you regardless of the rate. If you are already carrying more subscriptions than you can hold in your head, it is worth seeing what your software spend actually totals before you add another meter to it.

Which model should a small business pick?

If your usage is unpredictable and nobody on the team wants to monitor consumption, take flat rate. If you have a stable, well-understood volume of simple, repetitive interactions, per conversation is usually cheaper and easy to forecast. If the agent is a tool your staff use inside software they already live in, per seat is the least friction and the easiest to approve. Per action and credit systems are the right answer when you have someone whose job includes watching the meter, which in a company under twenty people is nobody.

The failure mode worth avoiding is choosing on the smallest published number. A tool advertising two dollars a conversation next to a tool advertising a hundred and fifty a month is not cheaper; it is quoting a line item against a total. Work out both totals for your own volume, then decide. We keep our own arithmetic in public on the AI agent pricing page, and the wider category, including the builders that bill by task run, is laid out on AI agent software.

Common questions

What is consumption based pricing for AI agents?

Consumption based pricing charges for what the agent does rather than who has access. The unit varies: an action, a conversation, a token, or a vendor-defined credit. It aligns cost with usage, which is fair in principle, and it makes budgeting hard in practice because the unit is usually something you cannot count before you deploy. Most consumption vendors offer a free tier specifically so you can measure your own rate first.

How much does an AI agent cost per month?

For a small business, real totals in 2026 cluster between roughly $30 and $500 a month for self-serve tools, with flat-rate agents around $149 and enterprise platforms running far higher once licensing and platform costs are counted. The spread is driven by the pricing model and by whether another subscription is required underneath, not by the quality of the agent.

Why do AI agent vendors keep changing their pricing?

Because underlying model costs move and early guesses about usage were wrong. Salesforce is a documented example: a help article from its October 2025 rate card update lists advanced prompts at a multiplier of 38, while the current rate card dated April 21, 2026 lists them at 16. Both were accurate when written. Always check the date on any pricing article, including this one, before budgeting from it.

Is flat rate always cheaper than per action pricing?

No. A light user running a handful of simple tasks a month will pay less on a consumption plan, sometimes far less, and a free tier may cover them entirely. Flat rate wins on heavy or uneven usage and on the cost of attention, since there is nothing to monitor, forecast or ration. Work out your own volume rather than assuming either direction.