On April 14, 2026, HubSpot cut the price of its Breeze Customer Agent from $1.00 per conversation to $0.50 per resolved conversation. Intercom's Fin charges $0.99 per resolution. Salesforce Agentforce charges $2.00 per conversation whether or not anything gets solved. Same category, a 4x spread, three completely different bets about what a customer should pay for. If you're trying to figure out how to price AI agents in 2026, that spread is your first clue that the number on the page is the easy part.
The hard part is the question almost no one asks out loud: is the price you picked profitable? You can charge $0.50 a resolution and print money, or charge $0.50 a resolution and lose it on every customer. The difference isn't the price. It's the cost and the resolution rate sitting underneath it neither of which appears on your pricing page.
So this playbook treats pricing AI agents as two decisions, not one. First, pick a model your customer will accept. Second, prove that model makes money as usage scales. Most teams nail the first and skip the second, then watch margin erode as their best customers use the product more. Let's do both.
Why per-seat pricing broke for agents
The seat assumed a human. One person logs in, does a bounded amount of work, you charge for the login. An AI agent has no seat it has API calls, tokens, tool calls, and completed workflows, and a single customer can trigger thousands of them overnight. As a16z put it, per-seat is no longer the atomic unit of software.
The math turned against seats quickly. The share of SaaS companies on pure seat-based pricing fell from roughly 21% to 15% in twelve months, while hybrid pricing a base fee plus usage climbed from 27% to around 41%. Gartner expects about 40% of enterprise SaaS to include outcome-based components by 2026, up from 15% in 2022. Zendesk is the cautionary tale: enterprise customers pay roughly $115 per support agent per seat, but when an agent resolves tickets autonomously the customer needs fewer humans, so seat revenue shrinks exactly as the software gets more valuable. That's structural, not a pricing tweak.
There's a second force that makes agent pricing different from old SaaS: variable cost of goods. Commodity inference has collapsed by one a16z estimate, from about $60 per million tokens in late 2021 to roughly $0.06 three years later yet frontier models hold premium pricing, and your cost per agent run moves every time you change models or a customer changes behavior. Traditional SaaS never had to model that. The marginal cost of one more seat was near zero. The marginal cost of one more agent run is real, variable, and yours.
The four ways to price an AI agent
Per-seat (demoted, not dead) Charge per user with access. It still works when the agent augments a human rather than replacing the workflow, when usage is naturally bounded, and when the buyer wants a predictable, procurement-friendly number. The margin risk is low when value roughly equals access, and high the moment one seat can drive unbounded usage that's when a flat fee quietly subsidizes your heaviest users. If you keep seats, add fair-use ceilings and watch cost per seat, not just seat count.
Usage-based (tokens, calls, compute) Charge for what's consumed. This is the default for AI API companies because cost and value both scale with usage, so margin can track consumption instead of fighting it. The risk is that raw usage pricing is commoditizing, gives you little pricing power, and exposes you to model-cost swings: if you mark up off last quarter's token price, a frontier upgrade can erase your spread. Mark up off a blended, fully-loaded cost per unit with headroom for price changes, and reprice as inference deflates rather than letting old margins coast.
Outcome-based (per resolution, lead, booking) Charge per result. This is the model every enterprise buyer says they want, and the marquee names are already here: Intercom at $0.99 per resolution, HubSpot at $0.50 per resolved conversation, Salesforce Agentforce at $2.00 per conversation, and Sierra running pure outcome pricing on its way past $150M ARR and a reported $10B valuation. Bret Taylor calls it "obviously the correct way to build and sell software."
He's right about the alignment and wrong to make it sound simple, because outcome pricing carries the sharpest margin trap of all: you can only safely price a result you can cost, and you eat the cost of every failed attempt. HubSpot can afford $0.50 a resolution partly because its CRM context drives a 65% resolution rate across 8,000+ customers it rarely pays for many misses. A startup without that context, charging the same $0.50 but resolving 40% of attempts, is paying model and infrastructure cost on all 100% while billing for 40%. Same headline price, very different P&L. Before you sell an outcome, know your cost per successful outcome failures included and make sure you can prove the result, or you'll lose the attribution argument at invoice time. This is exactly where teams discover whether their outcome-based pricing is actually profitable.
Hybrid (base + usage or outcome) A subscription base plus a usage or outcome layer on top. This is now the dominant real-world model somewhere around 41–46% of companies, depending on whose 2026 survey you read because it gives procurement a stable number to sign and gives you margin upside as usage grows. The trade-off is complexity: with two billing layers, margin leaks are easier to hide, especially in the usage tail. Instrument both layers from day one and keep an eye on the customers whose usage component is quietly outrunning their base.
A decision tree you can use
Work down this list and stop at the first honest "no" it points you to your model. * Can you measure cost per run accurately, today? If no, start with seat or simple usage pricing and instrument before you get fancy. * Can you prove the outcome, and will the customer accept your definition of "done"? If no, outcome pricing will bleed you in disputes use usage or hybrid. * Does the customer already measure that outcome and budget against it? If no, you'll spend the sale educating them, and hybrid lowers the friction. * Is usage bounded or unbounded per customer? Bounded leans seat or flat; unbounded demands a usage or outcome layer so cost can't outrun price. * Does procurement need a predictable committed number? If yes, wrap whatever you chose in a hybrid base.
For most AI agent companies in 2026, that path lands on hybrid with an outcome or usage layer which is also why hybrid won the adoption contest. If you want the full comparison, here's how to choose between usage-based, outcome-based, and hybrid pricing.
The decision everyone skips: pricing for profit
Here's the part the model debate ignores. Revenue is the visible half of the equation. The half that decides whether you survive looks like this:
Agent profit = value captured − (cost per run × number of runs)
You can see the left side on your dashboard the day you launch. You usually can't see the right side until margin erupts. Run the math on a support agent priced at $0.50 per resolved conversation. Say each attempt costs $0.18 fully loaded model tokens, tool calls, infrastructure, retries and your resolution rate is 50%. To resolve 500 conversations you attempt 1,000. Cost: $180. Revenue: 500 × $0.50 = $250. Your margin is 28%, not the 64% the sticker price implies, because you paid to attempt the 500 you never billed for.
Now add a power user who hammers the agent with harder, multi-step requests at a 35% resolution rate. Their cost per resolved conversation climbs above your $0.50 price, and that single customer goes margin-negative while your top-line revenue keeps rising. That's the cost eruption, and it's how a "profitable" AI product ends up losing money on your power users without anyone noticing until the quarterly numbers come in.
The lesson isn't "avoid outcome pricing." It's that the same price is healthy or fatal depending on cost per run and resolution rate, and those two numbers belong in the pricing decision from the start not in a post-mortem.
A checklist before you ship a price
Before you publish any agent price, confirm you can answer these: * What is the fully-loaded cost of one agent run, including retries and failed attempts? * What is that cost per customer and per use case, not just on average? * Does the price hold margin if your model provider raises prices or a customer's behavior shifts? * Are unbounded usage paths capped, banded, or metered so cost can't outrun price? * Is metering live before launch, so you learn from data instead of a surprise invoice? * Is there a repricing cadence on the calendar? Bessemer suggests revisiting pricing roughly every six months the market moves too fast for an annual review.
Pricing AI agents well in 2026 is less about picking the cleverest model and more about refusing to ship a price you can't defend on margin. Pick the model your customer accepts, then prove it makes money at the messy edges where your biggest customers live. The teams that do both will quietly out-survive the ones still arguing about which model is "right."
Paygent tracks the real cost and margin behind every agent, customer, and use case so the price you put on the page is one you can prove makes money. See your true per-agent margins before you scale.
