Is seat-based pricing dead in the age of AI agents?

Short answer: no, seat-based pricing isn’t dead. But it was built for a world where every user was a person, and that world is ending faster than most pricing pages have noticed.

Think about what a seat actually priced. It priced a human. A person logs in, works a few hours, opens some records, sends some emails, goes home. Even your busiest power user has natural limits: a working day, a pair of hands, a need to sleep. SaaS pricing was quietly built around those limits. The cost of serving one seat was small and predictable, so you could charge a flat monthly price per person and keep 80–90% gross margins.

Now put an AI agent in that seat. It doesn’t log in, it calls the API. It doesn’t work nine to five, it runs all night. It doesn’t open twenty records, it reads every record, then reads them again an hour later to see what changed. Jason Lemkin of SaaStr described one of his own agents making more than 30,000 API calls a day to a single product. No human seat was ever priced for that.

The squeeze from both sides

That’s why AI doesn’t just make seat pricing a bit worse. It squeezes it from both ends.

On the revenue side, the customer needs fewer people. If your agent answers tickets or qualifies leads, the same work gets done by a smaller team, and a smaller team buys fewer seats at renewal. Buyers know it: in BCG’s survey of IT buyers, 40% said cutting seats was their main way to reduce software spend. Gartner estimates up to $234 billion of application software spend, about a fifth of enterprise SaaS, is at risk by 2030 because agents can do the work without anyone opening the app.

On the cost side, the work didn’t disappear. It moved to agents, and agents are far more expensive to serve: model costs on every task, plus all the extra load from API calls, reads and syncing. Bessemer puts typical gross margins for AI companies at around 50–60%, against 80–90% for classic SaaS.

Fewer seats means less revenue. Heavier usage means higher costs. And the better your AI gets, the worse both become. Your product team is trying to make the agent do more, while your pricing model punishes you every time it does.

How vendors are reacting

Companies are feeling their way through this in three ways.

Some are putting a meter on agent access. Slack cut API rate limits for commercial apps sold outside its marketplace in May 2025, citing AI systems pulling out customer data. This month, Salesforce started charging Flex Credits for every call an external agent makes through its API or MCP server.

Some are treating the agent as a new kind of user. Microsoft’s Agent 365 gives agents their own identities in a company’s directory, licensed and shown in the org chart like an employee.

And some are pricing the result instead of the access. Intercom charges $0.99 per resolution for its Fin agent. Sierra charges only when its agent resolves an issue, or as Bret Taylor puts it, “let’s pay for a job well done.”

The first approach has a real risk, and Lemkin names it. Meter every agent call too hard and customers build around you: their agents copy your data into their own database once and read it from there for free. Less work flows through your product, it gets less sticky, and renewal gets harder. He calls it an agentic death spiral. The price has to bear some sensible relationship to cost and value, or customers will route around it.

What HubSpot did, in three steps

HubSpot is the clearest example of a company working this out in public.

In March 2024 it actually went towards seats. HubSpot moved every hub to seat-based pricing and made view-only seats free on paid accounts. Charge the people who do work in the product; let everyone else look for free. A seat is a fine unit when it really maps to a person doing work.

In June 2025 it added a meter for the AI. Alongside its Customer Agent, HubSpot introduced HubSpot Credits: a typical agent conversation cost about $1, and each plan came with a monthly allowance. Seats covered the humans; credits covered the AI. It also built the guardrails any meter needs, with spending limits and alerts as you approach them.

In April 2026 it moved to paying for results. The Customer Agent went from $1 per conversation to $0.50 per resolved conversation. HubSpot said it could afford to because the agent now resolves around 65% of conversations.

Notice the order: seats, then a meter on activity, then outcome pricing for one clean, countable job once the agent was reliable enough. HubSpot never threw seats out. It built on top of them.

Is hybrid a stepping stone, or the destination?

There are three ways this could go.

One is that we reinvent the seat for agents: a per-agent licence, priced like a digital employee. Easy to budget, but it has the same flaw as the human seat, only worse. One agent might run a few tasks a week; another, tens of thousands of calls a day. A flat price overcharges the first and loses money on the second. Add an allowance to fix that and you’ve built a hybrid anyway.

Another is that hybrid is a halfway house and everything ends up on pure usage or outcomes. The evidence points the other way. Salesforce tried per-conversation pricing for Agentforce, then credits per action, and by late 2025 Marc Benioff said seats bundled with credits had become the preferred option, because customers wanted a number they could budget for. Microsoft still sells its flagship Copilot at $30 per user per month. Finance teams don’t want a bill that swings with something they can’t see.

The third, and the one I’d bet on, is that hybrid is the destination, but the balance keeps shifting. Think of a mobile phone plan: a fixed line fee plus data. The fixed part gives the buyer a predictable number and covers your base cost. The variable part grows as agents take on more of the work. Over time, seats shrink to the people who set up and manage the AI, and the meter becomes most of the bill.

The data fits. Kyle Poyar’s 2025 survey of 240+ software companies found hybrid became the main model for 41% of companies, up from 27% a year earlier, while pure seat pricing fell from 21% to 15%. Only 5% used outcome pricing as their main model.

So the future isn’t pricing per human or per agent. It’s pricing the people by access, and the agents by the work.

How to work out your own price

This is where product marketing earns its place. The job isn’t to pick a number. It’s to find the unit the customer already cares about, prove what it’s worth, and make the price feel fair and predictable. Madhavan Ramanujam’s Monetizing Innovation and Thomas Nagle’s The Strategy and Tactics of Pricing both come down to the same idea: price from the customer’s value, not your cost. Applied to agents, it goes like this.

Start with the unit the customer already counts. Support leaders think in tickets, sales leaders in meetings, legal teams in contracts reviewed. Nobody budgets in tokens. HubSpot moved to resolved conversations because a resolution is what a support leader is measured on.

Work out what one unit is worth: what does the job cost the customer today without you? Then take a share of that value, not all of it, so switching is a clear win.

Then check your cost floor, which is where agents change the maths. With a human seat, cost to serve was close enough to zero to ignore. With an agent, you have to add up model costs, platform load and the failed attempts you don’t get paid for.

Here’s an example with made-up numbers:

Step	Illustrative number
Customer’s cost for a person to handle one ticket	$6.00
Your price per resolved ticket	$1.00
Customer’s saving per ticket	$5.00
Your cost per resolution, including failed attempts	$0.30
Your gross margin	70%

You capture about a sixth of the value, the customer saves $5 a time, and you keep a healthy margin. But if your agent only resolves half the conversations it starts, you pay for twice as many attempts per paid resolution, and cost can climb towards $0.80. Same price, 20% margin. With outcome pricing, your agent’s reliability is your margin.

Finally, package it so it’s predictable: a base fee or seats with an allowance included, an obvious overage rate, and alerts and caps live before the first variable bill.

That last step is where things go wrong. When Cursor switched its $20 plan from a set number of requests to $20 of model usage in June 2025, heavy users ran out fast, some were overcharged, and within three weeks the company apologised and offered refunds. A meter without clear limits reads as a hidden price rise. Usage pricing cuts the other way too: Snowflake cut its revenue forecast in May 2023 simply because customers used less. And outcome pricing only works for clean, countable jobs. As the team at Lago points out, pay an agent per meeting booked and you’ll get plenty of bad meetings.

How companies will make the change

Nobody should flip their whole customer base overnight. Going by what HubSpot and Salesforce have done, I’d expect most companies to move in this order: keep seats for people and add credits for AI with a generous allowance; meter agent access to the platform at a price that reflects real cost; move clean, countable jobs to outcome pricing once the agents are reliable; and over time redefine the seat as “the people who manage the agents”.

For existing customers, introduce the change at renewal, show each customer what they’d pay under both models using their own data, and talk to sales leadership early. If reps are paid on seats, they’ll keep selling seats.

So, is the seat dead?

No. It’s shrinking back to what it was always really pricing: a person. The people who set up and work alongside the AI will keep their seats. The agents doing the work will be priced on the work.

And pricing isn’t a one-off decision anymore. HubSpot changed its model three times in about two years. Ask the question every year: what is the customer actually paying us for now, and what does it cost us to deliver?