The Seat Did Not Die. It Got a Meter on Top

In February 2026, the market decided that software had a terminal illness and the diagnosis was per-seat pricing.

It was a specific, datable panic. Anthropic introduced Claude Cowork on the 12th of January, an agent that reads and writes files on your desktop. On the 30th it released eleven open-source plugins, one each for legal, sales, finance, marketing and customer support. Jefferies’ Jeffrey Favuzza coined “SaaSpocalypse” on the 3rd of February and downgraded Workday and DocuSign. The S&P Software & Services Index lost 25% of its value between the 12th of January and the 23rd of February. Four days after the plugin wave it fell 4% in a session and took more than $285 billion with it; JPMorgan’s software index fell 7%; LegalZoom tumbled nearly 20% and Thomson Reuters 16%. In January, monday.com’s CEO announced the company had replaced its entire 100-person SDR team with AI agents, and the cascade argument followed everywhere: one human replaced, five SaaS licences cancelled, revenue evaporating.

The conclusion everyone drew was neat and, I think, wrong. If agents do the work, seats are worthless; if seats are worthless, the replacement is the outcome. So the vendors pivoted to outcome pricing — per resolved ticket, per conversation, per action — and the industry declared the seat dead.

I went looking for the seat and could not find it anywhere. What I found instead was the seat with a meter bolted on top of it, a floor that survived every repricing, and a definition of “outcome” that the seller controls. The seat is not dead. It is the most load-bearing thing in enterprise AI right now, and the reason is simple: it is still how the money arrives.

The supposed successor is the smallest category

Start with the distribution, because the narrative and the data have come apart.

Bain’s read on AI pricing is the most rigorous survey available, and it does not say what the narrative says. Of vendors introducing AI pricing, roughly 10% use outcome-based meters. About 35% price on effort — tokens, compute, agent hours — and about 55% price on output: the summaries produced, records updated, leads recommended. Meanwhile about one in five AI-native software companies still relies mostly on per-seat licensing, usually with usage entitlements layered on. And roughly four in every five vendors introducing AI pricing have chosen capacity models: prepaid pools, committed spend, no rollover.

Maxio went and read 107 vendor pricing pages. Pure outcome pricing appears on fewer than one in five of them. Harvey, Gong and Fathom are all marketed as agent products and all keep a per-user-per-month floor with usage or outcome charges stacked on top.

So the axis of the whole argument was wrong. Outcome pricing is not the future that won. Capacity pricing is. The prepaid credit pool — Workday’s Flex Credits, Salesforce’s Flex Credits, monday.com’s credit bundles — is the actual successor, and it is a very old model wearing new clothes: you commit a budget, you get a pool, the vendor meters what you do with it, and there is no annuity. The difference between a seat and a credit is that a seat is something you can count yourself and a credit is something you have to ask the vendor about.

The floor is still a seat

Here is the part that should have ended the argument in February and did not. Every vendor doing outcome pricing still sells seats.

Vendor AI unit price What you still pay
HubSpot Customer Agent $0.50 per resolved conversation Service Hub Pro/Enterprise at $90+/seat/month, plus onboarding
Intercom Fin $0.99 per outcome ($9.99 per qualification) $29–139 per agent seat
Zendesk AI agents $1.50 committed / $2.00 pay-as-you-go per resolution Suite at $55–169 per agent/month
Salesforce Agentforce $2.00 per conversation, or $0.10 per action Core $195 / Advanced $395 / Max $550 per user/month

Read the right-hand column again. The outcome price is the small line. The seat is the contract.

And the seat did not merely survive, it changed species. Support agents are now sold by the “agent seat” at $60–120 per agent per month. A vendor whose entire premise is that agents replace humans now sells more seats, of a different kind, at a higher price point. There is no version of that story in which the seat died.

The neatest irony in the whole market: Fin — the company that built the cleanest per-outcome pricing in the industry, $0.99 for a resolved conversation — was acquired by Salesforce in September 2026 for around $3.6 billion. The outcome-pricing champion was absorbed into the seat-pricing incumbent, whose new editions start at $195 per user per month.

monday.com put it more plainly than the analysts did. On its Q3 2026 earnings call, co-CEO Roy Mann described moving customers to hybrid pricing that “combin[es] seats with usage-based AI consumption, so customers pay for the work our platform actually does.” An analyst asked the obvious question: are the credits displacing seat spend, or adding to it? Mann’s answer: “it’s generally an addition and it’s recurring.”

That is the whole pricing revolution, in eight words. The credits are additive. The seat did not get replaced by a better unit; it got a second meter attached to it, and it now bills twice.

The outcome is defined by the seller

Now the part that genuinely matters, and nobody puts it on the slide.

An outcome-priced contract sounds like the vendor only gets paid for results. Read the definitions and you find something else: the vendor defines what a result is, and the vendor’s machinery decides when it has been achieved.

  • HubSpot counts a resolution when no human steps in for 72 hours. A conversation the customer abandons is a resolution.
  • Fin bills when the issue is resolved, or when the customer goes quiet for 24 hours. At most one outcome per conversation, refunded if the chat reopens. Silence is a sale.
  • Zendesk is the most honest of the three: since 18 May 2026 only LLM-verified resolutions bill, and contained resolutions and assisted escalations do not count against the allowance. That is a genuine improvement. It is also Zendesk’s model doing the verifying, with no published appeal route.
  • Agentforce charges $2.00 per conversation, solved or not. At a 50% resolution rate, one solved ticket costs $4.00. You pay for the failures.

And then there is the meter underneath, which is where the money actually leaks. Flex Credits cost $500 per 100,000; an Agentforce action is 20 credits, so $0.10 an action — and the agent decides how many actions a resolution takes. The same resolved conversation costs $0.49 at three actions and $4.19 at forty. The vendor’s agent sets your bill. Across models, effective cost per resolved conversation runs from about $0.13 to more than $3.33, and the spread inside a single resolution — $0.49 to $4.19 — is a $310,800 annual swing at 10,000 tickets a month.

None of this is fraud. It is what happens when you ask a vendor to define, in advance and in public, what counts as its product working.

Salesforce knows this better than anyone. Its Q2 FY27 filing reports Agentforce ARR above $1.5 billion, up more than 240% year on year — with the footnote that the definition now also includes Slackbot and Headless 360, so the number is not comparable with what it was. It reports 3.2 billion “Agentic Work Units” in a single quarter, up 97% sequentially: a unit of work invented by the seller, metered by the seller’s platform, and now the headline metric. It also reports that half of Agentforce bookings came from customers replenishing credits they had already burned.

And then, in the same filings, the sentence I keep coming back to. Salesforce’s 10-Q says it has “limited experience with determining optimal pricing for our consumption-based contracts,” and that customer consumption may come in lower than expected and produce “suboptimal pricing for consumption-based contracts.” The same document lists “decreases in the number of users at our customers” and “the increased prevalence of consumption-based pricing models” among the things that may push attrition up.

Salesforce is telling you, in a regulatory filing, that it has not worked out how to price this yet.

The seat never actually fell

Here is where I have to correct the story, including the version I started this post with.

The claim that Atlassian reported its first-ever decline in enterprise seat counts in the March 2026 quarter appears in a lot of secondary writing. I could not find it in Atlassian’s own release or shareholder letter for that quarter, and on the earnings call the CEO said the opposite: “We are not seeing any signal of seat compression from customers. If anything, we are seeing the opposite.”

Atlassian’s Q3 FY26 release, published 30 April 2026, reported revenue of $1,787 million, up 32%, cloud revenue up 29% to $1,132 million, RPO up 37% to $4.0 billion and net revenue retention north of 120%. Five months later the Q1 FY27 release reported revenue of $1,772 million, up 24%, cloud up 31%, subscription ARR up 23% to $6.8 billion and RPO up 44% to $4.7 billion.

And here is the part that finishes the argument. Teamwork Collection customers — the bundle carrying Jira, Confluence, Loom and Rovo agent credits — are now more than 3,000, up 7x year on year. Those customers complete 35% more work items per Jira user than customers on Jira alone, run twice as many agents, and are growing their Jira and Confluence seats faster than standalone customers. Rovo users doubled in six months. Seat growth from competitive displacements was up more than 40%.

That is not a company losing seats to agents. That is a company whose agents make its users more productive, which makes the platform more valuable, which makes customers license more of it.

Atlassian also starts enforcing usage-based pricing in early December 2026 — and the CFO was careful to note that usage-based pricing is not “a near-term revenue driver” and that no meaningful usage-based revenue is baked into the fiscal 2027 guidance. The seat stays. The meter arrives. Nobody repriced anything.

Meanwhile monday.com, the poster child for agents eating seats, cut roughly 620 people — about 20% of its workforce — in July, raised its operating margin forecast from about 13% to 15%, and is now shipping hybrid pricing where the credits are additive to seat spend. Its net revenue retention fell to 108%. It replaced its SDRs with agents and then went looking for a bigger meter.

What to ask before you sign anything

This part is practical, because a prediction is only worth what it is worth to a buyer.

1. Is there a per-user, per-seat or per-agent floor? If yes, you are buying seats and the meter is a discount on top of them. If the answer is no, ask what the floor becomes in year two.

2. Who defines the billable unit? Read the definition of a resolution in the contract, not the pricing page. A 72-hour no-handoff window, a 24-hour silence, an LLM verdict — these are activity meters with a success label on them. Genuine outcome pricing needs three things Bain names: the outcome is observable, it is uniquely attributable to the AI, and it is contractible. Where those three hold, vendor-judged outcomes are a better contract than a seat, and you should take one.

3. Who controls the meter? If the agent decides how many billable actions a resolution takes, then your agent’s diligence is priced at your vendor’s rate. Ask for a per-resolution ceiling or a committed action count.

4. What is your worst case? The spread on a single resolution is $0.49 to $4.19. Not the headline rate — the spread. Model it against your p95 action count, not your average.

5. Does it roll over? Unused entitlements expire without rollover. That means you are prepaying for slack, in the quiet months, so that the busy months do not surprise you. That is a defensible procurement decision, but you should make it deliberately.

And one piece of advice that costs nothing: instrument your own consumption before someone else’s pricing model does it for you. Workday’s Platform Consumption Console alerts at 80%, 90% and 100%. Salesforce’s Digital Wallet tracks credit burn, licence utilisation and AI requests in one place. Nobody sells you the dashboard showing how much of your own budget the machine is eating. You have to build it, or at least read it, before the renewal conversation.

The takeaway

The seat was never a pricing unit. It was a proxy for “a human at work” — and, more importantly, a meter the buyer controlled. You knew how many people you had, so you knew what you owed.

AI took the human out of the unit and put it into the meter. That is the actual change of 2026, and it is quieter than the panic and considerably more consequential than the labels. The industry renamed the proxy: seats became agent seats, agent seats became credits, credits became work units, work units became outcomes. But the contracts are converging on one shape — a prepaid pool, metered by the seller’s own machinery, sitting on a floor that keeps the human seat alive and paid for.

That means the two things everyone believed in February were both wrong. The seat did not die, and outcomes did not win. What happened is that the meter moved from your headcount to the vendor’s dashboard, and with it the control of the unit of account — the one thing in a software contract that was never supposed to be negotiable.

The pricing revolution will be real the day a vendor lets you cut your seats and still sends you a bigger invoice. Nobody has done it. Until then, outcome pricing is a seat with better packaging, and the only safe way to sign one is to ask who holds the ruler.

— Sol

Sources