Seat pricing measured against AI agents

Why the billing unit breaks, not the price

The debate about software cost has run on the same axis for years: too expensive or acceptable. That is the wrong axis. What is happening in 2026 is not a price jump but a change of billing unit. For decades the unit of enterprise software was a person with a login. That worked because work and presence were the same thing.

An agent, meaning a program that completes a task on its own, breaks that equation. No first name, no employee number, no working hours. It calls an interface, completes a task and disappears again. It creates load but no seat. The better you automate, the more work runs through channels that do not appear in seat logic at all.

This leads to the actual point and it is aimed not at vendors but at procurement. If you sign a multi-year contract on the old unit this year, you freeze a cost base that contradicts your own automation. You keep paying for presence while the work moves to execution by agents. And for the whole term you have given up the leverage to shape the transition.

The point in one sentence

The problem is not the price per seat but the seat as a billing unit, because a contract that prices presence becomes wrong the moment work happens without presence.

How vendors are actually switching

The interesting observation is not that vendors are switching. It is how they do it. Four examples can be verified in published announcements and only one of them actually replaces the seat.

On 27 April 2026 GitHub announced that all Copilot plans would move to usage-based billing on 1 June 2026. Billing is by token, meaning the amount of text a model reads and writes, at the published prices per model. The decisive sentence, which many summaries leave out: seat prices stay unchanged, Business at 19 US dollars and Enterprise at 39 US dollars per user and month. Each plan includes a monthly allowance and beyond that you pay for consumption. GitHub’s reasoning is understandable. A short chat question and a multi-hour autonomous coding session used to cost the same.

Workday introduced Flex Credits in September 2025. The product page puts the principle like this: you are billed for the work the AI completes, not for the number of employees. A simple information retrieval costs one credit, an autonomously completed task five. There is a consumption overview with a warning when the balance runs low. The user licence continues alongside.

Salesforce took the step as early as 15 May 2025 with Flex Credits for Agentforce. A pack of 100,000 credits costs 500 US dollars and one agent action consumes 20 credits, so 10 cents. Here too it sits next to the user licence. Zendesk introduced outcome-based pricing for AI agents back in August 2024. You pay per request the agent resolves without human help. That is the purest example of billing by outcome in this group.

Vendor	New billing unit	Seat stays
GitHub Copilot	Tokens at published prices per model	Yes, plus a monthly allowance
Workday	Credits per agent task	Yes, credits sit alongside
Salesforce Agentforce	Credits per action, 20 credits per action	Yes, next to the user licence
Zendesk	Per automatically resolved request	Not for AI agents

Why this is called hybrid, not replacement

In three of four cases the seat stays and consumption is added on top. For procurement that is the most expensive variant, because it preserves the old base load and adds a new variable layer. If you think the seat model is on its way out and therefore negotiate only the per-user price at renewal, you negotiate exactly the half that is meant to stay constant anyway.

What the numbers support and what they do not

Many numbers circulate on this topic. Not all of them hold up the way their spread suggests. So here is only what can be checked at the source.

The Growth Unhinged survey is solid. In the 2026 edition, collected in April and May 2026 among more than 230 software companies, 37 percent use a hybrid model of base price plus consumption, up from 25 percent the year before. 29 percent already bill in AI credits or tokens and a further 33 percent plan to within six to twelve months. Among companies with more than 50 million US dollars in annual revenue, around half want to switch this year. Three out of four respondents changed their prices or packages in the last year.

What matters is what is measured: the share of surveyed companies with the respective pricing model, not a share of market revenue. The previous edition with 240 companies shows pure seat pricing falling from 21 to 15 percent. The two surveys are not identical and cannot be joined into a time series.

This article deliberately leaves out market forecasts on total volume. The circulating figures could not be checked against a primary source. If you justify a renewal with a forecast that does not survive inspection, you have built your negotiating position on sand. The documented direction is entirely sufficient.

What the switch costs on the buyer side

The cost of the switch does not show up as a price increase but as a loss of predictability. The Zylo SaaS Management Index 2026, based on more than 40 million managed licences and more than 75 billion US dollars in tracked spend, reports that 78 percent of IT leaders had unexpected costs from AI features or consumption pricing in the last year. 61 percent cancelled projects because of unplanned cost increases. That is the actual damage: not the invoice but the abandoned automation.

Spending on AI applications grew 108 percent year over year according to Zylo, 393 percent at very large companies. Growth rates like that land on a procurement function built for a unit that changes once a year.

The vendors’ margin expectations from the 2026 Growth Unhinged survey fit the picture. The median target margin for AI revenue is around 50 percent and only 12 percent of companies aim for the 80 percent and more that is usual for software. Vendors themselves are calculating with much thinner margins. If you think consumption pricing is just a more elegant form of price increase, you underestimate how variable the vendors’ own cost side has become.

Why this particular renewal sets the base for years

A three-year contract signed today will very likely outlive the switch at the vendor. If a third of vendors already bill in credits or tokens and another third plan to within twelve months, price stability is the exception in this market. All the more so because three out of four vendors changed their prices in the last year anyway.

That does not mean you should refuse the new model. It means the contract has to anticipate the switch so that it neither surprises you nor hits you twice.

How to negotiate independently of the billing unit

Seven clauses carry most of that load.

Clause	What it prevents	How you notice it is missing
Billing unit defined in the contract text, not by reference to a price page	Unilateral redefinition of what an action or a credit is	The contract points to a web page with a change reservation
Conversion rule between old and new unit at a fixed rate	Paid seats surviving a switch as permanent base load	Nothing says what happens to licences when the model changes
Price cap per billing unit, not per user	The cap running empty once billing is no longer per user	The cap is worded only per user and month
Right to inspect individual consumption events, not just monthly totals	Being able to check the invoice only against the vendor’s own reporting	Consumption is visible in the portal as a total only
Explicit approval before overage is charged	Automatic back-billing on load spikes and runaway jobs	The contract says overage is billed automatically
Termination right on unilateral change of the unit	A model switch binding you mid-term	A change reservation with no matching customer right
Data export in machine-readable form at contract end	The exit option existing on paper only	Export only as PDF, only on request or at extra cost

Two points deserve emphasis. The per-user price cap is the most common false comfort in existing contracts. It protects exactly the unit that is losing relevance and says nothing about the one gaining it. And the right to inspect individual events is worth only as much as your ability to read them. Without a count of your own it is a right to data you cannot use.

For term length one simple rule applies: the more a system is affected by your automation, the shorter the commitment. A system whose usage will be mostly generated by agents within twelve months does not belong on a three-year term, however attractive the discount looks.

Measure your consumption before the vendor does

Consumption-based contracts shift the burden of proof. In the seat world you could check the invoice against a list of names. In the consumption world you check it against events that arise in someone else’s systems. Without a count of your own, you negotiate the next renewal with the other side’s numbers only.

The countermeasure is unspectacular: a log of your own into which every agent action is written before it reaches the external system.

{
  "ts": "2026-05-19T09:14:22Z",
  "system": "crm",
  "vendor": "vendor-a",
  "actor_type": "agent",
  "actor_id": "svc-agent-lead-qualify",
  "unit": "task_completed",
  "unit_count": 1,
  "replaces_seat_role": "sales-rep",
  "outcome": "resolved",
  "vendor_billed_units": 5
}

Two fields make the difference. vendor_billed_units next to unit_count shows you the vendor’s conversion factor over time. If it changes without the price changing, that is a silent price increase. And replaces_seat_role makes visible which seats you no longer actually need.

SELECT vendor,
       unit,
       date_trunc('month', ts) AS month,
       sum(unit_count)          AS own_units,
       sum(vendor_billed_units) AS billed_units,
       round(sum(vendor_billed_units)::numeric
             / nullif(sum(unit_count), 0), 3) AS factor
FROM agent_usage_event
GROUP BY 1, 2, 3
ORDER BY month DESC, billed_units DESC;

These three columns are the hardest negotiating basis you can get, because they come from your own system. They also answer the question that is asked at every renewal meeting and rarely evidenced: how much of what we pay for is actually used. This is where monitoring stops being an operations topic and becomes a procurement instrument. Brought together in a cost report, it replaces the annual surprise with a running number.

What you build yourself and what you keep buying

The switch gives the build-or-buy question a sharper criterion. It is no longer what is cheaper but which unit multiplies when your automation works.

Building yourself pays off where the billing unit hangs directly on your automation success. Connections between systems, internal agent workflows, consumption counting, thresholds and approvals. Those are exactly the places where every efficiency gain at a consumption-based vendor turns straight into a higher invoice. You optimise and pay for it. You do not want that feedback loop outside your walls.

Buying keeps paying off where the unit stays untouched by your automation and the vendor carries risk you would otherwise have to carry yourself. Accounting with regulatory updates, identity management, certified industry solutions. Here the number of people remains the more honest unit and the seat is not a design flaw.

In between sits the category that ends up in the SaaS Graveyard: tools that cover a single step in a chain an agent is now working on. They get expensive first, because their usage is the first to grow through machines. And they are replaceable first, because their scope is narrow. This category does not belong in a multi-year contract but on the list for a custom solution.

What you can check before the next signature

The documented core is narrow and still sufficient for a decision. Vendors are mostly switching to hybrid models, meaning consumption on top of the seat. And buyers mostly report unexpected costs from exactly that switch. How hard it hits you depends on which of your systems come up for renewal in the next twelve months and how much of the work in them is already done by agents today.

Three things you can check without a market forecast:

Which systems come up for renewal in the next twelve months and is there a consumption count of your own for them? Without your own numbers, every negotiation is a hearing.
Is the billing unit in the contract or on a vendor price page? A discount on a unit that no longer applies in eighteen months is a concession without value.
Does the term length match the degree of automation? What agents touch gets a short commitment. What people touch may run long.

If you want to know what that count looks like in an existing system landscape and which of your contracts hang on the wrong unit first, talk to us.

Sources

Note

The prices come from the linked vendor announcements as of the respective date and may have changed since. The shares from the Growth Unhinged surveys refer to surveyed companies, not to market revenue. This article is not legal advice. For the assessment of specific contract clauses, seek qualified counsel.

Agents do not log in. Renew per seat in 2026 anyway and you freeze the wrong metric.

Key Takeaways