Agents Do Not Log In. Renew Per Seat in 2026 Anyway and You Freeze the Wrong Metric for Three Years.

Why the metric breaks, not the price

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

An agent breaks that equation. No first name, no employee number, no working hours. It calls an API, completes a task, disappears. It creates load but no seat. And the better you automate, the more work runs through channels that do not appear in seat logic at all.

Which produces the actual point, aimed not at vendors but at procurement: sign a multi-year contract on the old metric this year and you freeze a cost base that contradicts your own automation strategy. You keep paying for presence while your value creation moves into execution, and for that whole term you have no leverage to shape the transition.

The Core Claim

The problem is not the price per seat, it is the seat as a reference unit. A contract that prices presence becomes structurally wrong the moment work happens without it. This year's renewal decides which side of that break you spend the next years on.

What Gartner actually says

On 1 July 2026, Gartner put a number on the scale. By 2030, 234 billion dollars of enterprise application software spend is at risk from agentic arbitrage, roughly 20 percent of enterprise SaaS spending. At risk does not mean the spend disappears, but that its recipient and its basis of calculation can change.

George Brocklehurst, Managing Vice President at Gartner, names the mechanism precisely: “This breaks the link between user growth and revenue growth for many enterprise software vendors.” That coupling was the economic foundation of the subscription model and the reason vendors celebrated user adoption. When it breaks, a vendor loses not revenue first but its growth engine.

The consequence for the buyer sits in the same set of quotes, seen from the other side: “You are no longer buying software primarily for people; you are increasingly buying it for agents.” That is not a forecast, it is a procurement instruction. If the recipient is an agent, user count is the wrong question in the room.

Gartner names both sides. The shift is “an existential threat for vendors who are defending legacy dashboards and seat-based models” and at the same time a substantial revenue opportunity for others. A vendor defending its seat model is not defending its price, it is defending its way of counting.

One detail belongs in every vendor assessment: Gartner estimates roughly 70 percent of agents currently on the market are “agent-washing”. The metric shift comes with an offering that mostly does not deliver what the label promises. Buy on consumption and you may be buying consumption without result.

How vendors are actually switching

The interesting observation is not that vendors are switching, but how. Three patterns are evidenced in published announcements, and none is the clean replacement of the seat the debate assumes.

On 27 April 2026, GitHub announced that all Copilot plans would move to usage-based billing on 1 June 2026, billed on token consumption (input, output, cached) at published API rates per model. The sentence most summaries drop: seat prices remain unchanged. Consumption is layered on top, with a monthly credit allowance equal to the seat price, Business at 19 dollars per user per month with 19 in credits, Enterprise at 39 to 39. GitHub’s reasoning: the previous model is no longer sustainable, because a short chat question and a multi-hour autonomous coding session cost the same.

Workday takes the AI Flex Credits route: credits are consumed when agents work in production. The product page states the principle verbatim: “Flex Credits charge for the work AI completes on your behalf, not the number of employees in your organization.” Metering runs per agent skill, roughly one credit per information retrieval versus five per autonomously completed task, with a Platform Consumption Console and low-balance warning.

Salesforce moved as early as 15 May 2025 with its announcement of flexible Agentforce pricing: Flex Credits sit alongside the user license, a pack of 100,000 credits costs 500 dollars, and one Agentforce action consumes 20 credits, or 0.10 dollars. The earlier conversational pricing at 2 dollars per conversation remains in place next to it. Zendesk introduced outcome-based pricing for AI agents back on 28 August 2024, billing per automated resolution rather than per seat, the purest outcome pricing in this group.

Vendor	New billing unit	Seat remains	What you no longer control
GitHub Copilot	Tokens (input, output, cached) at published API rates per model	Yes, unchanged, plus credits equal to the seat price	How long an autonomous session runs
Workday	Credits per agent skill, 1 per information retrieval, 5 per autonomously completed task	Yes, Flex Credits sit alongside	How often an agent works in the background
Salesforce Agentforce	Flex Credits per action, 20 credits (0.10 USD) per action, 500 USD per 100,000 credits	Yes, alongside the user license	Which action costs how many credits
Zendesk	Per automated resolution	Not for AI agents	How many tickets the agent resolves

Why this is hybrid, not replacement

In three of four cases the seat stays and consumption is added on top. That is the most expensive variant for procurement: it preserves the old base load and adds a variable layer. Treat the seat model as legacy and negotiate only the per-user price, and you have negotiated exactly the half meant to stay constant.

What the numbers support, and what they do not

A note is required, because the figures circulating on this topic do not all hold up the way their distribution suggests.

The Gartner block is solid: 234 billion dollars, roughly 20 percent, by 2030. So is the Growth Unhinged survey. In the 2026 edition, fielded in April and May 2026 among more than 230 software companies, hybrid pricing sits at 37 percent versus 25 percent the year before. What matters is what is measured: the share of surveyed companies whose primary pricing model is that one, not a share of market revenue. The prior edition, with 240 companies, shows a move from 21 to 15 percent for seat-based pricing, 27 to 41 percent for hybrid, and 29 to 22 percent for flat fee. The two waves are not congruent and cannot be chained together.

What does not hold up is a figure that has made a career in the reporting: the forecast, attributed to Gartner, that at least 40 percent of enterprise SaaS spending will sit in usage-, agent-, or outcome-based models by 2030. That wording appears in a single secondary article and occurs neither in the Gartner press release of 1 July 2026 nor in four further syndications checked. Equally unsupported is the reading of the Growth Unhinged figures as a Gartner statement about vendor revenue share. And the frequently quoted claim that over 75 percent of AI vendors are uncertain about pricing agentic solutions could not be verified anywhere credible.

This is not a side issue. Justify a renewal with a forecast that fails inspection and you have built your position on sand. The evidenced direction is sufficient.

What the switch costs on the buyer side

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

The scale behind it: average SaaS spend of 55.7 million dollars, median 20.6 million, 305 applications per company on average. Spend on AI-native applications grew 108 percent year over year, 393 percent in large enterprises. Those rates land on a procurement function built for a unit that changes once a year.

Gartner worked the predictability loss through on a concrete case in a second press release on 24 June 2026, expecting AI coding costs to surpass the average developer salary by 2028, driven by rising token consumption. As the reason for poor predictability, Gartner explicitly names the lack of vendor transparency about how token consumption is calculated and billed. That is the clause missing from your contract, phrased as a market finding.

The vendor-side margin expectation fits, from the 2026 Growth Unhinged survey: the median target margin for AI revenue is around 50 percent, and only 12 percent aim for the 80 percent and above typical of SaaS. Vendors are modelling much thinner margins themselves. Treating consumption models as merely an elegant price increase underestimates how variable the vendors’ own cost side has become.

Why this particular renewal sets the base for years

Three developments coincide, and their overlap is the reason for the urgency.

First, the billing unit is moving right now. Growth Unhinged reports for 2026 that 29 percent of companies already bill in AI credits or tokens and 33 percent plan to within six to twelve months; above 50 million dollars in ARR, roughly half intend to switch this year. A three-year contract signed today will outlive that switch at the vendor, guaranteed.

Second, outcome-based pricing is small but moving in one direction: 5 percent of participants in the 2025 survey, with 25 percent expected by 2028. At the same time, 53 percent bundle AI features into the core product, and three out of four companies changed their pricing in the preceding year. Price stability is the exception here, not the rule.

Third, the shift has reached financial reporting. On 4 June 2026, Deloitte published a Technology Spotlight on accounting for outcome-based pricing in an agentic AI software product covering the ASC 606 questions it raises, in particular variable consideration and the timing of revenue recognition. When auditors write guidance, the transition is no longer a trend, it is a matter with a file number. And vendors rebuilding their revenue recognition rebuild their contract templates with it. Those templates are in front of you for signature.

How to negotiate metric-neutral

Metric-neutral does not mean refusing the new model. It means writing the contract so a change of billing unit neither surprises you nor hits you twice. 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, a resolution, or a credit is	The contract points to a web page with a change reservation
Conversion rule between old and new metric at a fixed rate	Paid seats surviving a model switch as permanent base load	Nothing states what happens to licenses 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 exclusively as “per user per month”
Audit right on raw consumption events, not just monthly totals	Checking the invoice only against the vendor’s dashboard	Consumption appears in the portal in aggregate only
Explicit approval before overage is charged	Automatic back-billing on load spikes and runaway jobs	Wording that overage will be billed automatically
Termination right on unilateral change of the metric	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 staying theoretical	Export only as PDF, 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 the metric losing relevance and says nothing about the one gaining it. And an audit right on raw events is worth only as much as your ability to read them; without your own count, it is a right to data you cannot use.

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

Measure your consumption before the vendor does

Consumption-based contracts shift the burden of proof. In the seat world you checked the invoice against a list of names. Now you check it against events originating in someone else’s systems. Keep no count of your own and your next renewal runs entirely on the other side’s numbers.

The countermeasure is unspectacular and buildable in a week: an event ledger of your own, written on every agent action 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 lets you watch the vendor’s conversion factor over time; if it moves while the price stays the same, that is a silent price increase. And replaces_seat_role shows which seats you no longer 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 asked at every renewal meeting and rarely evidenced: how much of what we pay for is used. This is where monitoring stops being an operations topic and becomes a procurement instrument, and a cost dashboard turns the annual surprise into a running number.

What you build yourself and what you keep buying

The build-or-buy question gets a sharper criterion from the metric shift. It is no longer what is cheaper, but which unit multiplies when your automation works.

Building yourself pays off wherever the billing unit hangs on your automation. Orchestration between systems, internal agent workflows, consumption metering, thresholds, approvals: exactly where every efficiency gain becomes a higher invoice at a consumption-based vendor. You optimize and pay for it. That feedback loop does not belong outside your walls.

Buying keeps paying off wherever the unit stays untouched by your automation and the vendor carries risk you would otherwise carry: accounting systems 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 covering a single step in a chain an agent now works on. They get expensive first, because machines grow their usage first, and they are replaceable first, because their scope is narrow. That category belongs not in a multi-year contract but on a list for building as a custom solution.

What you need to decide before the next signature

The evidenced core is narrow and still sufficient for a decision. Gartner puts 234 billion dollars and roughly 20 percent of enterprise SaaS spending at risk because the link between user count and revenue is breaking. Vendors respond predominantly with hybrids, consumption on top of the seat. And buyers largely report unexpected costs from exactly that transition.

Three decisions follow, and none needs a market forecast.

First, stand up your own consumption count for every system up for renewal in the next twelve months, before you negotiate. Without your own numbers, a negotiation is a hearing.

Second, negotiate the billing unit, not the discount. A point off a metric that will not govern in eighteen months is a concession without value.

Third, tie term length to 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 contracts hang on the wrong metric first, talk to us.

Sources

Note on sourcing: The Gartner press releases were blocked on retrieval; their content was cross-checked against four independent syndications. The pricing-model share figures come from the Growth Unhinged surveys and measure the share of surveyed companies whose primary pricing model is that one, not a share of market revenue; the 2025 and 2026 waves are not congruent and cannot be read as a time series. The Agentforce price points come from the Salesforce press release of 15 May 2025 and the Zendesk figures from Zendesk's own announcement of 28 August 2024; circulating claims of a Salesforce price per successful resolution could not be confirmed against the primary text and are not used here. For the Deloitte publication, title and date were verified, not the full text. The widely circulated claim that Gartner expects at least 40 percent of enterprise SaaS spending to sit in usage-, agent-, or outcome-based models by 2030 could not be confirmed against the primary text and is deliberately not used here. This article is not legal or tax advice; for the assessment of specific contract clauses, seek qualified counsel.

Key Takeaways