The seat license was never really a price. It was a bet on how work happens. The bet said one human sits at one screen, opens one tab, and generates one unit of value per month. Vendors billed for the seat. Buyers counted the seats. For twenty years nobody had a better proxy for consumption, so the proxy became the product.
That bet has failed in three separate places in the last thirty days.
On July 1, 2026, Microsoft raised the base-seat price of the Microsoft 365 suite so that the true all-in cost of a Copilot user landed at $69 per seat per month on E3 and $90 on E5, before any agent activity at all. On June 1, GitHub moved Copilot's chat and agent features to a usage-based credit pool. On June 16, Microsoft launched Copilot Cowork, which requires a $30 M365 Copilot license just to access, then bills the actual work through separate Copilot Credits at $0.01 each, stacked on model use, context retrieval, tool calls, and runtime. On June 30, the enterprise volume discounts many CIOs had built their forecasts around expired. Same company, same product, three pricing model changes in one month.
Then Salesforce ran the same experiment on itself, in public. Agentforce now offers a Flex Credits model (a $500 minimum for 100,000 credits consumed as agents perform actions), a $5-per-user Agentforce User License that gates access to the platform while metering agent actions through Flex Credits, and the Agentic Enterprise License Agreement (AELA), a flat unlimited-use fee across Agentforce, Data 360, MuleSoft, and Slack for two or three-year terms. The customer picks the pricing model. Salesforce is happy to sell them all three, sometimes to the same customer, sometimes on the same account.
The pricing is shattering.
What the seat was really pricing
Per-seat SaaS priced two things at once and hid the second one behind the first. The first thing was access, a license to open the app. The second was a bet that the human logging in would consume roughly the same amount of the vendor's compute, storage, and support as every other human logging in. The bet held because the workload was human. A human reads at a bounded speed. A human hits enter a bounded number of times per day. A human takes lunch. The distribution had a bell curve, and the vendor could price to the median with predictable margins.
Agents shatter that distribution. One agent can run a query loop overnight that generates more calls to the CRM than a hundred humans would in a quarter. Another agent can idle for a week and then process a batch. The bell curve becomes a long tail, and the tail sits on the vendor's infrastructure bill. Priced per seat, the agent is either free from the vendor's perspective (ruinous) or throttled from the buyer's perspective (useless).
The vendors know this. The response was to keep the seat, add a meter on top, and call it a choice. Nobody at the top of a SaaS revenue chart admits that the seat has failed. Admitting it would reprice the whole book.
Replacing software seats with AI agents is a workload problem first
Every conversation I have with a CIO in September 2026 starts with the same misframing. They think replacing software seats with AI agents is a contract negotiation. They ask what the right price is. They ask which vendor's model is fairest. They ask if there is a benchmark to anchor to.
There is no benchmark yet, because the pricing sits on top of a workload architecture that most buyers have not designed. Until you know what the agent actually does per unit of business value, no vendor's meter matches your revenue. Salesforce's Flex Credits meter drafted emails, scored leads, and generated summaries. Microsoft's Copilot Credits meter model use, context retrieval, tool calls, and runtime, with a model picker (Opus 4.8 and Sonnet 4.6 today, a lower-cost Cowork 1 coming) as the cost lever. These two meters are not comparable, and neither of them is your business.
Your business has its own unit. It is the closed deal, the resolved ticket, the shipped order, the reconciled account. That unit is what an agent needs to be priced against, and that price is not on any vendor's rate card. It is derived from the vendor's meter times the frequency of the agent's calls times the buyer's tolerance for variance in the tail. Nobody is doing that math at the procurement table, because the person at the procurement table is still shopping seats.
The Klarna receipt
Klarna is the canonical cautionary tale of 2026, and every executive evaluating an AI workforce strategy is now expected to explain how their plan avoids the Klarna outcome. It is worth being precise about what happened.
In 2024, Klarna announced that AI had replaced roughly 700 customer service agents and that the AI performed at human-equivalent quality. The company cut headcount to about half its previous size. It ended its service provider relationships with Salesforce and Workday and moved to internally built systems. The CEO, Sebastian Siemiatkowski, gave interviews explaining that SaaS was over.
By early 2026 Klarna was quietly rehiring customer service staff. Satisfaction scores had degraded on the complex, ambiguous, emotionally loaded cases where the AI could not follow the thread. The company did not reverse the AI deployment. It reversed the assumption that agents alone would carry every case.
The lesson every executive is taking from Klarna is the wrong one. The real lesson is that Klarna designed the workload for a fully autonomous system before it had measured which parts of the workload actually were fully autonomous. The distribution of customer service tickets is bimodal, not uniform. Some tickets an agent resolves in one call. Some tickets require a human because the customer is upset, the case is novel, or the outcome has legal weight. Klarna priced its whole workforce as if the whole distribution was the first kind. When the tail showed up, the tail was uncovered.
This is the exact mistake the CIO is about to make with seats. The seat license priced the whole workforce as one kind of user. The agent contract, if signed the same way, will price the whole workload as one kind of task. Both are wrong, and both are wrong for the same reason. Uniformity is a pricing convenience. Workloads are not uniform, and the failures live in the shape you did not price for.
The pricing chaos is a tell
Gartner has an estimate that $234 billion in enterprise SaaS spending is at risk between now and 2030 from agentic arbitrage. Industry analysts pencil 20 to 35 percent of enterprise SaaS seats disappearing by the end of 2027. On February 3, 2026, the SaaS sector lost roughly $285 billion in market value in a single trading session on the recognition that per-seat had decoupled from consumption.
These numbers are directional rather than precise. Directionally, they explain why every incumbent vendor is running two or three pricing models in parallel. Salesforce is doing it. Microsoft is doing it. GitHub is doing it. The playbook has become simple. Hold the seat as long as possible. Meter the agent work separately. Hope the customer pays for both, because the customer does not yet know how to compare them.
The customer often pays for both. This is the actual failure mode of 2026. The SaaS-pocalypse is a headline. The SaaS-double-tax is what shows up in the invoice. Seat count does not go down, because agents run as users, because SSO logs the agent in with a named identity, because audit trails need a human accountable name, because the vendor's contract language does not distinguish an agent from a person. Copilot Cowork requires a $30 Copilot seat to exist and then bills the work on top. Agentforce runs against a licensed Salesforce user object and then bills through Flex Credits on top. Every "replacement" is an addition until the contract language says otherwise.
The August 24, 2026 Proskauer legal analysis on whether an AI agent needs its own software license is a live procurement issue right now, and it is being litigated inside every enterprise agreement being negotiated as I write this. Vendors want to preserve the seat and add the meter. Buyers want to eliminate the seat and pay only the meter. Nobody has the contract language for that yet, and the party without the language is the party paying more.
What the buyer actually needs to know
The strategic move is to know your own workload before you sign anything. Demanding better vendor pricing is downstream of that, and downstream negotiations trade percentage points on the wrong number.
Five things, in order.
You need to know which business outcomes are actually driven by which agent calls. Not tickets closed, not emails drafted, not summaries generated. The revenue-shaped outcome. The closed opportunity, the retained customer, the recovered payment. Agents run against a lot of surface work that does not compound into outcome. You do not want to pay per surface call. You want to pay per outcome, and to pay per outcome you have to know which calls produce it. That takes tracing, not a vendor demo.
You need to know the variance. An agent on a well-scoped workflow costs about the same every day. An agent on an ambiguous workflow costs ten times more some days than others. Vendors price on the median. Buyers experience the tail. If your workflow is high-variance, a credit meter will surprise you in the wrong direction, and you will overpay by budgeting for the average. The remedy is a two-week baseline run against real data, on a small deployment, before signing anything at scale.
You need to know which parts of the workload are agent-native and which parts are not. The Klarna reversal was not a failure of agents. It was a failure to segment. Some customer service tickets should never touch an agent, because the cost of the tail exceeds the savings on the head. Segment first, then automate. The segmentation is the work most executives skip, because segmentation reveals which of their existing SaaS spend was overpaying for humans doing work that should have been automated years ago, and which was underpaying for humans doing work that should never be.
You need to know your fallback. Every agent deployment I have seen this year in production has a fallback path to a human, a rules engine, or a queue. The fallback is where the real cost sits, because the fallback is where the failures land. If you have not costed the fallback path, you have not costed the agent. The Klarna satisfaction data cratered on the fallback that had been unstaffed.
You need to know the exit. Every one of these pricing models locks you in a different way. The seat locks you on account count. The credit pack locks you on prepayment. The Salesforce AELA locks you on multi-year flat fee. If your workload changes shape in eighteen months, which contract punishes you least? That question does not appear on any vendor's proposal. It should appear first on yours, along with the exact number of credits your fallback path burns when the primary agent misfires.
The vendors are not neutral parties
The pricing chaos is vendor discipline, not vendor confusion. Salesforce runs three pricing models at once because the three models let Salesforce measure which one the customer converges on and price the next round accordingly. Microsoft raised seat prices and split off agent metering in the same quarter because the two moves together mean nobody can hold their spend flat while adding capability. GitHub moved Copilot to credits because the credits mask cost inflation better than a seat price bump would.
The buyer's job is to run their own experiment against the vendor's. That means running the same workload against two vendors on two different pricing models for a fixed period, measuring both cost and outcome, and locking a contract only after the measurement is in hand. It is real work. It costs a real quarter of time. It is also the only way to sign a contract in September 2026 that is still a good contract in September 2027.
Companies that skip this end up in one of two failure modes. The first is Klarna. Over-automated, tail uncovered, quietly rehiring, headline reversed. The second is the double-tax. Seats plus meters, no savings, and a vendor invoice that has gone up rather than down. Both failures are avoidable, and both are entirely upstream of the vendor negotiation. Both are also invisible to a procurement process that measures seat cost and stops there.
There is a third failure mode that is quieter and worse. The buyer signs the AELA flat-fee agreement because the flat fee is legible and the two-year commit locks in a price. Two months later, the buyer's workload evaporates, because a supplier reorganized or a product line was killed or a market shifted. The flat fee is now a floor under spend that the workload no longer justifies. The credit meter would have breathed with the workload. The flat fee does not. Flat fees look like discipline. In an emergent workload, they are a hostage contract.
What architecting looks like
Buying a tool solves the tool problem. It does not solve the workload problem. Replacing software seats with AI agents, as a strategic move rather than a procurement line item, requires redesigning the workload so the agent runs against the value-shaped unit and not the interface-shaped one.
That redesign has a shape. It starts with a workload audit that traces every business outcome to the calls that produce it. It continues with a segmentation that separates agent-native tasks from human-native ones with an evidence trail per category. It builds the fallback path before the automation, not after. It picks a vendor and a pricing model on measured cost per outcome, not on vendor pitch. It sets a review cadence on the contract itself, because in this market a good contract at signature is a bad one in six months.
None of this is on any vendor's roadmap, because no vendor is incentivized to help you unbundle. All three of Salesforce's pricing models make more revenue if you do not know your own workload. Microsoft's three simultaneous meter changes make more revenue if you are still budgeting on seats. GitHub's credit shift makes more revenue if you have not measured how many credits your team consumes at steady state. The vendor's roadmap ends at the contract signature. Yours has to start there.
The architecture work has to come from somewhere. If it does not come from an internal team with the mandate, the calendar, and the analytical depth to do it right, it comes from an outside partner who has done it before. Believing that either option is optional is how the seat renewal from three years ago ends up ported straight into an agent contract that quietly costs 40 percent more for the same output. The seat contract is expiring, whether it gets renegotiated or not.
Conclusion
The seat license was a proxy for how work used to happen. It failed because agents do not log in the way humans log in and do not consume the way humans consume. Vendors have responded by proliferating pricing models to keep the seat alive while adding meters on top. Buyers who negotiate seats in this environment pay twice. Buyers who negotiate meters without workload architecture pay the tail. Buyers who sign flat-fee agreements without variance analysis lock in a floor their business will fall below.
Replacing software seats with AI agents is not a procurement project. It is a workload architecture project that ends in a contract. The order matters. Do the architecture, then sign the contract. Sign the contract first, and the contract designs the workload, and the workload no longer belongs to you.
Agor AI Advisory does this work with executives who need to run the audit, do the segmentation, cost the fallback, and enter the vendor conversation with numbers rather than intent. The window for that work is now, because every seat renewal from here forward will bake in assumptions that will be hard to unwind by the next cycle, and every one of those cycles compounds against the buyer that did not measure first.
Sources
- The Doomed Evolution of Salesforce's Agentforce Pricing, Monetizely, 2026
- Salesforce AELA: The End of Per-Seat AI Pricing, THE DAILY BRIEF, 2026
- $234B SaaS at Risk: Why Agents Bypass Your Stack, THE DAILY BRIEF, 2026
- Klarna Reverses AI Layoffs: Why Replacing 700 Failed, Digital Applied, 2026
- Does Your AI Agent Need Its Own Software License, Proskauer New Media Law Blog, August 24, 2026
- Microsoft 365 Copilot Plans and Pricing, Microsoft, 2026
- SaaS Pricing Beyond Per-Seat: Salesforce's New AI-Driven Model, Briefstak Newsletter, 2026
- Service-as-Software 2026: The End of Seat-Based SaaS, LinesNCircles, 2026