Per-Seat SaaS vs AI Agent Pricing: Which Model Wins?

Which wins: per-seat SaaS pricing or AI agent pricing?

Per-seat pricing is losing ground, and no replacement has fully won yet. AI agents are replacing SaaS by breaking the pricing logic underneath it, not by shutting vendors down. The proof is already public: the industry's biggest per-seat vendor is running three different pricing models at once because none of them works alone anymore.

Is the seat already dying, even if vendors haven't said it out loud?

Start with Salesforce, the cleanest case study available. In about 18 months, Agentforce went from $2 per conversation at launch, to $0.10-per-action Flex Credits priced at $500 per 100,000 credits, to per-user licenses starting around $125 a month under its new Agentic Enterprise License Agreement. Salesforce didn't converge on one model. It kept all three running at once.

Marc Benioff said as much at Dreamforce 2025: "We had one pricing model a year ago... It's clear, one pricing model is not right for every company in the era of the agentic enterprise. So we've had to radically shift our pricing." That's not confidence talking. That's a vendor admitting its monetization model can't keep pace with how the product actually delivers value now.

Bain & Company reviewed more than 30 SaaS vendors and found the same pattern industry-wide: about 35% just raised per-seat prices to bundle in AI, roughly 65% bolted a usage meter on top of seats, and none had fully moved to usage- or outcome-based pricing. Nobody has solved this. Everybody is patching around it.

Why do agents break the per-seat logic?

A seat price only makes sense if a human is the unit of work. Per-seat licensing was never really about access. It was a proxy for how much value one person could extract from a tool in a month.

Agents break that proxy cleanly. When an agent resolves a support ticket or updates a CRM record without a human touching it, the seat count and the value delivered stop moving together. Peter Bendor-Samuel put it directly in Forbes: "the idea of a 'seat' starts to lose meaning, because the agents are effectively the users. Pricing based on the number of people logging in becomes arbitrary and increasingly disconnected from how value is actually created."

Salesforce is watching this happen inside its own customer base. Its sales engineers report seeing roughly a 10% reduction in seats and headcount across enterprise accounts, driven directly by AI making service agents more efficient. The vendor selling the seats is watching AI shrink demand for them in real time.

What pricing chaos did Dreamforce's agent-first pitch produce?

The $2-per-conversation to $0.10-per-action to $125-per-user sequence isn't three iterations toward a better answer. It's three answers stacked on top of each other because none of them is complete on its own.

• Conversation pricing punished high-volume, low-complexity use cases.
• Action-based Flex Credits made cost hard to forecast for anything conversational.
• Per-user licensing brought back the old seat problem it was supposed to solve.

The commercial traction, while real, is still small relative to Salesforce's core business. Agentforce ARR hit about $540 million in Q3 FY2026, up 330% year over year, across 18,500 deals, but only around 9,500 of those were paid. That's genuine momentum sitting on top of a company still doing roughly $41 billion a year the old way. The agent-first pitch is real. The pricing model to match it isn't finished.

What do Opus 5.5 and GPT-6 Sol actually change for builders?

The pricing scramble matters more this year than it would have two years ago, because the technical precondition for "just build it yourself" has been cleared. Cognition's FrontierCode 1.1 benchmark puts Opus 5.5 at the top, scoring 54.6% at roughly $0.80 per rollout versus $4 to $13 for competing models. On ARC-AGI-2, Opus 5.5 hits 93.3% and GPT-6 Astra hits 95.0%.

Those numbers matter less as leaderboard trivia than as a threshold crossing: frontier coding models are now capable and cheap enough that a non-specialist employee can prompt one into a working internal tool in an afternoon. That's the shift covered in Build vs. Buy in the Agent Era, and it's the same dynamic behind The End of the Large Software Team: one person with the right model now maintains what used to take a squad. It's also what makes a SaaS renewal look less like a foregone conclusion and more like a choice you're making by default.

Is the shadow AI economy the leading indicator?

If you want proof this is already happening outside procurement's view, look at MIT Project NANDA's 2025 study. Employees at more than 90% of companies use personal AI chatbot accounts for work, often without IT approval, while only 40% of companies have official LLM subscriptions in place. The study found this shadow AI economy often delivers better ROI than sanctioned enterprise AI initiatives.

That gap between 90% and 40% is the real story. Employees aren't waiting for a vendor to fix its pricing model or for IT to approve a rollout. They're building and using their own tools right now, which is exactly the dynamic behind Niche SaaS Is Bleeding Out as Clients Build It Themselves.

What does building it yourself really look like in the Klarna test case?

Here's where the piece needs to stay honest. Klarna's CEO said the company saved about $2 million after cutting ties with Salesforce, attributing it to AI-built data tools. That story spread as proof AI had replaced enterprise SaaS outright.

It's more complicated. Klarna later confirmed to CX Today that it replaced Salesforce and Workday with other SaaS products, not pure in-house AI tooling, using Deel for HR and a mix of alternative apps to cover CRM functionality. Some of the savings came from consolidating AI-built internal tools. A lot of it came from swapping one vendor for another.

The honest reading: build-your-own is real and it's saving real money, but it's rarely a full replacement in one motion. It's usually a mix of dropping the wrong vendor, picking a cheaper one, and owning the pieces that are cheap and fast to build yourself.

How much does per-seat SaaS actually cost today?

The status quo bill is what makes this urgent rather than theoretical. SaaS cost per employee rose from about $7,900 in 2023 to roughly $9,100 by the end of 2025, close to a 15% jump in two years, with SaaS inflation running nearly five times the general consumer inflation rate across G7 countries. One dollar of every eight a typical organization spends now goes to SaaS.

Meanwhile the waste inside that spend is enormous and well documented. Multiple analyses put unused or underutilized SaaS licenses at around 44% of total spend, with the average enterprise wasting tens of millions of dollars a year on licenses nobody is using. Per-seat pricing was never efficient. Agents just made the inefficiency impossible to ignore.

What pricing model replaces the per-seat license?

The vendors closest to AI-native products have already moved, and the results are hard to argue with.

• Intercom. Its Fin agent grew from $1 million to over $100 million in ARR on a $0.99-per-resolved-ticket model, and now handles more than 80% of Intercom's support volume.
• Sierra AI, co-founded by former Salesforce co-CEO Bret Taylor, crossed $150 million-plus ARR in about 21 months on pure outcome-based pricing.
• Zendesk shifted to $1.50 per automated resolution, or $2 pay-as-you-go.

These aren't ownership models. Customers are still renting, just against a unit that tracks value instead of headcount. But it's the market telling you seats were the wrong meter all along, which is the point Bain's data backs up too: outcome and usage pricing is the honest direction of travel, even if most vendors haven't committed to it yet.

Does building your own still need guardrails?

None of this is a free pass to let every team spin up its own agent and call it done. The same 90%-vs-40% gap that proves employees are building tools on their own is also a governance problem. An agent an employee built over a weekend to automate a workflow carries the same accountability risks classic shadow IT always had: no audit trail, no owner when it breaks, no plan for what happens when the employee who built it leaves.

Owning your software doesn't mean nobody is accountable for it. It means someone in your organization actually is, instead of a vendor's terms of service being the only accountability you have.

What build-vs-buy framework fits the agent-first era?

The practical question for any team facing a renewal or a new tool request isn't "can AI do this." Frontier models can write working software for most internal workflows now. The real questions are these:

1. Is this a systems-of-record workflow or a task workflow? If the tool needs to be the durable source of truth for customer or financial data across the whole company, the moat around buying a mature platform is still real. If it's a task, an internal dashboard, a report generator, an approval flow, build it.
2. Is the vendor pricing you for access or for outcomes? If you're still paying per seat for something an agent does autonomously, you're paying for a proxy that no longer matches the work. Push for outcome pricing or start building the alternative.
3. Who owns the agent once it exists? If nobody in IT or engineering can name the owner of a homegrown tool, you don't have ownership, you have unmanaged risk with better branding.
4. What does this actually cost you today, waste included? Run the real number against your SaaS stack, not the list price. At close to $9,100 per employee a year and 44% of licenses going unused, most companies are overpaying for access nobody uses.

The vendors are still figuring out how to charge for an agent-first world. That confusion is your opening. Teams building and owning software today are the ones setting the terms for the next renewal conversation instead of just signing whatever pricing model shows up this quarter.