Why Enterprise AI Pricing Is Quietly Leaving the Token Behind

How Salesforce, SAP, and Workday are rewriting the AI pricing model around outcomes instead of usage — and why that shift is harder than it looks

Opinion. Edgewisely's view: the token-based pricing that built the generative AI market is already becoming a liability for the vendors who rely on it, and the winners of the next two years will be the ones who figure out how to charge for results instead of activity.

For three years, buying enterprise AI meant buying compute by the token. That era is ending, and most vendors haven't finished building what replaces it.

The shift shows up first in the numbers. A 1H 2026 buyer survey found 43% of enterprise buyers now prefer consumption-based pricing models, while 27% favor outcome-based structures — and vendors still offering only flat, seat-based pricing risk immediate disqualification from enterprise deals, according to Futurum Group's research on the topic. That's a remarkable reversal from the SaaS pricing conventions that dominated enterprise software for two decades.

What's actually changing

Salesforce has introduced outcome-based pricing for its customer-service AI, charging companies only when a bot resolves an issue without human intervention, per PYMNTS' reporting on the shift across ServiceNow, SAP, and Workday. Workday has moved toward a consumption model through what it calls "Flex Credits," letting customers pay for the specific AI agents they actually use rather than a flat per-seat license. SAP, for its part, has been more cautious about outright outcome-based pricing but has said openly on recent earnings calls that it wants to monetize what it calls "the value of our agents" rather than the seats accessing them, a framing ERP Today traced through SAP's own cloud-backlog and AI-pricing commentary.

None of these are cosmetic changes. A token-based or per-seat model charges for activity — how much compute got used, how many people have a login. An outcome-based model charges for a result: a resolved ticket, a completed workflow, a closed deal. That's a fundamentally different unit of value, and it requires vendors to define, measure, and stand behind what "success" means for a given AI agent, in a contract, with real financial consequences attached.

Why this is happening now, not two years ago

The proximate cause is that token-based pricing has stopped matching how enterprises actually experience AI cost. Enterprise AI spending has hit "a wall of financial scrutiny," with CFOs across industries replacing open-ended experimentation with demands for measurable returns, largely because unexpected costs have derailed roughly a quarter of AI projects, per reporting compiled by MarketScale on enterprise AI's 2026 ROI reckoning. The root cause, according to that reporting, is the accumulated opacity of usage-based pricing stacked on overlapping tool subscriptions and shadow AI adoption happening outside formal procurement — buyers simply can't predict their bill, and CFOs are done tolerating that.

Outcome-based pricing is, in one sense, a vendor concession to that frustration: it re-anchors the price to something a CFO can actually defend in a budget review. It's also, less charitably, a bet by vendors that agentic AI performs well enough, consistently enough, that they can afford to be paid only when it works — and that the economics still favor the vendor even after absorbing that risk.

The math nobody's fully solved

Analysts see this as more than a pricing footnote. Gartner forecasts agentic AI will account for 30% of enterprise application software revenue by 2035, a figure north of $450 billion, cited widely in coverage of the shift including Planetary Labour's guide to agentic AI pricing across Salesforce, ServiceNow, UiPath, and Workday. Capturing that revenue at scale requires vendors to solve a genuinely hard measurement problem: what counts as a "resolved" ticket when a human touches it halfway through, what happens when an agent's output is technically correct but the customer disputes it, and how a vendor prices an outcome that depends partly on the customer's own data quality, which the vendor doesn't control.

That measurement problem is why hybrid models — blending consumption, seat, and outcome components — are more common in practice than pure outcome-based pricing. Adobe, Salesforce, and ServiceNow have all leaned toward hybrids rather than an all-or-nothing bet on outcomes, a middle path that hedges the vendor's exposure while still moving the conversation away from raw token counts.

What this means for buyers and builders

For enterprise buyers, the practical advice is to treat any outcome-based AI pricing pitch with the same scrutiny you'd apply to a performance-based marketing contract: get the definition of "outcome" in writing, understand what happens at the edges (partial completions, disputed results, human handoffs), and model what the bill looks like at both low and high usage before signing. A pricing model that sounds cheaper at your current volume can become expensive fast once an agent is actually good enough that you route more work to it.

For vendors, the lesson is that pricing innovation is now a genuine competitive differentiator, not back-office housekeeping. A company that can credibly price and deliver against outcomes — with the measurement infrastructure to back it up — has a sales argument that pure token-based competitors simply can't match in a CFO-led buying process. That's a moat built on operational maturity, not model quality, and it's one a lot of AI vendors are not yet equipped to defend.

The takeaway

The token was never really the product; it was a convenient proxy for compute cost during a period when nobody, vendor or buyer, had a better way to measure what AI was actually worth. That excuse is running out. As agentic AI gets good enough to complete real workflows unsupervised, the pressure to price it like a result — not a resource — is only going to intensify, and the vendors still selling tokens in three years will be selling to whoever's left who hasn't figured out how to ask for better.

Buyers stopped asking what AI costs. They started asking what it's worth. Vendors who can't answer that in writing are going to lose the deal.

Frequently Asked Questions

What is outcome-based AI pricing?

Outcome-based AI pricing charges customers based on results achieved — such as a resolved customer service ticket or a completed workflow — rather than on usage metrics like tokens processed or number of user seats. Salesforce has implemented this for its customer-service AI, charging only when a bot resolves an issue without human involvement.

Why are enterprises pushing back against token-based AI pricing?

Token-based pricing has become unpredictable and opaque for buyers, contributing to unexpected costs that have derailed roughly a quarter of AI projects. CFOs are demanding pricing models that are easier to forecast and tie more directly to measurable business value.

Which enterprise software vendors have shifted their AI pricing models?

Salesforce has introduced outcome-based pricing for customer-service AI agents. Workday has moved to a consumption-based "Flex Credits" system. SAP has signaled it wants to price around agent value rather than seats. Adobe, Salesforce, and ServiceNow have adopted hybrid models blending consumption, seat, and outcome components.

How big could the outcome-based AI pricing market become?

Gartner forecasts agentic AI will account for 30% of enterprise application software revenue by 2035, a figure exceeding $450 billion, though capturing that revenue depends on vendors solving significant measurement challenges around what counts as a completed "outcome."

Editor's note — sources: This analysis draws on reporting and research from Futurum Group, PYMNTS, ERP Today, MarketScale, and Planetary Labour. For related coverage of the enterprise AI stack, see Edgewisely's guides to the best AI gateways in 2026 and Obsidian's bet on the agent leash.