From Seats to Tokens: Why Enterprise AI Pricing Is Getting Harder to Budget

Enterprise software procurement has operated on a durable fiction for two decades: the per-seat subscription. A company pays a fixed annual fee per named user, the vendor grows by selling to more users, and the finance department books a predictable line item. For most enterprise software, the model worked because the software was, at root, a productivity tool for individual workers.

AI pricing per-token overview graphic

The shift to AI pricing per-token changes how finance teams forecast software spend, because usage now moves with every prompt. For the broader vendor picture, see our analysis of the enterprise AI software pricing war.

Generative AI is dismantling that fiction. An AI agent that handles 10,000 customer service conversations in a month is not a user. It has no seat. Its cost driver is compute — tokens processed, actions executed, API calls made — and those quantities bear no relationship to headcount. Enterprise software vendors face a structural choice: absorb variable AI compute costs inside a flat fee and compress margins as usage scales, or transfer that variability to enterprise customers as consumption-based pricing.

Most major vendors are choosing the latter.

The implications for CFOs, procurement teams, and technology leaders are significant. This is not a minor contract renegotiation. It is a change in who bears the financial risk of AI compute volatility — and the evidence so far suggests that enterprise buyers are absorbing that risk without fully understanding it.

Pricing pressure is also tied to infrastructure economics, and AI pricing per-token models depend on how providers fund capacity. Readers following the money can compare this with our piece on hyperscaler capex and debt financing.

How the Pricing Models Are Changing

Salesforce: Three Models, Simultaneously

When Salesforce launched Agentforce in October 2024, it priced autonomous AI agents at $2 per conversation. The logic was straightforward: customers pay for outcomes, not seats, yet underlying architectures still meter usage via AI pricing per-token consumption. The problem was definitional. What counts as a conversation when a single customer query triggers eight backend processes? Enterprise buyers could not model their costs, and procurement teams could not get purchase orders approved against open-ended variables.

By May 2025, Salesforce had introduced Flex Credits: $500 buys 100,000 credits; each standard AI action costs 20 credits, or roughly $0.10. Simultaneously, it added a $125/user/month per-seat add-on for employee-facing agents. Then, in December 2025, CEO Marc Benioff told analysts that seat-based pricing was becoming the norm for AI agents, pointing to its new Agentic Enterprise License Agreement as evidence of customers demanding predictability.

The result is that Salesforce currently sells Agentforce under three concurrent pricing architectures: per-conversation, per-action credits, and per-seat licences. Agentforce hit $540 million in annual recurring revenue by Q3 FY2026 at 330% year-on-year growth — but only approximately 8% of Salesforce’s 150,000-strong customer base has adopted it. The pricing complexity may be one reason adoption has not yet accelerated through the installed base.

Microsoft: Token Billing Inside a Subscription Stack

Microsoft 365 Copilot charges $30/user/month as of early 2026 as a flat add-on to existing E3/E5 licences. That flat model provides budget predictability, in sharp contrast to the volatility of consumption-based AI pricing per-token billing. What is less legible is the layer beneath it.

Microsoft Copilot Studio — used by enterprises building custom AI agents — charges $200/month for 25,000 credits, with pay-as-you-go at $0.01 per credit billed through Azure. Microsoft Fabric’s Copilot is metered per 1,000 tokens, with input and output tokens billed at different rates.

A single enterprise may now have AI costs flowing through its Microsoft 365 invoice, its Azure bill, and Copilot Studio credits — three separate metering systems, each with different units, requiring reconciliation to understand total AI spend.

ServiceNow: Consumption Layered onto Opaque Base Licences

ServiceNow does not publish pricing publicly — every plan requires direct negotiation. Its AI suite, Now Assist, has crossed $600 million in annual contract value, growing year-on-year at double its previous rate, through “Assist Packs”: a consumption add-on where AI summaries, response drafts, routing recommendations, and workflow automations each draw down a finite pool of assists. When that pool is exhausted, organisations purchase additional packs.

AI has become a recurring operational expense that can expand without a corresponding change in the base licence count.

The Risk Asymmetry

The common thread across these examples is where financial risk sits under variable AI pricing per-token models. Under per-seat pricing, the vendor shoulders demand risk: if customers use the software less than expected, the vendor still earns the same licence fee. Under consumption pricing, the customer shoulders demand risk: if usage scales faster than anticipated, the bill grows accordingly. The vendor’s margin floor is protected. The customer’s budget ceiling is not.

CIO.com analysis describes this explicitly: vendors are transferring the cost volatility of AI compute to customers while monetising customer-side productivity gains as margin, using pricing units — credits, interactions, events — that are often opaque enough to obscure the true value exchange.

A Salesforce Flex Credit, a Microsoft Copilot Studio credit, and a ServiceNow Assist are all billed against different definitions of what constitutes a unit of AI work; no cross-vendor benchmark exists to evaluate whether any of them is fairly priced.

Gartner’s assessment is starker: without proper cost modelling, organisations can make 500–1,000% errors in AI cost calculations. That is not a rounding error. It is the difference between a controlled pilot and a budget crisis. The pattern connects directly to broader findings on enterprise AI return on investment — the organisations reporting the weakest AI returns are often those that failed to model compute costs before scaling deployments.

What the Data Shows on Enterprise AI Spend

The survey evidence confirms the pattern. A West Monroe survey of 310 procurement, IT, and finance executives at organisations with at least $100 million in revenue found that nearly half saw licensing and subscription costs increase by more than 10% at renewal — above industry norms — and nearly two-thirds reported paying more for software assets than their peers. Only 3% believed their bills were lower.

A Metronome field report tracking 85% adoption of some usage-based pricing among SaaS vendors found that hybrid pricing — a subscription base plus consumption charges — surged from 27% to 41% of pricing models within a single year. Pure seat-based pricing fell from 21% to 15%.

The FinOps Foundation’s 2026 State of FinOps Report — based on 1,192 practitioners representing over $83 billion in annual cloud spend — found that AI spend management has moved from an emerging concern affecting 31% of FinOps teams in 2024 to being managed by 98% of teams by 2026. AI cost management ranked as the top new capability FinOps teams plan to build.

The infrastructure for governing consumption-based AI spend is being constructed at speed — which implies it was not there before. This shift is inseparable from the hyperscaler capital expenditure surge driving the underlying compute costs that vendors are now passing downstream.

Second-Order Effects

Budget processes are not designed for variable AI spend. Annual planning cycles work on fixed commitments, creating friction when managing unpredictable AI pricing per-token tiers.

A per-seat software bill renews at a known figure; a consumption-based AI bill can balloon mid-year if a new use case goes into production, a large deal requires intensive AI processing, or user adoption outpaces a conservative usage model.

McKinsey’s analysis of 17 global companies found that more than a third of high-performing organisations now direct more than 20% of digital budgets to AI — and recommends that CIOs make explicit decisions about which existing applications to retire to avoid AI consumption permanently inflating run costs.

The single-purchase-order model is breaking down. Enterprise procurement has relied on annual software purchase orders: one approval, one invoice, one budget line. Credits-based AI pricing is emerging as a partial solution — finance approves one large annual credit purchase, teams consume against it — but this creates its own problems: credits expire, rollover terms vary by vendor, and unexpectedly high usage in one business unit can drain credits intended for others.

Negotiating leverage is shifting. Historically, enterprise software buyers negotiated on seat count, discount tiers, and multi-year commitments.

Consumption pricing requires a different skill set: understanding usage patterns before signing, modelling cost scenarios at different adoption rates, negotiating spend caps, requesting credit rollover, and insisting on contractual transparency about what constitutes a billable unit. a16z’s enterprise AI survey found that AI procurement has matured to enterprise software standards — security checklists, benchmark scrutiny, more rigorous evaluations — but spend caps, credit rollover, and transparent billable-unit definitions remain an active negotiation frontier that buyers have not yet standardised.

A new vendor-customer dynamic. Vendors that shift to consumption pricing gain a structural revenue expansion mechanism: as enterprise AI usage grows, vendor revenue grows automatically, without requiring a separate renewal negotiation. The customer relationship becomes self-expanding from the vendor’s perspective. This is a meaningful change to the software contract as a governance instrument — one that procurement teams were not designed to manage.

The Counterargument: Consumption Pricing Can Be Fairer

There is a genuine case for consumption-based AI pricing per-token models that should not be dismissed. Per-seat licensing has its own distortions: companies over-provisioned licences routinely, paying for seats that went unused, while vendors extracted revenue unconnected to delivered value. Consumption pricing, in principle, aligns cost with benefit — organisations that use AI heavily pay more because they are presumably receiving more value.

The problem is not consumption pricing per se. The problem is the transition period: opaque unit definitions, multiple concurrent pricing models, immature enterprise tooling for tracking AI consumption, and budget processes built for a fixed-cost world. Organisations that build internal AI cost governance now — treating AI spend with the same FinOps discipline applied to cloud infrastructure — are better positioned to negotiate favourable terms and avoid budget surprises.

No large-scale public case studies of consumption pricing delivering the expected cost-value alignment have yet been published; the theoretical fairness argument remains ahead of the documented evidence.

SLA Contracting and Budget Risk Management for AI Pricing Per-Token Models

Managing enterprise technology spend under AI pricing per-token structures requires finance and engineering leaders to establish automated consumption guardrails. Unlike predictable per-seat SaaS licensing, the volatility inherent in AI pricing per-token agreements can trigger rapid budget overruns when recursive agentic workflows loop unpredictably. Enterprise procurement teams negotiating AI pricing per-token contracts must demand dynamic rate limiting and committed-use discounts.

Mitigating the cost volatility of AI pricing per-token models requires organizations to implement real-time cost telemetry across every application endpoint. By tagging inference requests with department IDs and setting hard spending caps, companies utilizing AI pricing per-token APIs maintain financial predictability. Furthermore, optimizing prompt caching reduces effective AI pricing per-token expenses by eliminating redundant token processing on repetitive context windows.

As enterprise adoption matures, hybrid contract models that combine baseline annual commitments with overflow AI pricing per-token tiers are becoming standard. Technology buyers who master AI pricing per-token unit economics will capture the agility of consumption billing while shielding corporate balance sheets from unbudgeted agentic compute spikes.

What to Watch

Outcome-based pricing is the next front. Rather than per-token or per-action charges, a small number of vendors are beginning to price on outcomes: per resolved support ticket, per completed transaction, per verified lead. a16z flagged this shift in late 2024. Outcome pricing transfers more risk back to the vendor — but also requires verification infrastructure that most enterprise deployments do not yet have.

Gartner’s 40% threshold: the firm projects that by 2030, at least 40% of enterprise SaaS spending will shift to usage-, agent-, or outcome-based models, according to its July 2025 report cited in Deloitte’s TMT Predictions 2026. Whether that share is reached sooner depends partly on how quickly FinOps-style AI governance spreads and how effectively enterprise buyers push back on opaque consumption units.

The Salesforce signal: Benioff’s December 2025 pivot toward per-seat framing for Agentforce — backed by a new enterprise licence agreement — is worth tracking. If customers at scale prove more willing to pay predictable per-seat rates for AI agents than variable consumption charges, other vendors will follow.

Benioff’s retreat from per-conversation pricing, driven by enterprise procurement resistance, is itself evidence that buyers who push back on opaque consumption models can shift vendor behaviour — suggesting enterprise buyers have more leverage to shape pricing norms than the current contract defaults imply.

This article was produced with AI assistance and reviewed by the editorial team.

Sarah Chen

Sarah Chen

Sarah Chen covers the economics of the technology industry for Next Waves Insight — venture funding, enterprise AI adoption, and the business models emerging around frontier technology. Her work traces capital flows and corporate strategy to explain not just what companies are building, but why, and what it signals about where the industry is heading. She has a background in technology and financial journalism.

5 thoughts on “From Seats to Tokens: Why Enterprise AI Pricing Is Getting Harder to Budget”

Leave a Comment