Salesforce helped make per-user SaaS subscriptions the default enterprise software buying motion for two decades. In 2026, Salesforce is selling Agentforce through a broader pricing stack: action-based Flex Credits at roughly $0.10 per standard agent action, user-based and enterprise agreement structures, and Agentforce Enterprise License Agreements designed to give larger buyers predictable, pre-committed access. Marc Benioff has framed the shift explicitly: one pricing model is not right for every company in the agentic enterprise. The company that helped establish per-user subscriptions as the industry default is now building the commercial architecture to move beyond them.
Salesforce is not alone. Uber reportedly exhausted its 2026 Claude Code budget by April. GitHub announced that every Copilot plan will shift to usage-based billing on June 1, with credits consumed based on token usage, while base plan pricing and code completions remain included. Anthropic more than doubled its estimated Claude Code token spend per active developer day, not as a pricing change, but because actual agentic usage patterns exceeded initial projections.
These are not isolated billing updates. They are the first visible fractures in a commercial model that was never designed for AI workloads. AI introduces a second unit of value (work performed) and a second unit of cost (inference consumed). The result is not the death of per-seat pricing. It is the rise of pricing stacks that layer access, usage, outcomes, and risk controls on top of each other.
Ten Salesforce licenses meant ten salespeople managing pipelines. A hundred Microsoft 365 seats meant a hundred people writing documents. The cost scaled with the headcount. The CFO could forecast it in a single spreadsheet column. That proportionality is what made per-seat the default.
AI breaks the proportionality in both directions. A single customer service agent can resolve thousands of interactions simultaneously, making per-seat overpriced for efficient deployments. A single developer running an agentic coding session can trigger millions of tokens in an afternoon, making flat-rate subscriptions ruinous for the vendor. That collision is what forced GitHub and Anthropic toward usage-sensitive pricing within months of each other.
Microsoft has reported 15 million paid M365 Copilot seats, roughly 3.3% of its 450 million commercial Microsoft 365 base, while saying there are "multiples more" users of free Copilot Chat. An enterprise paying $30 per seat for 10,000 licenses and seeing single-digit active usage rates is not paying $30 per productive user. It is paying several hundred dollars per productive user. Per-seat does not charge for value delivered. It charges for access granted.
Per-seat still buys something the alternatives do not: simplicity and budget certainty. But when adoption is uneven or when the workload does not scale with headcount, simplicity becomes an expensive illusion.
What replaces per-seat dominance is not a single successor. It is a pricing stack with four distinct commercial structures, each carrying a different risk profile.
You pay for what you consume, metered by actions, tokens, or compute. Salesforce Flex Credits, GitHub AI Credits, and every major inference API operate here. The advantage is auditability: cost tracks workload. The danger is the Jevons Paradox. Per-token costs have fallen by hundreds-fold in three years, but enterprise AI consumption has grown faster than costs have fallen. Organizations are spending more, not less. Budget controls, rate limiting, model tiering, and real-time spend visibility are not optional under consumption pricing. They are survival requirements.
Sierra, the most visible advocate, charges only when the AI achieves a predefined business result: a resolved support ticket, a saved cancellation, a completed upsell. No resolution, no charge. The model resonates with buyers who want vendor accountability. But the margin lives in one place: how "outcome" is defined. Buyers must specify whether abandonment, deflection, escalation, or partial resolution counts as an outcome. If the contract does not define those boundaries, the vendor will. The buyer's risk is not that outcomes are bad. It is that the definition and unit economics of an "outcome" become the contract. If resolution volume rises faster than avoided cost or incremental revenue, success can still create budget pressure. Volume caps, unit-economics tests, and shared-savings protections belong in every outcome-based deal.
Microsoft's new E7 Frontier Suite bundles Copilot, Entra, Agent 365, and M365 E5 into a $99 per user per month package. Microsoft's July 2026 packaging changes further embed Copilot Chat and related analytics into commercial suites. The per-seat anchor holds while the economics underneath shift to metered agents and studio credits. Enterprises get a floor they can forecast. The ceiling depends on how aggressively they deploy agents beyond the bundled allotment.
Less visible in the headlines but increasingly common in enterprise deals. Salesforce AELA agreements, Writer's enterprise subscriptions with unmetered elements, H2O.ai's annual capacity-based licensing, and private deployments sized by GPU allocation all sit here. The appeal is predictability without per-unit volatility. The risk is upfront commitment: you pay the fixed price regardless of utilization. When usage is high, this delivers the best unit economics in the stack. When usage is low, it becomes the most expensive form of shelfware.
The pricing model is not a billing detail. It is the single contract clause that determines whether your AI economics improve or erode at scale.
|
Pricing Model |
Best Fit |
Buyer Risk |
Vendor Risk |
|---|---|---|---|
|
Per-seat / hybrid |
Bounded productivity with predictable per-user usage |
Under-adoption: paying for idle seats |
Heavy-user margin erosion |
|
Consumption (tokens/credits) |
Agentic or developer-facing workloads with variable inference |
Budget volatility beyond one quarter |
Lower: the meter runs in vendor's favor |
|
Outcome-based |
High-volume customer-facing agents with clear success criteria |
Definition and volume risk: who decides "resolved," and at what unit cost |
Delivery failure: no outcome, no revenue |
|
Enterprise license / capacity |
Mission-critical deployments where budget certainty outweighs unit economics |
Upfront commitment regardless of utilization |
Usage overrun beyond modeled capacity |
Most enterprises running AI at scale will end up using more than one model across different vendor relationships. That is not a failure of strategy. It is the natural consequence of a technology whose cost profile varies by orders of magnitude depending on how it is used.
The wrong pricing model at the wrong scale is the most expensive AI decision most enterprises will make this year. Not because it shows up as a single line item. Because it compounds invisibly in every quarter that follows.
The 2026 Corporate Buyer's Guide to Enterprise Intelligence Applications evaluates 30-plus vendors across all four pricing models and maps them against the decision criteria enterprise buyers should use. It is built to help you see which commercial structure matches your workload, your risk tolerance, and your stage of AI maturity.
Sign up for launch day access at gaiinsights.com.