Seats to Credits: How AI Usage Pricing Changes the Cost of Switching
Two-thirds of B2B software companies in a June 2026 survey meter their AI features by usage, and more than half sell AI as credit or token bundles. That changes what you're comparing when you weigh a switch. A per-seat price is a number you can read off a quote. A credit price is a forecast, and the forecasting risk now sits with you. Here's how to bring it into the switching math.
Why the pricing model is now part of the switching decision
For most of the last fifteen years, comparing two SaaS tools on cost was mostly arithmetic: seats times price per seat times term, plus the one-off costs our hidden-costs framework covers. The AI layer is breaking that. Vendors increasingly price AI features by what gets consumed (actions, requests, tokens, conversations) instead of by who has a login. Often this sits alongside a seat price rather than replacing it.
That matters for switching in two directions. A credit-priced tool can look cheaper than a seat-priced incumbent on a quote and cost more in production once real usage lands. And the reverse: a team that pays for seats it barely uses may be overpaying compared with a metered alternative. Either way, the old comparison method of reading two price sheets side by side no longer gives a reliable answer on its own.
How widespread usage-based AI pricing is
The clearest recent data point comes from Nalpeiron's 2026 State of B2B Software Monetization survey, fielded in June 2026 among 255 B2B software companies' product and engineering leaders involved in pricing decisions. Asked how they price AI features (multiple answers allowed), 66% said usage-based metering, 52% said AI credit or token bundles, 49% said a per-seat plan with AI included, and 44% said outcome- or value-based pricing. Only 27% said they don't charge separately for AI at all. Most vendors combine several models, which is exactly what makes like-for-like comparison harder.
Analysts expect the shift to continue. CIO reported in June 2026 that Gartner predicts at least 40% of enterprise SaaS spending will shift to usage-, agent- or outcome-based pricing by 2030. A separate Gartner estimate, reported by CIO in July 2026, puts up to $234 billion of application software spending, about 20% of enterprise SaaS spend, at risk by 2030 as AI agents do work that used to require a person logging into each tool. Fewer human logins means less reason for vendors to keep charging per seat.
Three recent examples of the shift
None of these is a judgment on the products involved. They're useful because each one shows a different way usage pricing changes what a buyer has to model.
- Salesforce Agentforce (May 2025). Salesforce introduced Flex Credits, sold in packs of 100,000 credits for $500, with each agent action consuming 20 credits, or $0.10 per action. That sat alongside its existing $2-per-conversation model and per-user licenses, and a "Flex Agreement" let customers convert between licenses and credits. The lesson for buyers: the same product can now be bought three different ways, and which is cheapest depends entirely on your usage pattern.
- GitHub Copilot (June 2025). GitHub began enforcing monthly premium-request allowances on paid Copilot plans on June 18, 2025. As The Register reported, Business seats came with 300 premium requests per user per month and Enterprise seats with 1,000, with extra requests at $0.04 each, and more capable models consumed requests at a multiple (10x or more for some). The lesson: a seat price can carry a usage meter inside it, and which model your team picks changes the bill.
- Cursor (June–July 2025). The AI coding tool moved its $20/month Pro plan from a fixed number of fast requests to $20 of usage billed at API rates. Per TechCrunch, some users ran through their allowance quickly and some were charged overages they didn't expect. The company's CEO publicly apologized for the rollout and said it would refund unexpected charges. The lesson: a pricing change mid-subscription can change your cost base even when the list price stays the same.
What credit pricing does to the switching math
On the buyer side, the cost of that unpredictability is already showing up. Zylo's 2026 SaaS Management Index found that 78% of IT leaders surveyed had been hit with unexpected charges from consumption-based or AI pricing in the past year, and 61% had cut projects because of unplanned SaaS cost increases. As one strategy director put it to CIO, the shift moves forecasting risk from the vendor to the buyer. For a switching decision, that shows up in five places:
- Your baseline needs usage data, not seat counts. To compare a metered tool with your current one, you need to know how many actions, requests or conversations your team actually generates. Most teams don't track that. Without it, the comparison is a guess with a dollar sign on it.
- Prepaid credits can become sunk cost. Credit bundles are commonly sold upfront and often carry expiry or "use it or lose it" terms. Unused credits at the point you switch are money you don't get back, the usage-priced version of the contract-overlap cost our migration timeline guide covers. Check the expiry, rollover and refund terms before you buy a large bundle from a vendor you might leave.
- Model choice moves the price. Where different AI models burn credits at different rates, as in the Copilot example, the cost of a tool depends on configuration choices your team makes after signing. A pilot run on the cheapest model will understate production cost if your users then switch to a more capable one.
- Agent access is a new lock-in point. Gartner's George Brocklehurst told CIO that buyers should check whether an AI agent can do through a system's API everything a person can do on screen, and whether the vendor's terms allow it, because terms can restrict or price out third-party automation. He also called "who owns what the system learns from you?" a key contract question. If that learning accrues to the vendor's shared models, it is value you can't take with you when you leave, a newer form of vendor lock-in.
- Pricing can change under you. A vendor that reprices usage mid-term changes your total cost of ownership without changing the sticker price. A switch you decided against last year on cost grounds may be worth re-running.
A checklist before you sign or switch
The following steps make usage-priced AI tools comparable with seat-priced ones. They come from commonly circulated FinOps and procurement guidance, so treat them as practitioner recommendations, not audited research:
- Measure before you compare. Pull at least one to three months of real usage from your current tool (or a pilot) and price both options against the same usage profile, not against vendor-supplied "typical customer" estimates.
- Model three scenarios. Expected usage, 2x expected, and a spike month. A tool that wins only in the low scenario isn't really cheaper.
- Ask for hard spend caps and alerts. Know whether overage is billed automatically, blocked, or throttled, and who on your side gets notified. In the Cursor example above, some of the unexpected charges hit users who hadn't set a spend limit and didn't realize overage applied past the included allowance.
- Get credit terms in writing. Expiry, rollover, conversion between credits and licenses, and what happens to unused credits on termination.
- Lock the rate card. Ask for per-unit prices (and model multipliers, where they apply) to be fixed for the contract term, with notice requirements for changes. See our list of SaaS contract red flags for related clauses.
- Pilot on the configuration you'll actually run. Same models, same workflows, same number of users. Our guide to trials vs. pilots vs. sandboxes covers how to structure that.
If the tool you're evaluating sits on top of a third-party AI model, pricing is only half of the AI-specific switching cost. Prompts, fine-tunes and embeddings don't move cleanly between providers either; our breakdown of what breaks when switching AI vendors covers that layer.
Bottom line
Usage-based AI pricing isn't inherently better or worse for buyers than seats. It can be cheaper for light or uneven usage and more expensive for heavy, growing usage. What it changes is the work required to know which. A switching decision that compares seat prices alone now leaves out the line item most likely to surprise you. Collect usage data first, model a range rather than a point estimate, and treat credit expiry and overage terms as part of the switching cost, not fine print.
Open the switching cost calculator
This is a practical framework, not procurement, legal, or financial advice. Vendor prices, allowances and multipliers cited here are as reported at the time of each announcement and change frequently; check current vendor pricing pages before making a decision (last re-checked October 2026).