From Aryan Vatsa | Product & Market Analysis
Usage-Based Pricing Killed Predictability. The Terms Buyers Are Winning Back
On this page
79% of enterprises in a February 2026 survey of 500 finance leaders reported an AI cost overrun in the previous 12 months. The size of the bills is not the interesting part. Usage-based pricing removed the thing procurement exists to produce, which is a number you can put in a budget and still defend nine months later. Buyers have stopped arguing about rates and started rewriting terms.
Key takeaways
- 79% of surveyed enterprises reported an AI cost overrun in the past 12 months. The sample was 500 finance leaders at organisations of 1,000 employees or more in the US and UK, fielded by Sapio Research in February 2026, with a margin of error of 4.4 points.
- Better cost tooling reports more overruns, not fewer. In that same survey the most mature FinOps organisations reported overruns at 89%, with a mean overshoot of 30.9%, against 69% and 16.1% at early-stage teams.
- Vendors have conceded visibility, not liability. OpenAI shipped workspace, group and user credit limits in June 2026 and Anthropic shipped organisation and user spend limits with alerts in July 2026. Both cap your consumption. Neither caps your price.
- Rollover is the term almost nobody holds in writing. A full-text search of SEC filings returns the phrase "roll over unused capacity" in 9 filings from 3 companies, and returns no 10-K at all containing "true-forward".
What the unpredictability actually costs
Start with the number that makes the rest of this post necessary. In a survey of 500 finance leaders fielded in February 2026, 79% reported an AI cost overrun in the previous 12 months. Every respondent already had AI spend, so this is not a survey of the curious.
The pattern shows up in unrelated samples. A WitnessAI survey of 300 executives, reported by CFO Dive in July 2026, found 68% saying at least some AI initiatives ran over budget, and 33% saying it happened mostly or always. Two different questions, two different panels, roughly the same answer.
The overrun is not a rounding error
An overrun you can absorb is a variance. An overrun of a third of the budget is a planning failure with a name attached to it. The DoiT survey put the mean overshoot at 30.9% among the most mature FinOps organisations and 16.1% among early-stage ones.
Read that split carefully, because it inverts the obvious conclusion. The teams with the best cost tooling report the biggest overruns. The most likely explanation is that they can see the whole bill, and everyone else is comparing a partial view against an optimistic plan.
The renewal cost that nobody budgets for
Unpredictable bills do not only hurt the finance team. In the same WitnessAI sample, 27% said unmanaged AI usage led to a delayed or cancelled initiative, and 30% said it caused a cost overrun outright.
That is the number vendors should care about most. A meter that produces surprises does not just annoy procurement, it removes the internal sponsor's appetite to expand. The commercial model then becomes the main risk to the account, which is a strange place for a growth strategy to end up.
Why the bill moves when your usage does not
Procurement is used to negotiating a price for a unit it understands. Usage-based AI pricing breaks that habit in three specific places, and each one has a contract answer.
Model tier migration happens without a purchase order
A team upgrades a workflow from a small model to a frontier model because quality improved. Nobody raised a request, nobody changed a contract, and the per-call cost multiplied. The finance visibility arrives with the invoice, four weeks later.
This is why the vendor control that matters most is a default model set at the role or organisation level, not a warning email. The economics behind those per-call differences are covered in the piece on what inference actually costs and who absorbs it.
Agents multiply calls without multiplying work
An agent that retries, reflects and calls tools can produce dozens of billed events for one user action. Your usage in tasks stayed flat. Your usage in billable units did not.
Token prices falling does not fix this, and the two trends are often confused. That confusion is the subject of the post on why cheaper tokens keep arriving with larger bills, and it is the strongest argument for pricing the outcome rather than the call.
The unit is a vendor-defined abstraction
Salesforce prices Agentforce in Flex Credits. Its May 2025 announcement set one action at 20 Flex Credits, or $0.10, with credits sold in packs of 100,000 for $500. The company's own pricing help page repeats those rates.
Here is the part buyers miss. A credit is not a token, a call or a minute. It is a unit the vendor defines, publishes in a dated rate card, and can redefine in the next one. Negotiating 15% off a unit that can be re-specified is not a discount, it is a delay.
The five concessions buyers are actually winning
Across enterprise renewals, the asks have converged on a short list. They are ordered here by how hard they are to win, easiest first.
| Term | What it does | Evidence it exists | What the vendor wants back |
|---|---|---|---|
| Alert thresholds in the order form | Names the percentages and the recipients, so notification is an obligation rather than a courtesy. | Shipped as product by Anthropic at 75% and 90% for admins, July 2026. | Nothing. This costs the vendor no revenue. |
| Spend cap with an approval path | Stops consumption at a stated figure unless someone named approves the overshoot in writing. | Shipped by OpenAI in June 2026 as workspace, group and user credit limits. | Usually nothing, sometimes a commitment floor. |
| Invoice-grade usage data | Gives you the unit counts by team and model, in a form finance can reconcile. | Cost API at OpenAI, Analytics API at Anthropic, Digital Wallet at Salesforce. | Nothing. Vendors want you self-serving this. |
| Rollover of unused commitment | Stops an overestimated commitment turning into a payment for nothing. | Language appears in 9 SEC filings from 3 companies in the entire full-text index. | Term length, usually a multi-year renewal. |
| Frozen unit definition and rate card | Fixes what one billed unit means for the term, not just its price. | No public evidence found. Rate cards are published as separate dated documents. | The most resistance of the five. |
The evidence column is deliberately uneven. The first three are documented vendor features you can verify today. The last two are contract terms, and contracts are not published, so their absence from the public record is not proof that nobody has won them.
My position, stated plainly: the fifth row is the one worth spending your negotiating capital on, and it is the one almost every buyer trades away first. A cap protects this year. A frozen unit definition protects the renewal, because it is the only term that survives a repricing of the meter.
What vendors have already put in the product
Something changed in mid 2026, and it changed at both large model vendors within three weeks.
OpenAI announced usage analytics and spend controls for ChatGPT Enterprise on 18 June 2026. Admins got a consolidated view of ChatGPT and Codex credit consumption by user, product and model. The release added default workspace limits, group limits, individual overrides, an approval flow for credit requests, and a Cost API for finance systems.
Anthropic followed on 2 July 2026 with organisation and user spend limits, model defaults by role, an Analytics API, and spend-threshold alerts to admins at 75% and 90% of the organisation limit. Users get their own notifications at 75% and 95%, with a request path built in.
Why monitoring arrived before money
Look at what these features have in common. They control your consumption, they cost the vendor no revenue, and they convert a future billing dispute into a self-service problem you already agreed to own.
That is not cynicism, it is a fair read of the incentives, and the features are genuinely useful. It does explain the shape of the concession curve. Vendors gave away the cheap half of predictability quickly and are holding the expensive half, which is rollover, overage pricing and the definition of the unit itself.
Rollover is where the talk stops and the paper starts
Rollover of unused commitment is the concession buyers describe as standard and the market treats as rare. There is a way to check that instead of asserting it.
The SEC runs a full-text search across filings from 2001 onward. Queried on 20 August 2026, the exact phrase "roll over unused capacity" returns 9 annual reports from 3 companies. Six are Snowflake 10-Ks, including the one filed on 20 March 2026 for the year ended 31 January 2026. The others are TTEC Holdings and New Relic.
The related term buyers use for the opposite protection fares worse. A full-text search for "true-forward" across all 10-K filings returns nothing at all.
What this search does and does not prove
It proves the language is close to absent from public company disclosure. It does not prove rollover is rare in signed contracts, because contracts are not filed and vendors describe the same idea in different words.
The useful inference is narrower and still worth having. If a vendor tells you rollover is standard, ask them to point at where they have ever said so in public. Most cannot, and that tells you the term is a concession being traded, not a policy being applied.
What each term costs you, and what it is worth
Every concession has a price, and the price is rarely money. Alerts and usage data cost nothing, so ask for them in writing and stop treating them as wins. Caps usually cost a commitment floor, which is a fair trade when your floor is genuinely your floor.
Rollover costs term length. Vendors give it in exchange for multi-year renewals, because rollover only hurts them if you leave. If you are already planning a three-year relationship, rollover is close to free and you should ask for it before you ask for a discount.
The frozen unit definition costs the most political capital, because it constrains the vendor's own roadmap for repricing. It is also the term that decides whether your negotiated rate means anything in 18 months.
You will find negotiation guides quoting precise savings for each of these, such as cutting effective overage cost by 20 to 35 percent. Those figures come from consultancies selling negotiation services, with no stated sample and no methodology. Use the tactics if they fit, and do not plan a budget around the numbers.
Where this argument is weakest
Three things about the case above deserve stating before you act on it.
The evidence base is surveys, and surveys have sponsors
DoiT sells cloud cost management. WitnessAI sells AI governance. Flexera, whose 2026 research reported that 59% of respondents saw wasted AI software spend increase and only 31% had accurate visibility into AI software, sells IT asset management.
Each finding is convenient for the company that paid for it. The DoiT work is the strongest of the three because it names an independent fieldwork agency, a sample, a geography and a margin of error. It is still a survey of perceptions rather than an audit of invoices, and nobody has published the latter.
An overrun is not automatically a failure
If usage rose because adoption rose and the work produced value, the budget was wrong rather than the bill. The FinOps Foundation's 2026 survey of 1,192 practitioners representing more than $83 billion of annual cloud spend found 98% now manage AI spend, up from 31% two years earlier.
That is a young discipline attached to a fast-moving cost line. Some share of these overruns is forecasting immaturity, not vendor behaviour, and a contract term cannot fix a bad forecast.
The third caveat is about caps themselves. A hard cap that halts an agent mid-run during a launch week is an outage you chose and paid for. Prefer staged alerts with a fast approval path, and reserve the hard stop for workloads where a pause is cheaper than a surprise. Anyone selling caps as pure upside has not been on the receiving end of one.
How to run the negotiation
Bring numbers, not positions. Four of them decide every ask above, and you can pull all four from data you already hold.
| The number | Where it comes from | What it wins you |
|---|---|---|
| Peak month against median month, last 12 months | Invoice history | Sets the cap band and exposes whether your spike is seasonal or structural. |
| Share of consumption from your top 10% of users | Vendor admin analytics or usage API | Decides whether per-user limits work or only an organisation cap will. |
| Percentage of last term's commitment actually consumed | Order form against invoices | The entire rollover argument, in one figure. |
| Billed units per completed outcome | Your own logs, not the vendor's | Tests the unit definition and prices any move to outcome billing. |
Sequence matters more than the asks. Put terms on the table before price, because a rate cut on a redefinable unit is worth less than it looks, and vendors concede terms more readily while the number is still open.
Ask for alerts and usage data first and get them in the order form rather than the onboarding call. Then trade term length for rollover. Then, only then, discuss the rate. If you are consolidating vendors at the same time, the sequencing question gets harder, and the trade-offs are worked through in the piece on rationalising an oversized application estate.
One more thing worth saying to any vendor that resists all five. The shift from seats to meters was sold to your side as fairness, and it is covered in the analysis of what seat compression is doing to software pricing. Fairness runs in both directions, and a model that only surprises the buyer is not the model that was pitched.
Frequently asked questions
What are the main problems with usage-based pricing?
The core problem is variance, not price. Bills move because of things the buyer never approved, such as a team switching to a larger model, an agent retrying calls, or a vendor redefining what one billed unit contains. In a February 2026 survey of 500 finance leaders, 79% reported an AI cost overrun in the previous 12 months, and the mean overshoot at the most mature teams was 30.9%.
How do you cap spend on a usage-based contract?
Use both mechanisms. In the product, set organisation, group and user limits, which OpenAI and Anthropic both shipped in 2026. In the contract, name the cap figure, the alert thresholds, who receives them, and the person whose written approval is required before consumption continues past the cap. A control that exists only in an admin console can be changed without your agreement.
Can you negotiate rollover of unused credits?
Yes, and it is usually traded for term length rather than money. Rollover only costs a vendor if you leave, so a multi-year commitment makes it cheap for them to grant. Do not accept the claim that it is standard practice. The phrase "roll over unused capacity" appears in only 9 filings from 3 companies across the entire SEC full-text index, so ask for the wording in your own order form.
What is a fair overage rate in a consumption contract?
The defensible position is that overage should be billed at your committed rate, not at list or on-demand pricing. There is no published benchmark for what buyers actually achieve here, because contracts are private and the figures circulating come from consultancies selling negotiation help. Treat those numbers as directional, ask for parity with your committed rate, and get the fallback rate written down.
Do spend caps stop a vendor raising prices?
No, and this is the most common misunderstanding. A spend cap limits how much you consume, so it protects a single billing period. It does nothing about the rate card, the definition of a billed unit, or the renewal price. If you want price protection you need a separate term that fixes both the rate and the unit definition for the length of the contract.
What data do you need before negotiating a usage-based renewal?
Four things. Peak month against median month over the last 12 months, from invoices. The share of consumption coming from your heaviest 10% of users, from vendor analytics. The percentage of last term's commitment you actually consumed, which is the rollover argument. Finally, billed units per completed outcome, from your own logs rather than the vendor's dashboard.
Where to start this quarter
Pick your largest consumption contract and pull 12 months of invoices this week. Calculate two figures only: peak month divided by median month, and the share of your commitment you actually consumed. Those two numbers tell you whether you need a cap, rollover, or both, and they take an afternoon to produce.
Then read the order form for the term that governs the unit, not the price. If the rate card is a separate dated document that the order form points at, your price is only as fixed as that document, and that is the sentence to raise at the renewal call.
Related analysis
If you are deciding whether to buy the meter at all, the comparison in build versus buy for coding agents works through the same economics from the other side, and the piece on where measurable AI return has shown up covers what to measure once you have signed.
References
- DoiT with Sapio Research, AI spending survey, February 2026. 500 finance leaders, US and UK, organisations of 1,000+ employees, margin of error 4.4 points. Used for the 79%, 89%, 69%, 30.9% and 16.1% figures.
- CFO Dive, Nearly 7 in 10 firms report AI cost overruns, 22 July 2026. Reporting a WitnessAI survey of 300 executives. Used for the 68%, 33%, 30% and 27% figures.
- Anthropic, New analytics and cost controls for Claude Enterprise, 2 July 2026. Used for spend limits and the 75% and 90% alert thresholds.
- OpenAI, New usage analytics and updated spend controls for enterprises, 18 June 2026. Used for workspace, group and individual credit limits and the Cost API.
- Salesforce, New flexible Agentforce pricing, 15 May 2025, and the Agentforce pricing help page. Used for 20 Flex Credits per action, $0.10 per action and $500 per 100,000 credits.
- US Securities and Exchange Commission, EDGAR full-text search, queried 20 August 2026, and Snowflake's 10-K filed 20 March 2026. Used for the 9 filings from 3 companies and the absence of "true-forward".
- Linux Foundation, State of FinOps 2026 survey results, 19 February 2026. 1,192 respondents, more than $83 billion of annual cloud spend. Used for the 98% and 31% figures.
- Flexera, When AI budgets balloon, 20 July 2026. Used for the 59% wasted-spend and 31% visibility figures.
Weakest thing about this source base: four of the eight sources are surveys published by companies that sell cost management, governance or asset management software. Only the SEC filings and the vendor announcements are primary documents, and no audited study of actual invoices exists in public.
Related reading