AI Billing Units Are Now a Sales Blocker, Not a Finance Complaint

Every AI vendor invented its own billing unit, and none of them convert. A Copilot Credit, a Flex Credit, an outcome and an automated resolution are four different objects carrying four different prices. AI billing has quietly become the slowest part of an AI purchase. 16 to 20 weeks is now the most common buying cycle, against 7 to 10 weeks for ordinary software.

What a credit actually is, and why no two vendors agree

A controller can audit a seat. It maps to a person, a payroll record and a leaver process. A credit maps to nothing outside the vendor that issued it.

That is the whole story, and everything below is the consequence.

A credit is a private currency

Underneath almost every AI product sits a token meter run by a model provider. Tokens are at least comparable. A thousand input tokens is the same object at every lab, priced differently.

Vendors do not resell tokens. They wrap tokens, orchestration, retrieval, retries and margin into a unit they name themselves, then price that unit. The wrapping is legitimate. The naming is where the trouble starts, because each vendor picks a different noun and a different denominator.

Scott Grossman, CFO of Ensono, told diginomica in August 2026 that there is "no perfect way to forecast AI spend, because frankly, we're just learning what goes into the cost of a token". His team has had to build its own translation layer as vendors shift reporting from tokens to credits mid-contract.

Nothing converts between them

You cannot express a Copilot Credit in Flex Credits. You cannot express either in resolutions. There is no exchange rate, no clearing house, and no vendor with an incentive to publish one.

That absence does real work in a negotiation. When every account team quotes in its own currency, every account team gets to frame its currency as cheap. A buyer comparing $0.008 against $2.00 is not comparing prices at all. They are comparing a rounding error against a whole customer interaction.

Five vendors, five units, one consolidated bill

Here is the same category, priced five ways. Every rate below comes from the vendor's own documentation, read in August 2026.

Rates are list prices from vendor pricing pages and documentation. They are not comparable to each other, which is the point of the table rather than a flaw in it. Contracted rates, volume tiers and overage multipliers are negotiated and mostly unpublished.

The same question can cost 1 credit or 100

Cross-vendor confusion is the visible half. The harder half sits inside one vendor's own rate card.

The 1-to-100 spread inside one product

Microsoft publishes its Copilot Studio rates in full, which makes it the fairest available example rather than the worst offender. A classic answer costs 1 credit. A generative answer costs 2. An agent action costs 5. Tenant graph grounding costs 10. A premium text and generative AI tool costs 100 credits per 10 responses.

A buyer sizing a deployment is asked to forecast a number that depends on which of those paths an agent takes. That path is set by how the agent was configured, which knowledge sources it reaches for, and whether the orchestrator decided a step was needed. None of it is visible from the outside at signature.

Microsoft's own worked example makes the compounding clear. A support agent averaging four classic answers and two generative answers across 900 daily conversations runs at 7,200 credits a day. Add tenant graph grounding to the same agent and the per-response cost moves from 2 credits to 12.

Reasoning models bill on two meters at once

When an agent calls a reasoning-capable model, Copilot Studio bills the feature rate plus a premium tool rate of 10 credits per 1,000 tokens. Two meters, one response.

This is the detail that breaks a spreadsheet. A model upgrade that a platform team treats as a quality improvement arrives in finance as a second billing line nobody forecast. The analysis of what inference costs do to AI margins covers the supply side of the same pressure.

Outcome pricing moved the argument, it did not settle it

The obvious fix is to stop billing effort and start billing results. Several vendors have done exactly that, and it is a genuine improvement.

It also created a second definitional problem on top of the first.

Three definitions of one word

Zendesk bills an automated resolution when an AI agent handles an interaction with no human involvement and the customer does not reopen the ticket within a 72 hour quiet period. Intercom bills an outcome when the customer confirms resolution, when they stop asking after a reply, or when a workflow completes. Intercom states that handoffs count.

Those two definitions differ on the case that matters most commercially. Consider an interaction the agent partly handles before passing it to a human. One vendor bills it. The other does not. Both call the result a resolution.

My position, stated plainly, is that outcome pricing is better than session pricing and should not be sold as certainty. It converts a quality risk into a volume risk. Finance can budget for volume more easily than for model behaviour, so this is progress. It is not the predictability the marketing implies.

What this actually does to a forecast

Seat maths is arithmetic. Headcount times price, adjusted for a hiring plan you already own.

Consumption maths is a distribution. You need a volume forecast, a units-per-interaction estimate, an overage rate and a view on how the vendor's own product changes before renewal. Three of those four are outside the buyer's control.

The FinOps discipline has felt this arrive at speed. 98% of practitioners now manage some form of AI spend, against 31% two years earlier, and the survey names visibility, allocation and value measurement as the three main obstacles. The most requested capability is granular monitoring of tokens, model requests and GPU utilisation. That is a polite way of saying the bills currently arrive without a breakdown anyone can use.

Two structural details make the forecast worse than the headline rate suggests. Prepaid pools usually expire, so unused capacity is not a buffer. And overage enforcement is a cliff rather than a slope. Microsoft disables custom agents at 125% of prepaid capacity, with end users told the agent has reached its usage limit.

That second detail is the one buyers underprice. A budget overrun in seat software is an invoice conversation. A budget overrun in agent software is an outage. The survey of agent orchestration tooling shows how many meters a single workflow can now touch at once.

The evidence that opacity is now blocking sales

This is the part vendors should read twice. Confusing pricing is no longer a post-sale support cost. It is a pre-sale conversion cost.

Levelpath surveyed 300 US procurement and supply chain leaders at organisations actively buying AI software in June 2026. AI ranked as a top buying priority and as the slowest thing they buy. The most common AI purchase cycle was 16 to 20 weeks, against 7 to 10 weeks for standard software over $10,000.

The stakeholder count moved with it. 58% of AI purchases involve seven or more people, against 43% for standard software, and 28% involve eleven or more. Every additional approver is another person who has to be taught what a credit is.

Finance is in the room, and a token budget makes a veto more likely

G2 surveyed 1,038 B2B decision-makers in June 2026. Finance involvement in software buying rose from 31% to 46% in a single year, while information security involvement fell from 32% to 25%. The gate moved from whether the product is safe to whether the bill is survivable.

Then the counterintuitive finding. Among organisations that had set up a dedicated token or LLM budget, 54% reported a CFO blocking an approved purchase. Among those without one, the figure was 29%.

Read that carefully before drawing the obvious conclusion. Measuring token spend does not cause vetoes. It reveals the variance that was always there, in organisations already spending enough to need a budget line. The lesson for a vendor is not that transparency is dangerous. It is that the moment a buyer can see the variance, an unexplained rate card stops being tolerable.

The consequences are already priced in elsewhere in the survey data. 57% of procurement leaders reported spend problems with an AI vendor, 35% received higher-than-budgeted bills, and 9% terminated a vendor over price increases. Only 16% said they felt very or extremely confident managing AI costs.

Where this argument is weakest

Three objections, and the first is stronger than I would like.

The case for bespoke units is real

An AI vendor genuinely does have variable cost of goods. Seat pricing forces the vendor to absorb usage risk it cannot control, which either raises the floor price for light users or bankrupts the vendor on heavy ones. A credit is an honest attempt to pass through a cost that actually varies.

Buyers have noticed. G2 found that preference for outcome or usage-based pricing doubled from 11% to 23% in one year. Nearly half, 49%, had already been offered variable-cost alternatives. 52% said variable pricing improved their view of the vendor. The market is not asking for seats back. It is asking for units it can model.

So the complaint in this post is narrower than "consumption pricing is bad". It is that vendors invented incompatible units and then declined to publish the distribution of how many get consumed.

The second objection is about attribution. Not all of the extra ten weeks is pricing. Security reviews outrank contract negotiation as a reported cause of delay, at 58% against 52%, and data residency questions are genuinely new work. Pricing complexity is a substantial share of the gap, not the whole gap.

The third is about the evidence itself. Both buyer-side surveys here are vendor-sponsored research from companies that sell into this problem. Levelpath sells procurement software and G2 sells to software marketers. Sample sizes and fielding dates are disclosed, which is more than most, and the findings point the same way as the FinOps Foundation's independent survey. Treat the exact percentages as directional and the direction as well supported.

What a vendor should publish tomorrow

If you sell an AI product, the fastest available win is not a price cut. It is a denominator.

Publish the full list of billable events and their rates, as Microsoft does. Then publish the thing almost nobody does: the observed distribution of units consumed per interaction across live deployments, with sample size and time window. Not the average. The average is the least useful number in a skewed distribution, and consumption distributions are always skewed.

A buyer who can see that the tenth percentile fires 3 actions and the ninetieth fires 40 can build a scenario. A buyer given a single average builds a worst case instead, and then asks their CFO to approve it. That is how a deal reaches week 16.

Three more things cost nothing and remove weeks. Say whether unused units expire. Say what happens at the overage threshold, in operational terms rather than commercial ones. And offer a spend cap that throttles rather than disables, because the current alternative asks a buyer to choose between an unbounded bill and an unplanned outage.

The vendors that do this will win deals against better products. That is the uncomfortable implication of finance sitting in 46% of buying decisions. The same dynamic runs through the piece on seat compression rewriting SaaS pricing and the assessment of Salesforce in the agent era.

Frequently asked questions

How to price-proof the next contract

Two moves, both available before your next renewal call.

Build your own translation layer, because no vendor will build it for you. Take your three largest AI line items and restate each as cost per completed business event: cost per resolved ticket, per drafted document, per qualified lead. That number is comparable across vendors even though their units are not. It is the only column that belongs in a board pack, and it is the discipline applied in the analysis of where measurable AI return has shown up.

Then send one email to each vendor asking for the consumption distribution behind their quote, with sample size and period. The answer, or the absence of one, tells you more about your renewal risk than the rate card does. A vendor that has measured its own product can produce it in a day.

Related reading