AI credits explained: what they are, how they work, and what one actually costs

Short answer: an AI credit is a prepaid unit of consumption that a vendor sells you instead of selling you the work directly. You buy a pool of credits, and every action the AI takes draws some number of them down. The catch is that two numbers decide your bill, and most vendors publish only one of them: what a credit costs, and how many credits an action consumes. Microsoft publishes both. Salesforce publishes both. Glean publishes the consumption and not the price. Clay splits your plan across two separate credit meters that behave differently. Until you have both numbers for your vendor, you do not have a forecast, you have a hope.

Last updated August 2026. Written for US businesses evaluating credit-metered AI tools. Every vendor figure below was read on that vendor's own pricing page, documentation or terms document, checked on August 20, 2026. Credit rates change often, so confirm before you sign.

What are AI credits?

AI credits are a synthetic currency. Rather than charging you per user, per hour or per outcome, the vendor invents a unit, sells you a pool of those units, and then charges a certain number of them for each thing the software does. Asking a simple question might cost one credit. Running a research task across six systems might cost a hundred and forty.

Vendors like credits for three reasons, and it is worth understanding all three because they explain most of what follows. First, credits let one price cover wildly different costs, since a cheap model call and an expensive multi-step agent run can both be billed in the same unit. Second, credits are prepaid, so the vendor collects the money before delivering the work. Third, and least discussed, a credit breaks the link between what you spend and what you understand you are spending. Nobody has intuition for what 3,750 credits means. Everybody has intuition for what $30 means.

How do AI credits work?

Three mechanics do all the work, and they are the three things to pin down for any vendor you are evaluating.

The conversion rate is what one credit costs in dollars. Microsoft sells 25,000 Copilot Credits for $200 a month, which is $0.008 per credit. Salesforce sells Flex Credits at $500 per 100,000, which is $0.005 each. Some vendors never publish this at all.

The consumption rate is how many credits each action costs. A Copilot Studio agent action is 5 credits, so four cents. A standard Agentforce action is 20 Flex Credits, so ten cents. Same category of action, two and a half times the price, and you only see that by converting both into money.

The allowance and the overage rule decide what happens at the edges. Some included volume usually comes with the subscription, and then behavior diverges sharply. Salesforce states plainly that there is no overage penalty and bills in arrears. Microsoft disables custom agents once your entitlement is exhausted, and your users see the message "This agent is currently unavailable. It has reached its usage limit."

What one credit actually costs at each vendor

This table converts four vendors into the same units, which is the single most useful thing you can do before a procurement conversation. It is also the thing no vendor will do for you.

Two things jump out of that table. The first is that Glean is the odd one: it publishes an unusually precise consumption rate card and has never published what a FlexCredit costs, so you can calculate your usage exactly and still not know your bill. We took that apart in full on our Glean pricing breakdown. The second is the naming collision. Salesforce sells Flex Credits and Glean sells FlexCredits, and they are unrelated currencies in unrelated products. If both vendors are in your evaluation, label them carefully in the spreadsheet.

What are Microsoft AI credits?

Microsoft's unit is the Copilot Credit, and it replaced the old "messages" currency on September 1, 2025. You buy capacity packs of 25,000 credits at $200 a month, or use pay-as-you-go against a linked Azure subscription, or prepurchase a one-year pool that saves up to 20 percent. Microsoft's own FAQ states there are no feature differences between the payment routes, only differences in timing.

The rate that catches people out is the prompt meter, which bills per thousand tokens at three tiers: basic 0.1 credits, standard 1.5, premium 10. Converted, that is $0.80, $12 and $80 per million tokens, and reasoning models bill on two meters at once. The full licensing picture, including the four separate routes to a Copilot Studio license, is on our Copilot Studio pricing page.

What are AI credits in Google?

Google is the exception in this group, because Gemini Enterprise does not use a credit currency at all. It charges per seat, from $21 a month for Business and from $30 for Standard, and bills Agent Platform consumption separately on standard Google Cloud rates underneath. There is no pool to buy down and no credit balance to watch.

That sounds simpler and in one way it is, but it moves the unpredictability rather than removing it. With credits you cannot forecast the meter; with seats you cannot forecast how many people will actually use the thing you bought for everyone. The edition and seat-cap detail sits on our Gemini Enterprise pricing page.

What are monthly AI credits?

Monthly credits are an allowance that refreshes on a billing cycle rather than a balance you own outright. The important question is what happens to the unused portion at the end of the month, and the answers are not consistent enough to assume.

Salesforce Flex Credits do not roll over. Microsoft Copilot Credits do not roll over. Clay splits the difference in a way worth knowing: Actions never roll over, while Data Credits bank up to twice your monthly allowance, and Clay charges a 30 percent premium on top-ups bought mid-month. Glean's FlexCredit Terms take a third position again, expiring credits one year after purchase or issuance and making them non-refundable, with all sales final.

So "monthly credits" can mean use-it-or-lose-it in thirty days, bankable to a cap, or a twelve month cliff, depending entirely on whose contract you are signing. Anyone who bought a large pack for the discount should know which of those three they agreed to.

Do AI credits expire?

Often, yes, and the expiry rule is usually in a separate terms document rather than on the pricing page. Glean's FlexCredit Terms, version 9.25.25, state that credits expire one year after the date of purchase or issuance if unused and are not refundable except where required by law. That is a real commitment to model, because a discount for buying two years of usage up front is not a discount if half of it evaporates.

There is a practical consequence for finance teams here. A prepaid credit pack is committed spend that sits outside your normal per-seat software line, and it is easy for it to go untracked until someone notices the balance is gone. It needs to be visible in whatever system already handles your other committed purchase order approvals, with the expiry date attached, or it will quietly become a write-off.

Why credit pricing gets expensive exactly when the tool is working

This is the structural problem with consumption billing and it is worth stating plainly. A credit meter charges you more in the months the software is doing the most for you. The month your team finally adopts the tool is the month the bill spikes. The quarter you run a big campaign is the quarter you blow through the allowance.

That is backwards from how most small businesses budget, and it produces a specific pathology: teams learn to use the expensive features less. We have watched buyers turn off premium models to protect a credit balance, which means they bought a capability and then rationed it. Glean's own rate card shows why the temptation is strong, with premium models costing roughly five times standard models on the same query type.

The multipliers compound in ways nobody models up front. A single toggle in an admin console can be a five times cost change. We wrote up the specific mechanisms in why AI agent bills exceed estimates, and the underlying pricing models are compared in AI agent pricing models explained.

Questions to ask before you sign a credit-based contract

Six questions, and get the answers in writing rather than in a call.

1. What does one credit cost in dollars? If the vendor will not put a number on this, you cannot forecast anything and you should say so.
2. What does each action I actually plan to run consume? Not the cheapest action on the card. The one your workflow depends on.
3. What is the spread, not the average? Glean publishes a 50th and a 90th percentile, and on agent runs those differ by more than sixteen times. A budget built on a median is a coin flip.
4. What happens when the pool runs out? Billed in arrears, hard stop, or degraded service. All three exist in the market today.
5. Do unused credits roll over, and when do they expire? Get the exact date rule, not "they carry over".
6. Which settings let a user increase consumption? If an admin toggle multiplies cost by five, that toggle is a budget control and it should be governed like one.

The alternative: not buying a meter at all

Credits are not the only way to buy AI, and for a lot of small businesses they are the wrong way. The whole apparatus above exists to allocate a variable cost across customers with wildly different usage. If your usage is fairly steady and fairly predictable, you are paying the complexity cost of a system built for someone else's problem.

WorkAgent is flat at $149 a month for the account. No seats, so it does not multiply by headcount, and no credit meter, so the bill does not rise in the months the agent is busiest. There is nothing to forecast, no pool to ration and no admin toggle that quietly changes your run rate. We do no voice work at all, and intake is queued rather than instant, so it is not the right tool if you need calls handled or access this afternoon.

That trade is worth making only if the shape fits. If you need permission-aware search across a hundred enterprise systems, buy the enterprise search product and negotiate the credit rate properly. If what you need is a recurring job taken off your plate and returned finished, the meter was never buying you anything. The honest comparison is laid out on AI agent pricing, and what the work looks like day to day is on AI assistant for business.