When the invoice moves

Affiliate software for AI SaaS: what changes when you bill on usage

By Jimi Barkway · Published 5 September 2026 · Part of the comparison pages

In brief

Affiliates get paid a percentage of what your customer paid you. If part of that payment is usage the customer burned, you are paying commission on your own model bill. Take a $20 plan with $10 of usage on top that costs you $6 to serve: a 20% commission takes $6 instead of $4, so you keep a fifth of the usage against four fifths of the plan. Once your model costs reach 80% of what you charge for usage, that part of the bill earns you nothing at all. No tool I could check handles this, including ours: 165,709 bytes of one vendor's documentation never mentions it. The fix is to sell the usage as its own product in Stripe and write one rule paying 0% on it.

What changes when your product bills on usage?

In shortAffiliates get paid a percentage of what your customer paid you. That is fine when your price is fixed. It stops being fine when part of that payment is money you already owe OpenAI, because the percentage comes off your supplier's bill as well as your own margin.

Most SaaS charges the same amount every month. An AI product often does not. Usage-based billing means the bill moves with how much the customer used, rather than sitting at one price. The extra on that invoice is tokens you had to buy from a model provider before you could sell them. You are reselling something you paid for.

Affiliate software does not know any of that. It sees one invoice and takes its percentage of the whole thing.

I checked AffiliateRail's own code on 5 September 2026 rather than trust the marketing. The number it pays commission on is the invoice total. There is no special handling for usage anywhere in it. The only thing you can subtract is tax. Not what the usage cost you to serve.

Two more things from that read. When an invoice has several lines on it, the rules only ever look at the first line, so whichever line Stripe happens to put first is the one that decides. And a sale is labelled either subscription or one_time. There is no label for usage at all.

How much of the commission lands on your AI bill?

In shortMore than founders expect. On a $20 plan with $10 of usage on top, a 20% commission takes $6 instead of $4. You keep 80p in the pound on the plan and 20p in the pound on the usage.

Let's say your plan is $20 a month. A heavy customer adds $10 of usage. That usage costs you $6 in model calls. Your affiliate rate is 20%, paid every month. The $6 is a stand-in for your real cost, and it is the first number to swap for your own.

A $20 plan with and without $10 of metered usage, at 20% commission and $6 of inference cost
Base onlyBase plus usage
Invoice$20.00$30.00
Commission at 20%$4.00$6.00
Inference on the metered line$0.00$6.00
After commission and inference$16.00$18.00

Look at the bottom row. The busy customer paid you $10 more and left you $2 better off. That $10 of usage paid $2 of commission and cost $6 to serve, so you kept $2 of it. The plan is the opposite: $20 comes in, $4 goes out in commission, and you owe a model provider nothing.

So the plan keeps 80% and the usage keeps 20% (arithmetic, 5 September 2026). The heavy month is still worth having. It is worth a quarter as much per dollar as it looks.

Now the rule to remember, which matters more than the example. Add what the usage costs you to what you pay in commission. When those two reach 100% of what you charge for the usage, that part of the bill is earning you nothing. At a 20% commission, that happens when the usage costs you 80% of what you sell it for, so $8 on $10.

Past that point you are losing money on every extra call, and selling more does not save you. Twice the usage is twice the cost.

So work out one number before you set a rate: what your usage costs you, as a share of what you charge for it. For example, charge $10 for something that costs you $3 and it carries a commission comfortably. Charge $10 for something that costs you $8 and it does not.

This is why module 1 of the manual starts from your margin rather than what everyone else pays. On a fixed-price SaaS the difference is small. When a third of the invoice is somebody else's cost, it decides whether affiliates work for you at all.

Do any of these tools handle usage billing?

In shortNot that I could find. Tolt publishes its entire documentation, 165,709 bytes of it, and the words "metered", "usage-based", "per-unit" and "overage" appear zero times. Dub's index returns nothing on the same terms. Neither does AffiliateRail's own guide. That is two vendors and us, not the whole category.

You can check this yourself, which is why I am putting the method in. Both vendors publish their docs in a form you can search in one go: Tolt's full text and Dub's index. I searched five terms on 5 September 2026 and got nothing from either. The two tests are not equal, though. Searching the full text is strong evidence. Searching an index of page titles is weak.

Our own docs come out the same way. AffiliateRail's flows guide lists everything a commission rule is allowed to look at: amount, currency, billing type, billing interval, product id, product name, first payment, payment number, coupon code, how the sale was tracked, months since conversion, customer status and customer country. Thirteen things. None of them tells you whether a line was usage.

One vendor deserves credit here. Polar documents usage-based billing properly, including meters, credits and how to count LLM calls, and it has an affiliate feature too. Explaining usage billing is not the same as explaining how to pay commission on it, and there is no page joining the two up. It is still the closest anyone gets, and leaving it out would be dishonest.

So when a tool says it supports usage-based SaaS, what that means is the invoice arrives and it takes a cut. True, and not the same as handling usage.

What can you actually do about it?

In shortSplit the usage out in Stripe so it is a separate product, then write one rule that pays 0% on it. Product name and product id are the two things every tool here can check, and they are the whole answer.

A Price in Stripe is the record of what something costs and how often you charge for it. Most founders put the plan and the usage on one. Put them on two instead and your affiliate tool can finally tell them apart, because it can see the product name.

A commission flow is the rule deciding what a sale pays, and it is where those two Prices get treated differently.

This is a billing change, not an affiliate one, and your accountant will thank you for it regardless.

Then the rule is one line. Pay your normal percentage on the plan. Pay nothing on the usage. Your partner still earns on every customer they bring and every upgrade to a bigger plan, which is the part they influenced. They stop earning on how much a customer burns, which they did not.

Two warnings before you rely on it. Remember that only the first line of an invoice gets read, so the split has to be real in Stripe and not something you track in a spreadsheet. And if your usage is baked into the plan price, there is nothing to separate, so the margin has to come out of your commission rate instead.

Set the rule once, on the product that earned it

Commission conditions on product, plan, interval, coupon and country, with a plain-English trace on every commission. Fourteen days, no card.

Start free trial

What does Stripe recommend for usage billing now?

In shortMetronome, for anything new. Stripe says to use Billing Meters only if you are already on them, and calls adding usage pricing to a flat-rate plan "the clearest signal" to choose Metronome instead.

This changed recently enough to matter, because most articles on the subject were written before it. Stripe's usage-based billing page, read on 5 September 2026, gives you two routes and is blunt about which one to take. Metronome is "Stripe's primary usage-based billing platform, recommended for all new integrations". Billing Meters "remains fully supported for existing integrations". The same page says Stripe now owns Metronome.

Metronome is Stripe's primary usage-based billing platform, recommended for all new integrations. Basic usage-based billing, built on the Billing Meters API, is a lower-level primitive that remains fully supported for existing integrations.

Basic usage-based billing, Stripe docs, checked 5 September 2026

One line on that page matters for commissions. Billing Meters "only reconciles usage at invoice time". That is the mechanism behind this whole article: the usage gets totted up and dropped into the invoice, and the invoice is what your affiliate tool takes its cut of. How you bill and what you pay commission on are one decision, made in Stripe.

What happens when a customer refunds?

In shortA commission you have not sent yet gets cancelled. One you already sent comes off the partner's next payout as a negative. The setting that protects you is the hold period, which is how long you sit on a commission before you are allowed to pay it. On a product people quit fast it needs to be longer than your refund window.

A clawback is money you take back from a partner after you have already paid it, and it is awkward every time. Far better to not have paid it yet. AffiliateRail cancels a commission you have not sent and books one you have as a negative against the next payout. A customer only counts as refunded once every one of their sales is.

The category barely talks about this. Across two vendors' own teaching material it amounts to one line in one academy and one bullet in a checklist. For a product where somebody can join, burn a month of model calls and then dispute the charge, that is an odd thing to skip.

Set your hold at least as long as your refund window, which the one-week guide covers on day one. Then watch refund rate partner by partner rather than as one number. A partner sending you the wrong sort of customer shows up there months before it shows up in churn.

So which tool should an AI SaaS use?

In shortOn usage they are all the same, because none of the ones I could check handles it. So pick on the things that do differ: whether a rule can check the product, how long you can hold a commission, and whether the tool takes a cut of your payouts.

Checking the product is the important one. If a tool cannot write a rule against product name, the split above is closed to you, and your margin has to come out of the headline rate for every partner.

Hold length matters more here than in most categories. If people cancel quickly you want a long hold, and a tool with a short fixed hold has decided your refund policy for you.

The payout fee is arithmetic rather than taste, and the flat-fee page has all thirteen platforms with the date each price was read. A percentage of payouts stacks on top of the problem above: you are already paying commission on your supplier's bill, and then a fee takes a cut of that commission.

Start with your own number, not with the tool. Work out what your usage costs you against what you charge for it, decide whether that part of your revenue can carry a commission at all, and only then go shopping. The calculator handles that last step.

Common questions

Does the commission really come out of my inference cost?

Out of the invoice that contains it, which comes to the same thing. The base is what your customer paid and there is no line-level exclusion. So if a third of the invoice is metered usage, a third of the commission is paid on that usage. Tax is the only built-in deduction.

Can I pay a lower rate to compensate?

You can, and it costs you the good partners. A rate low enough to survive your heaviest usage month underpays on the base plan every other month, and the base plan is the part a partner sold. Splitting the Price keeps the rate honest.

What if usage is bundled into my plan price?

Then there is nothing to condition on, and margin has to come out of the rate. Work it from a heavy month rather than an average one. The average hides exactly the months that hurt, and module 1 of the manual has the method.

Which billing approach should a new AI product use?

Stripe's own answer is Metronome for all new integrations, with Billing Meters kept for existing ones. Its page calls adding usage pricing to a flat-rate plan the clearest signal to pick Metronome. Read it before you build: this guidance changed recently.

How long should the hold period be on an AI product?

At least as long as your refund window, and longer than you would set for a conventional SaaS. Fast cancellations are common on consumption products. A hold that expires before the disputes arrive turns every refund into a clawback conversation.

Is any tool better than the others at usage-based commissions?

None of the two whose docs I could search documents it, and neither does AffiliateRail. So choose on three things that do differ. Whether a rule can test the product, how long the hold can be, and whether the payout fee is a cut of what you pay out.

Where these facts come from

Fact-checked and reviewed by Jimi Barkway on 5 September 2026. Every figure above was read off the document named here on the date beside it. To contact AffiliateRail about one, email support@affiliaterail.com and the figure is corrected and the date moved.

Start

Pay commission on the plan, not on your token bill

Rules that condition on product, plan, interval, coupon and country, and 0% of every payout. No card for fourteen days. Full refund within 30 days of your first payment.