Dashboard

How to Run an Affiliate Program for an AI Product

Your cost of goods is not zero, so commission set on revenue can pay a partner more than the account earns you. Here is the margin-based version.

Manuele Estivo
Manuele Estivo
Growth & SEO Lead
6 September 20261 min read

Running an affiliate program for an AI product breaks the standard playbook in one specific way: your cost of goods is not zero. A normal software affiliate can pay 30% recurring forever because the marginal cost of another user is a rounding error. An AI product pays inference costs on every referred user for as long as they stay, which means a commission set on revenue can quietly hand a partner more than the account earns you.

The fix is not to pay less. It is to set commission on gross margin and to structure the payout so it clears your payback period rather than preceding it.

Work out what you can actually pay

Three numbers, and you probably have all of them.

  • Average revenue per account per month. Not list price, actual realised revenue after discounts.

  • Variable cost per account per month. Model API spend, plus anything else that scales per user: vector storage, transcription minutes, egress.

  • Gross margin per account. Revenue minus variable cost.

Now the number nobody calculates: gross margin over the expected lifetime of the account, which is monthly gross margin divided by monthly churn rate.

Conventional SaaS

Typical AI product

Monthly revenue per account

£40

£40

Variable cost per account

£2

£13

Gross margin

£38 (95%)

£27 (68%)

Monthly churn

4%

6%

Lifetime gross margin

£950

£450

Safe first-year commission at 20% of lifetime margin

£190

£90

Same headline price, less than half the affordable commission. A partner offered "30% recurring" on the right-hand column is being paid £12 a month against £27 of margin, before you have covered any support cost, and while your inference bill is the one number in the table most likely to move.

If your variable cost per account is a mystery, that is the prerequisite job, and working out what it costs to run an AI-built app is where to start.

Pick a structure that matches your margin shape

Four structures, and the right one depends on whether your costs are front-loaded or flat.

Flat one-off bounty. A fixed amount per converted paid account, paid once. Simplest to explain, easiest to budget, and it caps your exposure at exactly the moment your costs are least predictable. Best default for a product under two years old.

Recurring percentage with a term cap. For example 20% of revenue for 12 months, then nothing. Attractive to partners, and the cap means a customer who stays four years stops costing you commission at the point they become genuinely profitable.

Percentage of gross margin. Honest, aligns incentives properly, and almost nobody will accept it because partners cannot verify your costs. Useful for a small number of close partners, unworkable at scale.

Tiered bounty by plan. Different amounts for different plans, weighted by the margin each plan actually carries. This is the one that handles the usual AI-product problem of a heavy-usage plan being your worst margin, not your best.

Whatever you choose, publish it as a number, not a range. "Up to 40%" reads as a haggling invitation and attracts exactly the partners you do not want.

The clauses that save you later

Write these into the terms before you have a single affiliate, because adding them afterwards looks like a retreat.

  • A cooling-off period before payout. Commission clears 30 to 45 days after the first payment, not on signup. This is your only defence against signup fraud and it costs you nothing with honest partners.

  • Clawback on refund or chargeback. If the account refunds inside the window, the commission reverses. Anyone running a subscription product will meet this eventually, and handling chargebacks on an AI subscription is the related mess worth reading first.

  • No bidding on your brand terms. Otherwise you pay commission on traffic that was already yours, twice: once to the ad platform, once to the partner.

  • No coupon-site placement without approval. Coupon sites intercept purchase intent at checkout rather than creating it.

  • Self-referral is void. Say it explicitly, including for the partner's own company accounts.

  • You may change rates with 30 days notice, existing referrals honoured for their term. Your inference costs will change. This clause is how you survive that without breaking trust.

Attribution, kept boring

Last-click, 30-day cookie, first-party where you can. Resist longer windows: a 90-day cookie mostly means paying for customers who would have converted anyway.

Track the affiliate ID against the account at signup and store it on the account record permanently. You need it at the point of refund, at the point of upgrade, and at the point somebody disputes a payout, and reconstructing it from analytics later is not possible.

Give partners a dashboard showing clicks, signups, conversions and pending versus cleared commission. The single most common reason good affiliates go quiet is that they cannot tell whether they are being paid correctly.

Where affiliates fit, and where they do not

Affiliate programs work when there is an existing audience that trusts a recommendation in your category: newsletter operators, YouTube reviewers, agencies serving your ICP. They work badly as a first growth channel, because a partner needs an audience already, and if you have not yet found your first hundred users you have nothing for them to point at.

Agencies and consultants who deploy your product for clients are a different and usually better relationship. They want margin on delivery, not a referral fee, which is closer to reselling than affiliating, and turning a happy client into referrals is the lower-effort version of the same instinct.

For the wider set of options before you commit to this one, AI monetization strategies covers what else is on the table.

On the compliance side, affiliates in the US must disclose the relationship, and the FTC's endorsement guides put responsibility on the advertiser as well as the publisher, so a disclosure requirement belongs in your terms rather than in your hopes.

FAQ

What commission rate is normal for an AI product?

There is no normal, because variable costs differ enormously. Work from lifetime gross margin and pay a fifth of it in year one as a starting point, then adjust once you have real churn data.

One-off bounty or recurring commission?

One-off if your inference costs are large or volatile. Recurring with a 12-month cap if your margins are stable and you need the program to be competitive against other offers a partner is weighing.

Should I use an affiliate platform or build it?

Use a platform. Tracking, tax forms and payouts across countries are more work than they look, and none of it differentiates your product.

How do I stop affiliates from cannibalising organic signups?

Ban brand-term bidding, exclude coupon sites, and check whether referred accounts show different retention from organic ones. If they retain worse, the traffic was intent you already had.

How did this land?

About the author

Manuele Estivo
Manuele Estivo

Growth & SEO Lead

Manuele covers distribution: SEO, content strategy, and how AI-built products find their first thousand users. He tests everything he recommends.

Share

Get the next post in your inbox

One email a month. Product updates, engineering posts, and the best of Built with Swarmz.

I agree to receive emails about AI building tips and Swarmz product news. Unsubscribe any time.