How do I know if I'm losing money on AI features?

Pull last month's model-provider invoice, divide by paid seats, and compare to your per-seat ARPU. If AI cost is more than ~15% of ARPU on average — or more than 100% for your top 5% of users — you're subsidizing power users out of margin.

Shouldn't AI just be included in the base plan?

Including it is fine; pricing it as if it costs zero is not. Token consumption is power-law distributed: the top 5% of users consume 50–80% of inference. A flat seat fee guarantees the heaviest users are unprofitable, and the lightest users subsidize them.

What markup should I apply to AI usage?

20–40% is the typical range in 2026. Below 20% and you have no buffer for model price changes; above 50% and customers feel gouged once they price-compare. Start at 30% with a transparent breakdown.

Do I need to be a Merchant-of-Record to rebill AI usage?

No. The cleanest pattern is: you pay the model provider, you meter per end-user, you forward metered events to your existing Stripe Billing as a usage line item. No Connect, no MoR, no PayFac.

How do I prevent one user from running up a huge bill?

Per-user spending caps enforced at the gateway. Once the cap is hit, the next tool call returns a friendly upgrade prompt instead of running. This must be enforced server-side, not in client code.

Are you giving your AI features away for free?

If you shipped AI features on your existing seat price and never instrumented per-user metering, you are almost certainly losing money on your top 5% of users. Token consumption is power-law distributed: a few heavy users eat the majority of inference cost, and a flat seat fee guarantees they're unprofitable. The fix is four moves — meter per end-user, mark up 20–40%, enforce per-user caps, and forward metered events to your existing Stripe Billing.

The quiet margin leak

Most SaaS companies that shipped AI in 2024–2025 are running the same accidental experiment: AI features ship on the existing seat price, model invoices grow every month, and nobody on the finance team can point to the revenue line that pays for it. That's not a free feature — that's a margin transfer from your investors to your heaviest users.

Five-minute diagnostic

Pull last month's invoice from your model provider.
Divide by paid seats. That's your average AI cost per seat.
Compare to your ARPU. If AI is > 15% of ARPU on average, your gross margin is bleeding.
Group inference by end-user. The top 5% will be 50–80% of total tokens.
For those top users, AI cost frequently exceeds the seat price entirely.

Power-law reality

Token consumption follows a power-law, not a normal distribution. Pricing flat is pricing for the median user while paying for the tail.

The fix: meter, mark up, cap, rebill

The architecture is four small pieces, and you don't need to be a Merchant of Record:

Meter input/output tokens per end-user at the gateway layer.
Mark up by 20–40%. Surface the breakdown — buyers respect transparency.
Cap per-user spend with a hard server-side limit. Default low; let admins raise it.
Rebill by forwarding metered events to your existing Stripe Billing as a usage line item.

That's it. No Stripe Connect. No PayFac. No new entity. You stay the merchant of record; the AI line item shows up alongside the seat line on your existing invoices.

Common objections, answered

"Usage-based pricing will confuse my customers"

Two patterns defuse this: bundle a generous monthly allowance per seat (covers 95% of users with zero overage), and show usage live in-product. Customers don't hate metering; they hate surprise.

"My competitors include it free"

They are either VC-funded to do so temporarily, capping invisibly (and breaking power users), or losing money. Watch which strategy survives 2026.

"Building this is a quarter of work"

It used to be. In 2026 the metering pipeline is a managed product — see Monetize. Plug it in front of any model provider, attach your Stripe meter, ship in a week.

What good looks like

Every paid plan has an included AI allowance.
Overage is metered and shown live in-product.
Per-user caps prevent runaway scripts from bankrupting a workspace.
Gross margin on metered AI revenue ≥ 60% after model cost.
Finance can point to one Stripe line item and say "that's the AI business."

The bottom line

AI is the first SaaS feature in a decade with a real marginal cost. Pricing it like it doesn't is the most expensive default you can pick. Meter it, cap it, mark it up, and turn the cost center into a revenue line.

FAQ

How do I know if I'm losing money on AI features?: Pull last month's model-provider invoice, divide by paid seats, and compare to your per-seat ARPU. If AI cost is more than ~15% of ARPU on average — or more than 100% for your top 5% of users — you're subsidizing power users out of margin.
Shouldn't AI just be included in the base plan?: Including it is fine; pricing it as if it costs zero is not. Token consumption is power-law distributed: the top 5% of users consume 50–80% of inference. A flat seat fee guarantees the heaviest users are unprofitable, and the lightest users subsidize them.
What markup should I apply to AI usage?: 20–40% is the typical range in 2026. Below 20% and you have no buffer for model price changes; above 50% and customers feel gouged once they price-compare. Start at 30% with a transparent breakdown.
Do I need to be a Merchant-of-Record to rebill AI usage?: No. The cleanest pattern is: you pay the model provider, you meter per end-user, you forward metered events to your existing Stripe Billing as a usage line item. No Connect, no MoR, no PayFac.
How do I prevent one user from running up a huge bill?: Per-user spending caps enforced at the gateway. Once the cap is hit, the next tool call returns a friendly upgrade prompt instead of running. This must be enforced server-side, not in client code.