How do I price an AI product when my costs are variable?
Price on the value of the outcome, then defend that price with a cost floor you have actually measured. Unlike ordinary software, your marginal cost per customer is real, it varies by an order of magnitude between your lightest and heaviest accounts, and it moves when a supplier changes prices — so instrument spend per account before you publish a number, and bill on a unit the customer can predict. The decision founders skip is what happens when someone exceeds what their plan pays for: hard cap, soft cap, or metered overage. Make that choice before launch, because it is what decides whether your heaviest user is a reference customer or a loss.
Why this is not ordinary SaaS pricing
Classic software pricing advice rests on an assumption that stops being true the moment a model is in the loop: that serving one more customer costs approximately nothing. With an AI product it costs something you can measure, it differs wildly between accounts doing nominally the same thing, and it is set by a vendor who can change it without asking you. Gross margin stops being a fact about your business and becomes a live variable you have to watch.
Two things follow. First, you cannot price purely on value and check the margin later, the way SaaS pricing normally works — you need a cost floor per account before you publish a number. Second, the distribution matters far more than the average. Most accounts will use much less than they pay for and a few will use much more; the mean hides both. Look at the ninetieth percentile, because that is the customer who decides whether your pricing survives success.
Pick a unit the customer can predict
The unit you bill on is a communication decision as much as a financial one. Customers forgive a high price. They do not forgive a bill they could not have predicted.
| Model | Best when | Where it breaks |
|---|---|---|
| Flat tier with an included allowance | Usage is bounded and you can size the allowance safely | Heavy accounts quietly erode margin unless the allowance is enforced in the product |
| Per seat | Value scales with the number of humans, not with compute | One seat running an agent all day costs what ten browsing seats do |
| Usage or credits | Consumption varies wildly and your buyers are technical | Unpredictable bills suppress usage — people ration the thing you want them to love |
| Per outcome | The unit is obvious to the buyer: a published post, a qualified reply | You absorb retries and failures, so quality problems become margin problems |
| Hybrid: tier plus metered overage | Most AI products, once you have real usage data | Two numbers to explain; needs a visible usage meter or it reads as a trap |
What we do, and what it cost us to learn
At AgentCeres — the AI Growth Officer at agentceres.com — we landed on the hybrid: every plan carries a monthly model-spend allowance, usage beyond it is metered, and the running total is visible in the dashboard rather than discovered on an invoice. Three specifics are worth copying.
The allowance is deliberately not equal to the price. Above our middle tier the included model spend sits well below what the plan costs, because an allowance sized to the sticker price means zero gross margin at full utilisation — match them one to one and you have not priced a product, you have resold tokens at cost. Our free trial takes the opposite stance: it is card-less with a hard spend cap that pauses scheduled work when it is reached, enforced in the product rather than reconciled on a bill. An uncapped card-less trial is an invitation, and the cap is also the cheapest abuse control we have. If you are still choosing an entry point, free trial versus freemium is the related decision.
The thing we did not plan for was a supplier price change. A model we depend on shipped its generally-available version at three times the input price of the preview it replaced. We absorbed it with a per-model billing multiplier rather than repricing customers mid-cycle — cheaper than the trust cost of a surprise increase, and only possible because we were already recording spend per model per account. If you take one thing from this page: build the meter before you need it. You cannot retroactively discover which customers were expensive.
FAQ
- Should I charge for AI features separately or bundle them into existing plans?
- Bundle when the AI is how the product works, and separate it when it is a genuinely optional add-on with its own cost profile. Bundling fails by raising everyone's price to cover a feature a minority uses heavily; separating fails by putting a paywall around the part that makes the product worth buying, so adoption stalls. If you are unsure, bundle a modest allowance into your existing tiers and meter beyond it. That keeps the product whole while making the heavy accounts visible.
- Is usage-based pricing better than per-seat for an AI product?
- It matches your costs better and it communicates worse. Per-seat pricing is legible to a finance team and disconnected from what you actually spend; usage pricing tracks costs precisely and makes customers ration the product. Most AI companies converge on a hybrid because it is the only structure that addresses both: a predictable base that covers typical use, plus metering for accounts that leave the typical range.
- How do I stop a single customer from destroying my margin?
- Caps and visibility, in that order. A per-account spend limit enforced by the product — pausing background work rather than silently absorbing the cost — turns an unbounded liability into a support conversation. Then show usage to the customer while it accumulates, so reaching the limit is expected rather than a shock. Nobody minds a ceiling they can see coming, and everybody minds a surprise.
Want this done for you?
AgentCeres is a managed AI marketing team — specialists draft the work, you approve what ships. 14-day free trial, from $39/month.