How to Price an AI Agent (So You Don’t Lose Money on Every Conversation)
The single fastest way to lose money on an AI product is to charge a flat price for a cost that is variable per conversation. Your subscription is fixed at $49; your model bill is not. One power user with long, tool-heavy chats can cost you more than they pay — and you won’t notice until the invoice lands.
Pricing an AI agent isn’t guesswork. It’s three steps: know your cost floor, pick a model that tracks that cost, and cap the downside. Here’s the whole thing.
Step 1 — Find your cost floor (per conversation, worst case)
You can’t price what you can’t measure. The unit that matters is cost per conversation, not per message — and it’s dominated by input tokens, because every turn re-sends your system prompt, tool definitions, retrieved context and history. We broke the full arithmetic down in The Real Cost of One AI Conversation, and you can plug in your own numbers with the AI agent cost calculator.
Two rules when you do this:
- Model the worst-case month, not the average. Price against the heavy user and the seasonal spike, because that’s when a flat price hurts.
- Include every lever: model tier (premium can be ~9x a flash model), number of tools (each definition is re-sent every call), and retrieval depth. Miss one and your floor is fiction.
Step 2 — Pick a pricing model that tracks the cost
There are three honest ways to price an AI agent. Each trades predictability against margin safety.
| Model | How it works | Best when | Watch out for |
|---|---|---|---|
| Flat subscription | One price, usage included | Costs are stable and you’ve capped them | A heavy user erases your margin silently |
| Usage-based (metered) | Charge per token / message / conversation | Usage varies wildly between customers | Customers can’t forecast it — it’s a hard sell |
| Hybrid (flat + fair-use cap) | Flat plan with a usage ceiling, overage above it | Most SaaS — predictable for buyers, safe for you | Set the cap from real data, not hope |
For most products the hybrid wins: buyers get a predictable bill (which is what they actually want — metered AI is a miserable thing to budget for), and you’re protected above the cap. Set the included allowance at or just above the worst-case you modeled in Step 1, and price overage at your real marginal cost plus margin.
Step 3 — Protect the margin you just priced
Pricing sets the ceiling on revenue; these keep the cost underneath it.
| Lever | Effect on cost |
|---|---|
| Route easy turns to a cheaper model | Biggest single lever — flash vs premium can be ~9x |
| Trim tools the agent never uses | Every tool definition is re-sent on every call — cut the dead weight |
| Cap retrieved context and history | Input tokens are the bill; smaller context = lower every turn |
| Set a fair-use cap per user | Turns an unbounded liability into a known number |
Routing alone often decides whether a product is profitable. Most commerce turns — FAQs, lookups, recommendations — never needed the expensive model; see the model-routing math for why “don’t upgrade, route” usually beats paying for the premium model on every call.
Step 4 — Sell predictability, not the lowest rate
Here’s the counter-intuitive part: the winning price is rarely the cheapest one. Usage-metered pricing looks cheap and reads as risky — buyers can’t forecast it, so they hesitate. A flat plan with a clear cap is easier to say yes to even at a higher number, because the buyer knows exactly what they’ll pay. Predictability is a feature you can charge for.
The short version
- Measure cost per conversation at the worst case, not the average.
- Default to a flat plan with a fair-use cap; meter only the overage.
- Protect margin with model routing, tool hygiene, and context caps.
- Price the predictability — it closes more deals than a low rate.
This is exactly how we price WisWes: flat monthly plans with model usage included, billed around recovered sales rather than a meter that surprises you at month-end. If you want to see the cost side before you set your price, run your numbers through the cost calculator — it shows the raw model spend, so the margin math is yours to set on top.