How to Price an AI Product Without Killing Your Margin

Every time someone uses your AI product, you pay. That one sentence breaks most of what founders learned about pricing software.

Classic SaaS pricing worked because the marginal cost of one more user was close to zero. Server capacity was already paid for. A customer who logged in fifty times a day cost you roughly the same as one who logged in twice a month. That assumption is dead for AI products. Every call to a model burns real compute, and that compute shows up in cost of goods sold rather than in a fixed engineering budget. So the question of how to price an AI product isn't a marketing exercise. It's a survival calculation.

Here's the part most first-time founders miss. You can grow revenue fast, celebrate, and still be losing money on your best customers. Let's look at what the numbers actually say in 2026, and then at what to do about them.

Why is pricing an AI product different from pricing SaaS?

Because your costs scale with usage, and traditional SaaS pricing was built on the assumption that they don't. In SaaS, cost of goods sold typically runs 10 to 25 percent of revenue. At scaling-stage AI B2B companies, inference alone averages about 23 percent of revenue. That means for every million dollars of AI product revenue, roughly $230,000 disappears into model calls before you've paid for hosting, support, or anything else.

Stack the rest of COGS on top and the picture gets uncomfortable. AI-native companies are running total COGS around 40 to 50 percent, which leaves gross margins of 50 to 60 percent. Mature SaaS sits at 70 to 90 percent. One 2026 projection puts the AI-native average at about 52 percent, which is real improvement from 41 percent in 2024, but still nowhere near what investors and founders were trained to expect.

If you're a SaaS company bolting AI features onto an existing product, the typical hit is 12 to 17 points of gross margin depending on how hard you've optimized. That's not a rounding error. That's the difference between a company that can afford a sales team and one that can't.

What gross margin should an AI product actually target?

Aim for 60 percent or better, and treat anything under 50 percent as a problem you need a written plan to fix. Not a vague intention. A plan with dates.

The useful mental model is to stop thinking of your AI feature as software and start thinking of it as a service with a cost per unit delivered, the way a manufacturer thinks about cost per widget. Pick the unit that matters in your product: a resolved support ticket, a generated report, a document reviewed, a call transcribed. Then work out what that unit costs you at current model prices, at realistic usage, including the retries and the failed attempts that nobody puts in the spreadsheet.

Do that math before you pick a price, not after. I've watched founders set a price by copying a competitor, then discover four months later that the competitor is running a different model at a tenth of the cost. Their price was never the right price for your cost structure.

What are the AI pricing models founders choose from in 2026?

There are four real options, and most companies end up blending two of them.

Per seat. Flat fee per user per month. Predictable for the buyer, dangerous for you, because a single heavy user can consume twenty times what an average one does at exactly the same price. Per seat is also losing ground fast: its share of SaaS pricing fell from 21 percent to 15 percent in twelve months. A Cruxy survey of 300 SaaS CEOs in April 2026 found 97 percent planning to retire seat-based pricing within two years.

Usage based. Customers pay for what they consume, usually metered in credits, tokens, or actions. Roughly 80 percent of customers say usage-based pricing lines up better with the value they get. The catch is budget volatility. Even as token prices fell about 80 percent year over year, total AI spending grew 320 percent, which tells you exactly how buyers behave when a meter is running and the product is good.

Outcome based. Customers pay per result delivered. Intercom's Fin charges $0.99 per billable outcome, HubSpot dropped its Customer Agent to $0.50 per resolved conversation in April 2026, Zendesk's AI agents run roughly $1.20 to $1.50 per verified resolution, and Salesforce Agentforce launched at $2.00 per conversation. This is the cleanest value story you can tell a buyer. It's also the hardest to define, because you and your customer have to agree on what counts as a resolution.

Hybrid. A predictable base fee plus metered charges for AI work. This has quietly become the most common model in 2026, and for good reason. It gives you recurring revenue you can forecast and a variable component that covers your variable cost. Intercom does exactly this: seats at $29 to $139 per agent per month, plus $0.99 per outcome, plus a 50-outcome monthly minimum.

How do you price an AI product when you have zero customers?

Build the cost model first, then pick a price that clears your target margin at your worst realistic usage, not your average one.

Concretely, four steps:

1. Define your billable unit and measure what it costs. Run 50 real tasks end to end and record actual token spend, including failures and retries.
2. Multiply by your expected monthly volume per customer, then multiply that by three. Heavy users are not an edge case, they are your most engaged customers and they will find you.
3. Set the price so that even at 3x expected usage you still clear 60 percent margin.
4. Put a usage ceiling in the plan from day one. Adding one later feels like a price increase. Having one from the start is just how the product works.

This is the moment where the cost model belongs next to the rest of your financial planning, not in a separate file nobody opens. A spreadsheet works. So does Notion, or a planning tool like Foundra that walks first-time founders through the projection alongside the rest of the business model. What matters is that your pricing page and your financial model are looking at the same numbers.

One more thing on early pricing. Charge something from the first customer. Free tiers on AI products are a direct transfer from your bank account to a model provider, and the usage data you get from people who pay nothing tells you almost nothing about willingness to pay.

Should you charge per seat, per usage, or per outcome?

Match the model to how variable your costs are and how measurable your value is.

Your situation	Model that fits
Light AI features, usage roughly even across users	Per seat, watch the outliers
Usage varies wildly between customers	Hybrid: base fee plus metered overage
You can point at a countable result the customer cares about	Outcome based
Developer tool or API	Pure usage based
Enterprise buyer who needs budget certainty	Committed contract with a usage pool

The honest test for outcome pricing is whether you could write the definition of a successful outcome on one line and have your customer agree to it without negotiation. "Ticket resolved without human escalation" passes. "Improved productivity" does not. Zendesk's May 2026 move to bill only for LLM-verified resolutions is a sign of where this goes: buyers stopped accepting vendor claims about what counted, so the definition had to get tighter.

How do you keep one heavy user from bankrupting you?

Cap it, meter it, and make the limit visible in the product before the customer hits it.

The specific controls worth building early:

• A hard usage ceiling per plan, with a clear upgrade path when someone reaches it
• A live usage meter in the app, so nobody is surprised at the end of the month
• Per-account alerts on your side when a customer crosses a margin threshold
• Model routing, so cheap requests go to a cheap model and only the hard ones hit a flagship

That last one is where the real money is. The price gap between model tiers is enormous right now. Google's Gemini 3.1 Flash runs $0.10 per million input tokens and $0.40 per million output, while flagship models sit at $5 to $30 per million. If 70 percent of your requests are simple and you're sending all of them to a frontier model, you're burning margin on purpose.

And the ground keeps shifting in your favor. Frontier-class pricing per million tokens is roughly a fifth of what it was two years ago. Andreessen Horowitz's analysis found the cost for an LLM of equivalent performance dropping about 10x per year, faster than compute during the PC era or bandwidth during the dotcom boom. Your margin problem this quarter may partly solve itself next year. Partly. Don't build a business that depends on it.

What pricing mistakes are AI founders making right now?

The expensive ones are all variations on the same theme: changing the meter without preparing the customer.

Cursor is the case study everyone in this category should read. In mid-2025 the company moved Pro users from request-based billing to monthly usage credits, and described the change using the phrase "rate limits," which almost nobody understood. Users started reporting unexpected charges of $10 to $20 a day. One team burned through a $7,000 annual plan in a single day. The CEO apologized, refunds went out, spending controls and usage visibility got better, but the old plan never came back. Cursor survived it because the product is loved. A seed-stage company with 40 customers does not get that grace.

The other recurring mistakes:

• Inventing a credit currency nobody can price. If a customer can't answer "what will this cost me next month" in under ten seconds, your pricing page has failed.
• Pricing against a competitor's cost structure instead of your own. They may be running a distilled model on their own hardware.
• Forgetting retries and failures in the cost model. These are frequently 20 to 30 percent of real token spend.
• Giving unlimited usage to land a logo. The logo is worth less than the bill.
• Treating the AI feature as free marketing inside an existing SaaS plan. That's the 12 to 17 point margin hit, and it usually arrives without anyone noticing for two quarters.

How often should you change your pricing?

Review the cost side monthly, and revisit the price itself every six to twelve months in the first two years.

Monthly review is not about changing the price. It's about watching cost per unit and margin per account so you catch a problem while it's small. Model prices move, usage patterns drift, and a feature you shipped in March can quietly double your inference bill by June.

When you do change the price, give existing customers notice and grandfather the early ones for a defined window. Early customers took a risk on an unproven product. Protecting them costs you very little and buys you the references you'll need later.

Key takeaways

• AI products run 50 to 60 percent gross margins, not the 70 to 90 percent of mature SaaS. Plan around that number, don't fight it.
• Inference alone averages about 23 percent of revenue at scaling-stage AI companies. Model your cost per unit before you pick a price.
• Hybrid pricing, a predictable base plus a meter, is the most common 2026 model because it matches recurring revenue to recurring cost.
• Outcome pricing tells the best value story but only works when the outcome is countable and both sides agree on the definition.
• Build usage caps, live meters, and model routing before you need them. Retrofitting a limit reads as a price increase.
• Token costs are falling roughly 10x a year for equivalent performance. That helps, but never build a plan that requires it.

FAQ

What is a good gross margin for an AI startup?

Sixty percent or better is a reasonable target for an AI-native product in 2026. The category average is around 52 percent. Below 50 percent you're in territory where growth makes the cash problem worse rather than better.

Should I use usage-based or subscription pricing for my AI product?

Most companies land on both. A base subscription covers your fixed costs and gives you forecastable revenue, while a usage or outcome component covers the compute that scales with each customer. Pure subscription only works if usage is close to uniform across your customers.

How much does it cost to run an AI product?

It depends entirely on the model tier and request volume. Cheap models run about $0.10 per million input tokens, flagship models $5 to $30. The practical answer is to measure your own cost per billable unit across 50 real tasks rather than estimating from published rates.

What is outcome-based pricing?

Charging per result delivered rather than per user or per unit of consumption. Intercom's Fin charges $0.99 per resolved conversation, HubSpot $0.50, Salesforce Agentforce $2.00. It aligns your revenue with customer value, but it requires a definition of "outcome" that survives a procurement conversation.

Can I raise prices after launch?

Yes, and most AI companies will have to. Give notice, explain what changed, and grandfather existing customers for a set period. What damages trust isn't the increase, it's changing how the meter works without making the new math easy to understand.

Do I need to charge from day one?

Charge early. Every free user of an AI product costs you real money, and free-tier usage data tells you nothing about what someone will actually pay. A small price from your first customer is better research than a thousand free signups.