AI Product Pricing Strategy: Seats, Usage or Outcomes?

The right AI pricing strategy depends on where your costs and your customer's value both scale. Seat pricing works when value tracks the number of people using the product and inference costs per user are predictable. Usage pricing works when value and cost both scale with volume (tokens, API calls, documents, minutes). Outcome pricing works when you can measure a result the buyer already pays for, such as a resolved ticket or a qualified lead. Most AI startups end up on a hybrid: a platform fee or seats for predictability, plus usage or outcome charges above an included allowance, with prices set so gross margin stays above roughly 60 to 70 percent after inference costs.

That last part is what separates AI pricing from classic SaaS pricing. Every extra unit of use costs you real money in model calls. Price like a 2015 SaaS company and your best customers can become your least profitable ones.

Why AI pricing is different from SaaS pricing

Traditional SaaS has near-zero marginal cost. One more user barely changes your hosting bill, so per-seat pricing with 80 percent-plus gross margins was the default.

AI products break that in three ways:

• Variable cost per use. A heavy user can cost 20 to 50 times more in inference than a light one on the same plan.
• Falling model prices. Model costs for a given capability have dropped sharply year over year. Price too close to today's cost and you leave money on the table. Lock in too generous an allowance and you're exposed if you move to a pricier model.
• Fewer seats, more automation. If your agent replaces work a person used to do, the buyer will have fewer seats over time. Seat pricing then punishes your own success.

The four pricing models compared

There's no universally right answer. There is usually a wrong answer for your specific product, and it's often whatever your closest competitor picked without thinking.

Margin math: the formula to run before you publish a price

Before you set a number, work out your cost to serve one unit of value. Here's a simple version.

Inference cost per unit = (avg input tokens x input price + avg output tokens x output price) x calls per unit
Cost to serve per unit = inference cost + retrieval/vector DB + hosting + third-party APIs + support allocation
Gross margin = (price per unit - cost to serve per unit) / price per unit
Target: gross margin of 60 to 70 percent or better at your median customer, and positive at your 90th-percentile heavy user

A worked illustration

Say you run an AI contract-review tool. These numbers are invented to show the math.

• One review uses about 6 model calls.
• Each call averages 8,000 input tokens and 1,000 output tokens.
• Assume your model costs $3 per million input tokens and $15 per million output tokens.

Per call: 8,000 x $3 / 1,000,000 = $0.024 input, plus 1,000 x $15 / 1,000,000 = $0.015 output. Total about $0.039 per call. Six calls makes roughly $0.23 per review in inference. Add retrieval, hosting and a support allocation and call it $0.35 per review.

Now test three pricing options:

The seat plan looks friendly and loses money on your best customer. The pure usage plan protects margin but gives the buyer an unpredictable bill. The hybrid keeps margin healthy and gives finance teams a floor they can budget.

Run this on a spreadsheet with your real token counts from production logs, not estimates. Then rerun it every quarter, because model prices and your prompts both change.

How to choose: a decision checklist

• Value scales with people (lean seats) or with work done (lean usage)
• A clear outcome exists that the buyer already pays for today (test outcome pricing)
• Cost to serve varies a lot between light and heavy users (avoid pure seats)
• Procurement needs a fixed annual number (add a platform fee or committed spend)
• You sell to developers (show transparent usage pricing and a free tier)
• Your product reduces the buyer's headcount (don't tie revenue to seats)
• You can explain the price in one sentence (if not, simplify)

Typical starting points by AI category

These are common patterns, not rules. Use them as a first draft, then run the margin math.

If you're using a free tier or trial to drive adoption, product-led growth for AI covers how to cap free usage so inference costs don't eat the funnel.

Plan for model prices to change

Model prices tend to fall, and you'll also switch models as better ones arrive. Price on the value of the unit (a review, a resolution, a report), not on the token cost underneath. That way a cheaper model widens your margin instead of forcing a price cut, and a pricier model doesn't force an awkward increase. Review your cost-to-serve sheet every quarter and before any model migration.

Pricing page anatomy for AI products

A pricing page is a sales page. For AI products it also has to defuse the fear of a surprise bill.

A pricing calculator ("estimate your monthly cost") is one of the highest-converting elements you can add to a usage-based page. It turns an abstract per-unit price into a number a buyer can take to their manager.

Pricing is a marketing message

Your pricing model tells the market what you think you are.

Charge per seat and you're telling buyers you're a tool their team uses. Charge per outcome and you're telling them you're a worker who gets paid for results. Charge per token and you're telling them you're infrastructure.

That framing shows up everywhere: in how analysts and journalists categorise you, in which budget line pays for you, and in which competitors buyers compare you against. An AI agent priced per resolved ticket competes with a support team's headcount budget, which is much larger than a software line item. The same agent priced per seat competes with every other help desk tool.

When I work on launch positioning with AI founders, pricing comes up early for exactly this reason. The pricing model is often the clearest proof of a positioning claim. "We only get paid when the ticket is resolved" is a stronger headline than most taglines. If you're refining positioning alongside pricing, the positioning statement template for AI startups is a good companion.

Common AI pricing mistakes

• Copying a competitor's per-seat price without knowing your own cost to serve.
• Unlimited plans on a product with real inference costs.
• Charging in tokens to non-technical buyers who have no idea what a token is.
• No spend caps, so one runaway integration produces a bill that ends the relationship.
• Never raising prices. Early customers can be grandfathered, but new customers should pay for the value you've added.
• Hiding pricing entirely at seed stage. For self-serve and mid-market buyers, a missing price page often means a missing deal.

Testing your pricing without breaking trust

Pricing experiments are fine. Surprise price changes for existing customers are not.

• Test with new signups only, by cohort or by segment.
• Run pricing interviews: show three plan structures and ask which they'd pick and why.
• Watch conversion from pricing page to trial and from trial to paid, not just total signups.
• Track net revenue retention by plan. Usage and hybrid plans usually expand faster.
• Give existing customers at least 60 to 90 days' notice on any increase.

Pricing also shapes your whole go-to-market plan, from which channels can pay back to how much you can spend acquiring a customer. The free GTM planner builds a 90-day plan by category and stage, and the marketing budget calculator shows what your pricing means for CAC and payback. For the metrics side, see CAC, LTV and payback period explained.

The cheapest model to serve is rarely the cheapest model to sell. Pick the one your buyer can say yes to and your margins can survive.

Want to pressure-test your pricing and launch story together? Build your 90-day plan with the free GTM planner.