Seats, Usage or Outcomes: Choosing a Pricing Model When AI Does the Work
If your product replaces the person, you cannot keep charging per person. You are billing for the thing you are deleting.
Day 3 was about the level of your price. This one is about the shape. They are different problems, and the shape is the one founders get wrong quietly for years, because a bad pricing model does not produce an error message. It produces a business where your best customers pay the least and your invoice shrinks as your product improves.
The shape has a name: the value metric. It is the unit you charge for, and it decides who can afford you, who expands, and whether growth in your customers' business shows up in yours.
Choose the unit that grows when your customer wins. Everything else in pricing is decoration.
Per-seat pricing dominated software for twenty years because it worked. Value scaled with the number of people using the tool, seats were easy to count, and buyers understood it immediately. That logic held right up until software stopped assisting the person and started doing the task.
Now run the per-seat logic on a product that handles support tickets automatically. The customer starts with 20 agents. Your product works, so next year they need 12. Your revenue falls 40 percent as a direct result of doing your job well. You have built a business that is punished for performance, and no amount of sales effort fixes a metric that points the wrong way.
Those three numbers frame the decision. Andreessen Horowitz's analysis of AI businesses found gross margins materially below classic software, because inference, human review and cloud spend scale with usage instead of sitting flat. Meanwhile Stanford's AI Index reported that the cost of inference for GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024. Your unit costs are real, and they are falling fast. A pricing model that ignores either fact leaves money on the table or takes on losses you cannot see.
The three shapes, and when each one is right
Per seat: charge for the people who log in
Use it when your product makes an individual more effective and the customer adds people as they grow. Design tools, CRMs, collaboration software.
- Buyers understand it instantly and budget for it annually.
- Revenue is predictable, which matters when your cash buffer is thin.
- It breaks when the product reduces headcount, and it encourages password sharing at the exact moment you want more usage.
Test: if your best case study says "we did the same work with four fewer people," per-seat is working against you.
Per usage: charge for the work done
Use it when the product does a countable unit of work: documents parsed, calls transcribed, orders routed, messages sent.
- Revenue grows automatically with the customer's volume, so expansion needs no sales conversation.
- It aligns your bill with your own variable cost, protecting margin.
- It makes budgets unpredictable for the buyer, which slows enterprise procurement, and it exposes you to their seasonality.
Fix the downside with a committed floor: a base fee that includes a volume allowance, then a rate above it.
Per outcome: charge for the result
Use it when the outcome is unambiguous, attributable to you, and measured in a system both sides trust: a recovered payment, a filled shift, a resolved ticket, a qualified meeting.
- The easiest thing in the world to sell, because the customer only pays when they win.
- Highest revenue capture when you genuinely drive the result.
- Attribution fights are brutal. If the customer can argue they would have got the outcome anyway, you are now in a debate instead of a renewal.
Only take this shape if you can measure the counterfactual. Otherwise you have signed up to be paid at the discretion of someone else's analyst.
Per seat charges for access, per usage charges for work done, per outcome charges for results delivered. Most durable models combine two: a platform fee for access plus a meter on the work. Tabs switch with arrow keys and all three panels are readable with scripting off.
Test your metric with four questions
Write your candidate value metric on a page and put it through these. A metric that fails any one of them will cost you more than a bad price level.
Does it grow when the customer wins?
If your customer has their best year ever and their bill does not move, your metric is disconnected from value and you have capped yourself at whatever you charged on day one.
Can the customer predict it?
A buyer who cannot estimate next quarter's bill within twenty percent will either not sign or will build a spreadsheet to police you. Publish a calculator, offer a cap, or include a generous allowance in the base.
Can you both measure it the same way?
The metric must live in a system both sides can see. "Hours saved" is a beautiful story and an ugly invoice, because it is computed from an assumption. "Documents processed" appears in a log.
Does it track your own cost?
If serving a customer costs you 30 cents per transaction, a flat monthly fee means your worst-margin customer is your heaviest user, and you will not notice until they are 40 percent of your volume.
The margin math nobody runs until it hurts
Software founders inherit an assumption from an earlier era: that serving one more customer is free. For anything touching inference, transcription, human review or third-party APIs, it is not. Run the numbers per unit, then decide.
Negative margin. Every unit you sell costs you money. Raise the price or cut the cost before you sell one more account. Under 40 percent. You are running a services business with software branding. Either automate the cost out or price for it honestly. 40 to 70 percent. The realistic band for products with real inference or review costs. Watch it monthly, per customer. Over 70 percent. Classic software economics. You can afford to buy growth, so the constraint is distribution, not margin.
Gross margin equals revenue minus variable cost minus the fixed cost of serving the account, all over revenue. At 40 per thousand units against 14 of cost, 30,000 units and 200 of fixed cost, the account runs at 48 percent margin and 580 of monthly gross profit. Push volume up and margin improves as the fixed cost spreads. Push price down 20 percent and watch how fast the whole thing inverts.
The third output is the one to keep. Volume needed to cover the fixed cost is your minimum viable customer, and it tells you exactly which deals to stop chasing. Most companies discover this number by accident, two years in, when someone finally allocates support hours per account.
Falling costs are an opportunity, not a discount
Inference costs have fallen steeply and will keep falling. The reflex is to pass it through immediately. Think about it for a week first.
If you priced on value, your customer is paying for a resolved ticket, not for tokens. A 60 percent cut in your unit cost is a margin expansion that funds your sales team, your support quality and your next product. If you priced on cost-plus, you just handed your entire improvement to the customer and taught them to expect an annual reduction.
The exception is competitive: when a rival passes savings through visibly and buyers can compare unit rates directly, you follow or you lose on the spreadsheet. Which is another argument for a metric your competitors do not share. If you charge per resolved case and they charge per million tokens, nobody can run that comparison at all.
Take your ten largest customers. In a spreadsheet, put what each pays today in one column. In the next, put what they would pay under a different value metric: if you charge per seat, price them per unit of work done; if you charge flat, price them per transaction. Use a rate that keeps your total revenue across the ten roughly the same. Now look at the differences per customer. You will find two things every time. A handful of customers are getting enormous value for a small fee, which is where your next price increase lives. And a couple are paying well above the value they receive, which is where your churn is coming from next quarter. That single sheet is worth more than a pricing consultant, and it takes an afternoon.
Structure beats cleverness
The model that survives contact with real buyers is usually the least exotic one: a platform fee plus a meter.
The platform fee covers your fixed cost to serve the account, gives the buyer a predictable line in their budget, and gives you revenue that does not evaporate when their volume dips in December. The meter captures growth without a renegotiation. Include an allowance in the base so small months feel fair, set the overage rate so that heavy users stay profitable, and publish both.
Three guardrails make it work in practice. Cap the first year's overage exposure for enterprise buyers, because procurement will demand it anyway and offering it first is worth a point of price. Alert customers before they cross a threshold, since a surprise invoice is the fastest route to a cancellation and a bad review. And never bill for units that came from your own errors, retries or timeouts, because the day a customer audits that and finds it, you lose the account and the reference.
How closely each shape ties your revenue to your customer's success. Alignment is not the only criterion: outcome pricing scores highest and carries the worst attribution risk. The practical winner for most companies is a base fee plus a per-unit meter, which sits in the aligned band without the measurement fights.
Changing the model on existing customers
Repricing an installed base is a trust exercise, not a billing exercise. Grandfather existing customers on their current model for a defined period, twelve months is standard, and put the end date in writing at the start. Move new customers to the new model immediately so you gather evidence. When you migrate the old base, show each customer their own numbers under both models, and cap the first year's increase.
Expect to lose some accounts. The ones you lose are usually the ones extracting the most value for the least money, which reads as painful and is usually a margin improvement. Measure the result in gross profit, not logo count, or you will talk yourself out of a good decision.
The takeaway
- The value metric decides who expands and whether your revenue grows with your customer's success.
- Per-seat pricing inverts when your product removes seats. Charge for work done instead.
- Measure gross margin per account with variable costs included. AI-heavy products run well below classic software margins.
- Falling inference costs are margin expansion, unless you priced cost-plus and taught buyers to expect a rebate.
- Base fee plus meter, with an allowance, alerts and a cap, beats every clever structure in practice.
Frequently asked questions
What is a value metric in pricing?
The unit you charge for, chosen so the bill grows as the customer's value grows. Good ones are easy to understand, easy to measure in a shared system, and rise with the customer's own success: documents processed, tickets resolved, orders shipped. Bad ones let value grow without payment, or punish usage.
Why is per-seat pricing a problem for AI products?
Per-seat assumes value scales with the number of people doing the work. A product that does the work needs fewer people, so performing well shrinks your own invoice. Products that automate should charge for work completed, such as cases handled or documents processed.
What gross margin should a software business target?
Classic software runs 60 to 80 percent. Andreessen Horowitz's analysis of AI businesses found materially lower, often 50 to 60 percent, because inference, review and cloud costs scale with usage. If you have per-transaction costs, measure margin per transaction and price with that cost visible.
Should I use a platform fee plus usage, or pure usage pricing?
Base fee plus usage is usually stronger for a small company. The base covers your fixed cost to serve and steadies revenue; the meter captures growth. Pure usage exposes you to your customers' seasonality and makes forecasting hard, which is dangerous on a thin cash buffer.
The pricing model was wrong for three years. Nobody said anything.
Kill My Startup examines the decisions that quietly cap a company long before the market does.
Buy on Amazon →Sources
- Andreessen Horowitz, "The New Business of AI and How It's Different from Traditional Software," on gross margin structure.
- Stanford HAI, AI Index Report, on the fall in inference cost for GPT-3.5-level performance between November 2022 and October 2024.
- Price Intelligently / ProfitWell research on value metrics and pricing structure.
- Standard software finance definitions of gross margin, cost of goods sold and committed-use contracts.