Pricing an AI feature
Price an AI feature per seat, per use or in credits, show the margin from the token bill, and pull four levers in order from a 10% margin to 70%.
On this page
An AI feature has a cost every time it is used, which most software features do not, so its price is a margin decision before it is a marketing one. The drafting add-on sells at $5 a seat to 12,000 seats, $60,000 a month, against the $54,000 token bill from module 4: a 10% gross margin on a feature that works. Four levers, pulled in order, take it to 70%, and the choice between charging per seat, per use or in credits decides who pays for the heaviest users.
The mistake: a price with no cost line
The price was set the way add-on prices usually are: the help desk's other add-ons sell at $5 a seat, so this one did too. Nobody wrote the cost line under it.
Gross margin is revenue minus the cost to serve, divided by revenue. The cost to serve is the token bill, priced at module 4's assumed $3 and $15 per million tokens; servers, retrieval and the gateway round to zero beside it, as reading an architecture showed. So ($60,000 − $54,000) ÷ $60,000 = 10%: per seat, $5 of revenue against $4.50 of tokens. Assume finance holds every add-on to a 70% gross margin; substitute your own company's line.
And 10% is an average. A draft costs $0.009 as specified, and drafts per seat are not equal. A seat drafting 10 a day costs 10 × $0.009 × 30 = $2.70 a month, a 46% margin. A seat drafting 60 a day costs $16.20, more than three times its price. A per-seat price is set on the mean and paid for by the tail; per seat or per use works the whole distribution.
Four levers, in order
Seventy percent margin means cost at most 30% of revenue: $18,000 a month against $60,000. Three levers are the token levers from tokens as a cost model, in the order that lesson pulls them, and the fourth is the price.
- Cap output on short tickets. 70% of tickets need a reply under 150 tokens. The bill falls to $44,550 and the margin to ($60,000 − $44,550) ÷ $60,000 = 26%. It costs a line in the prompt.
- Route the short tickets to a small model at a tenth of the price, once the eval puts it within 2 points on that slice. The decision doc carries $27,500, which keeps the routed output at the mid-tier price as a margin for escalations. Margin: 54%. It costs an eval run, a router, and a re-run on every model update.
- Cache the fixed 1,100-token prefix. Assume a cache read at a tenth of the input price. On the 60,000 drafts a day still on the mid tier, input falls from 60,000 × 1,500 × $3 ÷ 1,000,000 = $270 a day to about $92, roughly $5,300 a month; call it $5,000. Bill $22,500, margin 62.5%. Not counted until the vendor's page has been read and the engineers confirm the prefix is byte-identical on every call.
- Price the seat at $6.25. At $22,500 of cost, 70% needs revenue of $22,500 ÷ 0.3 = $75,000, which is $6.25 across 12,000 seats. Without lever 3 it needs $27,500 ÷ 0.3 = $91,700, $7.64 a seat.
The order is the point. Pricing first would need $54,000 ÷ 0.3 = $180,000 of revenue, $15 a seat, which asks the buyer to pay for tokens the product is wasting. Each cost lever has an eval behind it; the price lever has a customer behind it, and should be the smallest one left. It also has to survive the value check: in positioning, one minute on 40 replies a day is worth about $440 a seat a month at an assumed $30 an agent-hour, so $6.25 is about 1.4% of it.
Per seat, per use, or credits
The same drafts can be charged three ways, and each moves a different number.
Per seat. The buyer's bill is fixed and approved once a year. The margin falls as use rises, and the lightest seats subsidise the heaviest. It works when usage is narrow.
Per use. A price per draft. At 70% margin on the as-specified cost, $0.009 ÷ 0.3 = $0.03 a draft; after three levers, $22,500 ÷ 6,000,000 drafts a month = $0.00375, so $0.0125 a draft. The margin is fixed per draft, and the buyer's bill moves every month. Every draft now has a visible price, so admins ration it, and the drafts they ration first are the short tickets that routing made cheapest. Per use puts a price on the behaviour the North Star counts.
Credits. A seat includes an allowance of drafts; blocks above it are sold at a per-draft rate. The bill is fixed up to the allowance, the margin is protected above it, and only the heaviest users ever see a price on use. A fair-use cap without a block price is the same idea enforced as a rate limit, which protects the margin and annoys exactly the customers the product fits best.
What decides between them is the shape of usage. 200,000 drafts across 12,000 seats is about 17 a seat a day; if the median seat drafts 10, heavy seats are pulling the mean up, and a seat price set on the mean overcharges the median and undercharges the tail. The wider that gap, the more an allowance earns its complexity.
The price is a formula
Model prices change and a seat price signed for a year does not. Write the price in the decision doc as a formula — cost per draft, drafts per seat at the median and the 95th percentile, target margin — so a vendor price change is one recalculation, not a new argument. Margin also feeds payback: CAC, LTV and payback multiplies revenue by it, so a 10% add-on barely repays its own acquisition.
In an interview
"How would you price this AI feature?" is an estimation question with a decision at the end. What is being tested is whether you start from the cost per unit, compute the margin at the mean and say what happens in the tail, order the levers by what they cost to pull, and pick a charging model by who it makes pay. "$0.009 a draft, 17 drafts a seat a day, $4.50 against a $5 price: a 10% margin, and negative for the heaviest seats" is the opening sentence.
The mistake that loses points is pricing from the competitor's list with no cost line, then quoting one margin as if every seat cost the same. The second is reaching for the price before the cost levers, which charges the customer for waste.
Check yourself
1. Finance wants 70% this quarter at $5 a seat, with no price change. Can the levers get there?
70% at $60,000 means a bill of at most $18,000. Levers 1 to 3 reach $22,500, and caching the prefix on the routed 70% as well saves another 140,000 × 1,100 × ($0.30 − $0.03) ÷ 1,000,000 ≈ $42 a day, about $1,200 a month: $21,300. Still $3,300 short. Not with the levers costed; offer an allowance with credits, or $6.25 a seat, and say which further cost lever, with its eval, would close the gap.
2. A new tenant brings 3,000 seats at 40 drafts a seat a day and asks for the $5 list price. Sign it?
120,000 drafts a day costs about $16,500 a month after levers 1 and 2, against $15,000 of revenue: a negative margin before lever 3, and about 10% after it. Not at list. At the routed cost of about $0.0046 a draft, 70% needs about 1.5 cents a draft; 3.6 million drafts a month is about $55,000, roughly $18 a seat. Offer seats with an allowance and a committed block of credits priced to that.
3. The vendor halves its input price next quarter. Do you cut the seat price?
After levers 1 and 2, input is about $10,000 of the $27,500 bill (60,000 × 1,500 × $3 plus 140,000 × 1,500 × $0.30, per million, about $333 a day). Half of that saves $5,000 a month and lifts the margin from 54% to about 62%, still under 70%. Keep the price; the cut is margin, not a discount. Revisit when the bill is under $18,000.