Pricing AI Products: The Margin Moved and Nobody Repriced
If you charge per seat for software that replaces seats, your best outcome is your worst commercial result.
Summary
- Classic software runs at 75 to 85% gross margin because serving one more customer costs almost nothing. AI-native products commonly sit between 40 and 65%, because every request costs money.
- Per-seat pricing contradicts the product's own promise: sell something that does more work with fewer people, charge per person, and your revenue falls exactly when you deliver most.
- The structure that works is a platform fee, an included allowance of a usage-linked unit, a published rate beyond it, and an annual commitment at a better rate.
Software pricing has been built on one quiet assumption for twenty years: serving one more customer costs you almost nothing.
That assumption is what made classic software businesses work. Gross margins of 75 to 85%. Free trials that cost nothing to run. Unlimited plans that were safe to offer, because usage barely moved your bill. Almost every pricing habit the software industry has came from that world.
AI-native products don't live in that world. Every request costs you money. Heavy users cost you a lot of money. And most of the pricing habits inherited from classic software quietly stop working.
Plenty of Indian AI companies we meet are running on those inherited habits, and haven't yet noticed what it's doing to them.
The number that changed
A classic software business runs at 75 to 85% gross margin. Hosting, storage, a few third-party services: real costs, but small relative to revenue and largely fixed.
AI-native products frequently land between 40 and 65%. Sometimes lower in the early months before anyone has optimised anything.
That gap isn't a detail. It changes four things at once.
What you can spend to acquire a customer. At 80% margin you can spend heavily and recover it. At 50% you need roughly twice the revenue to recover the same spend, over the same period.
What sales model you can afford. A field sales team is expensive. Some margins support one and some don't.
Whether free trials and free tiers make sense. A free user in classic software costs you almost nothing. A free user on an AI product costs you real money every time they use it, and the heaviest free users cost the most.
What multiple you get valued at. Investors know the difference between a 78% margin business and a 52% one, and they price it.
None of that means AI businesses are bad businesses. It means the playbook has to change, and the pricing is where it changes first. It also means your real floor needs computing properly rather than assuming.
The per-seat problem, stated plainly
Here's the contradiction sitting inside most AI product pricing.
You sell something that lets a customer do more work with fewer people. That's the promise. It's what gets you in the room.
And then you charge per person.
When the product works exactly as intended, your revenue stays flat or falls. You delivered more value and earned less for it.
So when the customer handles twice the volume with the same team, or the same volume with half the team, your revenue does not move with them. Meanwhile your costs went up, because all that extra work ran through your models.
More value delivered. Less revenue earned. Higher cost to serve. That's not a pricing problem you can fix with a discount policy, it's the shape of the whole model.
What your price should move with
The rule is simple to state and uncomfortable to act on.
Whatever drives your cost has to be visible in your price.
Your cost moves with how much the product is used. So your price has to move with something that tracks usage, not perfectly, but directionally.
That doesn't mean billing raw tokens. Customers hate that, they can't forecast it, and it exposes them to changes in your model costs that have nothing to do with them. It means finding a unit that sits between the two: something the customer understands and values, which also happens to correlate with what you spend. That choice, what you charge per, is the hardest pricing decision to reverse.
Depending on what you do, that might be documents processed, conversations handled, tickets resolved, hours of audio, reports generated, cases reviewed, or records enriched.
The test is whether a customer can look at that unit and say yes, that's what I'm getting from this.
The structure that usually works
Almost every AI product we look at ends up in the same shape, for good reasons.
A platform fee that covers your fixed cost of serving that account, support, success, integration, and gives the customer a predictable base they can budget for.
An included allowance of whatever your unit is. Generous enough to feel fair, sized around what a normal customer actually uses.
A rate beyond the allowance, clearly published.
An annual commitment at a better unit rate, which gives you forecastable revenue and gives them a discount they can justify internally.
This structure solves the tension that breaks pure usage pricing in Indian enterprise. Finance teams need to know what they'll spend. A pure usage bill is unpredictable, and unpredictable gets treated as a risk by procurement, which slows deals and shrinks them. The platform fee and the allowance give them a number they can put in a budget, while the overage protects you from the customer whose usage triples.
Things to build in from the start
Fair use is not a policy, it's a number. Unlimited is a promise you will regret. If you must say unlimited for positioning reasons, define the fair use limit clearly and enforce it quietly.
Check margin per customer, not on average. Rank your customers by how much they use. If your heaviest users are your thinnest margins, you have a structural problem that will get worse as you grow. This is a monthly check, not an annual one, and it is the same discipline net revenue retention depends on.
Price for the model you run today, not the one you hope to run. Model costs have fallen and may keep falling. If they do, that's upside. Pricing on the assumption that they will is how companies end up with a base of customers they can't serve profitably.
Keep the right to adjust. Build into your contracts that included allowances and overage rates can be revised at renewal with notice. Not to punish anyone, but to make sure a pricing decision made in 2026 doesn't govern a cost base from 2029.
Outcome pricing: the tempting one
There's a lot of enthusiasm right now for charging per outcome: per ticket resolved, per rupee recovered, per hire made.
The appeal is obvious. It aligns you with the customer, it removes their risk, and it's a strong way for an unknown vendor to get in the door.
Two conditions have to hold, though, and they're harder than they look.
The outcome has to be measurable without argument. If there's any room to dispute whether it happened, every invoice becomes a negotiation, and you'll spend more on collections than you gained on price.
The starting point has to be agreed in writing, before you go live. Without a baseline both sides signed, your customer will attribute the improvement to their own work, and they will often be partly right.
Where both conditions hold, outcome pricing is excellent. Where they don't, it's a slow-motion argument you've contracted yourself into. Our general advice: use it as an optional layer on top of a normal structure, not as the whole model.
For Indian AI companies specifically
Two things worth naming.
Selling globally versus selling domestically. If you sell the same product in India and abroad, your costs are the same in both places. Model costs are in dollars and don't care where the customer is. That means a heavy India discount eats your margin much faster than it would for a classic software company, and there's less room to absorb it. If you discount domestically, know exactly what it costs you and put a real fence around it.
Free tiers need more thought here. A generous free tier in classic software is a marketing expense with a small bill attached. In an AI product it's a real, growing cost, and the users who consume the most are the least likely to convert. Either cap it tightly or use a time-limited trial instead.
Where to start
Take last month. Work out your cost per customer, every model call, every API, every bit of compute, and put it next to what each of them paid you.
Then sort that list by usage.
If the customers at the top of the usage list are at the bottom of the margin list, you have found the thing that will limit this business, and you've found it while it's still cheap to fix.
Our Pricing Maturity Assessment covers what you charge per, your cost floor, and the rest of the layers, so you can see where the gap actually is before you change anything.
Frequently asked questions
What gross margin should an AI product have?
Lower than classic software. Where traditional SaaS runs at 75 to 85%, AI-native products commonly sit between 40 and 65% because model usage is a real, variable cost. The number matters less than whether it improves as you scale. If it gets worse with growth, your pricing structure is wrong.
Should I charge per seat or per usage for an AI product?
Usually not per seat alone. If your product lets customers do more with fewer people, per-seat pricing means you earn less exactly when you deliver more. Most AI products end up on a platform fee plus an included allowance of a usage-based unit, with a published rate beyond it.
Is outcome-based pricing a good idea for AI?
Only when two things are true: the outcome can be measured without dispute, and the starting baseline is agreed in writing before you go live. Where both hold it's a powerful way to sell. Where they don't, every invoice becomes an argument. Treat it as an optional layer rather than your whole model.
How useful was this article?
One tap. It tells us what to write more of.
The next one
Get what we publish next, by email.
Working notes on raising, borrowing, protecting, growing and structuring capital in India. One email a week at most, and you can leave any time.
We use your address only to send this. See our privacy policy.
We store your address to send you these emails and nothing else. See our privacy policy.
Related reading
