Your average customer costs $25 a month to serve. Your median customer costs 25 cents. Those are real numbers from an AI product I launched pricing for — same plan, same price, a hundred times apart underneath.
Almost every argument I get pulled into about AI pricing — tokens, credits, outcomes, agents — is downstream of that one gap.
For twenty years of SaaS you could underwrite a deal off the org chart. More people, more seats, more cost, more revenue, and the ratio held well enough that you could price on value and never look at COGS again.
In AI, cost tracks behavior. Behavior has nothing to do with the size of the company writing the check.
The logo won't tell you. Neither will the seat count. A twelve-person startup can be your single largest cost line. An enterprise with four thousand seats can cost you almost nothing. I've watched both happen in the same quarter, in the same product.
Then the use case shifts and the whole distribution moves under you. Summarization and multi-step agent runs are two different businesses sharing one price list.
You're not pricing a product. You're pricing a distribution you can't see.
Every generous cap you ever wrote was safe for a reason nobody said out loud. A person gets tired.
They click twice a second, get bored, go to lunch, take a holiday, and come back Monday having used a fraction of what you were braced for. Your "unlimited" plan was never unlimited. It was capped by attention span, and biology enforced it for free.
Call it the attention-span subsidy. For a decade your margin was funded by human fatigue and it never once showed up on the P&L.
Now point an agent at the same endpoint. It runs at 3am on a Sunday and it doesn't get bored of your API.
Every soft limit you were quietly relying on gets found and flattened by Tuesday.
Compute gets cheaper every quarter. It's the one prediction in this industry I'd bet real money on.
So when you price off your cost per call, you've promised to hand every efficiency your engineers ship straight to your customer. Not in a negotiation, not in a renewal you can reopen. In the structure itself.
That's the efficiency giveaway. Your team does the work, your customer gets the savings, and nobody ever held a meeting about it.
Cost tells you when to walk away from a deal. It has never once told you what to charge.
The tell shows up in the first fifteen minutes of most AI pricing calls I take.
The price came from nowhere. Someone looked at a competitor, looked at the compute bill, split the difference and shipped it. No customer conversation, no willingness-to-pay work, no test.
Credits are where this hides best. A credit sounds like a unit, but it's a placeholder for one. It buys you time while you work out what the customer is actually paying for. Skip that work and you've issued a currency with no exchange rate. You know what a credit costs you. Your buyer has no idea what it buys them, so every renewal turns into a negotiation about what the money was worth in the first place.
Vibes pricing survives longer in AI than it does anywhere else, because there's no competitive anchor to embarrass you. Nobody can prove you're wrong, which is a much weaker position than it feels like from the inside.
The instinct when your cost is chaotic is to make the price simple. One number, flat fee, don't make the buyer think about it.
Wrong lever. Customers will pay a premium for predictability and transparency. Simple is not what they're buying.
Say three things and put them in writing: what you charge for, how you charge for it, and when the charge moves. A buyer who can answer all three will accept real structure: tiers, thresholds, overages, fair use lines, whatever the product needs. A buyer who can't answer them asks for a spend cap instead.
That's the worst outcome on the board. Your customer buys less than the value they'd get, and your growth is now capped at whatever number made someone nervous in a room you weren't in.
Nobody asks for a spending limit on a thing they can count.
You don't invent the unit. You go find it.
A design tool I worked with was deciding whether to meter by pages or by projects. Pages was the tidy answer, because pages are what the software makes and the database already counts them.
Their customers were agencies. An agency doesn't sell pages. It invoices per client, and one project is roughly one client.
Meter projects and your price lands in a column where money is already moving. Every new client they win is a bigger bill they're glad to pay. Meter pages and you're an expense with nothing on the other side of their ledger to justify it.
Most of us pick a unit off a database schema, then spend two quarters running research to find a number the customer was already using.
Go read an invoice your customer sends someone else. It's faster than the study, and it's free.
Grafana Labs. $450M+ ARR. New AI product, no competitive anchor to price against. Average cost per user around $25, median 25 cents. Self-serve usage looked nothing like contracted usage. Every condition I've just described as hard, all at once.
I did not solve the unit problem first. I shipped a hybrid seat model.
Seats, with usage handled underneath, and iteration built into the plan from day one. Not because seats are the right long-term answer for software that thinks. Because a seat was a number the buyer could already forecast, and we needed deals more than we needed a perfect meter.
I'd push back on that model in a client engagement today. I'd also ship it again under those same conditions, and I've stopped feeling conflicted about that.
The work that made it hold wasn't the model anyway. It was one conversation with the CEO, the CFO and the CTO before launch: how much burn do we tolerate, and for how long?
That's a decision, not a spreadsheet. Once those three agreed on the number and the window, pricing stopped being a debate and became a plan. Then I got into the deals to close them.
What I'd tell a founder sitting where I was: don't overindex on margin or the mechanics of the model this early. Set a logo target. 25, 50, whatever fits your stage. Do what it takes to close them and make it repeatable.
You will not learn the right unit from a whiteboard. You learn it from a hundred sales conversations, and the only way to get those is to sell something.
Margin discipline comes after, and it has its own tool. Fair use is how you protect the floor without touching the 80% who were never the problem — I wrote the build-it-yourself version separately.
Usage by account. Median, P95, P99. If your mean is nowhere near your median, you don't have a pricing problem. You have a concentration problem, and it needs its own lever.
One meeting, three people, two numbers: how much are we willing to lose per customer, and for how many quarters. Write it down.
What was already a line in their business before you existed? Incidents handled. Contracts reviewed. Clients served.
When usage is genuinely unpredictable, stop guessing at the annual number. Run a paid pilot, watch 60 to 90 days of real usage, and size the 12-month contract off what you see. You stop negotiating against a hypothetical, and so do they.
Not in a deck — in the contract and on the pricing page, where the buyer can find it without asking you. Transparency is a feature, so charge like it.
In that order.
The number you can see is your cost, and it's the least useful one on the table. The number that decides your business is sitting in your customer's budget, already allocated to the job your product does.
You don't find it by thinking about it. You find it by walking toward it until somebody flinches.
30 minutes, no pitch — just an honest look at your cost distribution, your unit, and what you can actually defend in a budget meeting.
Book a free call →