When you sit down to price an AI product, the argument lands on the metric almost immediately.

Per seat? Per message? Per agent? Per resolved ticket?

It's a fair thing to argue about, and it does matter - but it's the second question you should be asking, not the first.

The first question is this: who carries the cost uncertainty?

With classic software, serving one more user costs you close to nothing, so you could afford to be a little careless here and still come out fine.

But AI has taken that cushion away.

Every time a customer hits the button, you pay a real bill. And at this stage neither of you knows how hard they're going to lean on it.

It's tempting to assume falling token prices will rescue you here. They won't.

Yes, token costs have dropped about 97% in a single stretch - OpenAI went from something like $5 per million input tokens to around 15 cents - but it settled nothing, because the models got better at the same time.

People started asking them harder questions, and the token requirement climbed right back up.

So the price per unit fell while the number of units multiplied to meet it. The ground under your model is still moving - and engineering a precise, permanent price on top of moving ground is like playing three-dimensional chess against an opponent who hasn't sat down at the board yet.

That leaves you two honest options. The whole game is choosing between them deliberately, instead of drifting into one by accident:

• You carry the uncertainty. Flat fee or license, unlimited usage, a price the customer can plan around. Wonderful for adoption - until a power user discovers everything your product can do and takes a knife to your margin. Go this way and you need an off switch you can reach: fair-usage limits, a hard cap, a tripwire on consumption. Never ship the unlimited tap without one.
• The customer carries it. Pure usage - meter everything, invoice at the end. Your margin is protected, but now they're holding a bill they can't forecast, and finance hates an unpredictable bill more than a large one. So they ration it: go easy on the tool, adoption stalls, and everyone admires the clean unit economics while the product goes unused. The meter that guards your margin is the one strangling your growth.

This is why credits have taken over so much of the AI conversation.

A credit model is just usage-based pricing paid upfront: buy a block, spend it down, roll the unused balance into the next period. That roll-over is the whole point - take it away and you've rebuilt a license.

You get cash early, the customer still feels like they only pay for what they use, and the risk of under-usage is shared instead of dumped on one side.

I've designed credit systems since before AI made them fashionable, and for anything complex they're still the best tool we have. Your CFO won't love the revenue recognition - but it's a headache worth taking on.

The catch is that you don't actually know how your AI product gets used yet. Almost nobody does - it's too early. So don't lock in a permanent model. Build a bridge.

Take three to five existing customers and put them on an unlimited license - flat fee, no meter, an off switch you control. See this as a low-risk bet: worst case it costs you a little, and you can always wind the arrangement down later.

Here, you're not buying revenue, you're buying data. You're looking for how the product gets used when nobody's punished for every click, what the real value metric is, and whether the thing blocking adoption is your price or your product.

Then you design the real model around what you learned instead of what you guessed. That bridge buys you a year or two - long enough to get the real model right, instead of betting on a guess today.

There's one last move I would recommend, too:

When you land on the real model, charge as close to the end of the value chain as you can measure and invoice.

Picture the customer's process as a rainbow. There's a team, some data, a few models - and then the gold at the far end: the credit scores, the resolved tickets, the closed deals.

The gold is what they actually want. They'll happily pay for the inputs along the way - the seats, the setup, the onboarding - but only as the toll to get there.

Charge at the very start and you become a cost they'll try to minimize. They'll sit through the whole project before they feel any return, which on an enterprise timeline is months before they're ready to expand.

Charge closer to the gold and two things happen at once. The ROI math gets easier, because your fee comes out of the value created rather than piling onto the cost. And the gap between "they pay you" and "they get paid back" shrinks, so expansion arrives sooner.

So don't tax the beginning of the process. Tax the end.

The whole process:

• Decide who carries the cost uncertainty - either you (a flat fee or license, your margin at risk) or the customer (usage, adoption at risk). Don't drift into the answer and find it on a margin report three quarters from now.
• If you don't know your usage curve, you're probably not ready for the permanent model - and most teams don't, which is fine.
• Run a license bridge on three to five customers - unlimited usage, a flat fee, an off switch you control. You're buying data, not revenue.
• Default to credits for the real model - cash now, risk shared, the customer still feels like they only pay for what they use. Set a sensible expiry so you're not carrying the liability forever.
• Charge as close to the gold as you can measure and explain - but no closer. If you can't count it and put it on an invoice, you can't price on it.
• Build the model so it's allowed to be wrong in two years - because it will be. Make sure your contracts give you room to change it.

One idea...

The only four ways to charge - and what each one does to your cash and your risk.

Everyone argues about what to charge for. Far fewer stop to think about how - and the "how" is what I call the modality.

Whatever you decide, it takes one of four shapes:

1. Flat fee. One price, no meter. Predictable on both sides, but the usage risk is all yours, and it doesn't expand - the customer grows while your revenue sits still. Best as a base layer, not the whole model.
2. License. The customer buys a set volume - ten seats, a million API calls - and pays upfront, used or not. Cash comes early, the price is predictable, and the under-usage risk sits with them. The catch: it's a harder sale, and the under-adopter shows up at renewal wanting a discount.
3. Usage-based. The mirror image. Access now, meter what they use, invoice at the end. What they pay matches what they get - but the cash arrives late, the bill is unpredictable, and the under-usage risk flips back to you.
4. Credits. The hybrid: usage paid upfront, the unused balance rolling over. Cash now like a license, "pay for what you use" like usage, risk split down the middle. More to explain, and you need expiry terms - but for anything complex or AI-shaped, nothing else does this much work.

In practice you almost never use just one. The B2B models I work on usually stack two or three - a flat fee for the account, a license for the seats, usage for the infrastructure - and each of those is its own call on cash and risk. Read more?

Willingness To Pay News

We roasted some pricing at SaaSiest - and the guidebook is coming.

At SaaSiest 2026 in Malmö, I gave a keynote with the diplomatic title "How to F*ck Up Your Pricing" - a step-by-step guide to ruining your pricing, role by role. (Turns out everyone in the C-suite gets to help.)

The CPO makes it technical and sells features instead of outcomes. The CRO, CFO and CEO each get their own way to break it - and yes, the CEO is the f*ck-up of last resort. The serious point under the jokes: pricing isn't one person's job. It's a value stream, and any single link can break it.

The talk is becoming a short guidebook - "How to F*ck Up Your Pricing" - out very soon. Watch this section.

Our Podcast "Pricing Page unPacked"

Episode 5: Intercom's Fin AI agent.

Fin was stuck on a seat-based model growing around 4% a year, against a SaaS median of 19%. The team rebuilt it around a single metric - the resolved conversation, at $0.99 - so the customer pays for the result, not the tool. Fin has since passed $100M ARR.

Rob and I take the pricing page apart, including where it's still exposed: the price war coming for $0.99, and the fact that a resolved ticket is worth far more to a bank than to a gaming app.

The fifth teardown, after Figma, Notion, Slack and HubSpot.

Listen here

This is the last episode of Season 1, we will be back in August with Season 2.

Spotify | Youtube | Apple | Amazon

PS: If you do one thing this month, pull your three highest-usage AI accounts and look at what they're doing, not what they're paying. The gap between those two numbers is your real pricing model trying to get your attention.