Pricing is the feature you forgot to ship

For the first time, your product's success isn't moving your revenue in the same direction. The more people use the thing you built, the worse the numbers look.

Here's why. In traditional SaaS, one more user costs you close to nothing to serve. Everyone costs roughly the same. That single assumption is what makes per-seat pricing work — it's why nobody ever had to think about it.

AI breaks it. Two users on the same plan, paying the same price, can cost you wildly different amounts. One runs a few short prompts a week and costs you fifty cents. Another runs long agentic workflows all day and costs you five hundred dollars. Your revenue line stayed flat. Your cost line didn't.

So the gap you're seeing isn't a billing bug. It's the pricing model doing exactly what it was designed to do, on a product that no longer matches the assumption it was built on. At best, you've opened a gap between the value you deliver and the money you capture. At worst, every interaction with that feature now costs more to serve than it earns — and your own success is draining your revenue.

That's the trap. Not that the feature failed — that it worked, on a pricing model that can't survive its own success.

But there's another way to read that same situation. The pricing model that's now exposed was never a law of nature. Someone chose it. And the moment a model stops fitting reality is the moment a market is open to be disrupted — not by adding more feature capability, but by changing the way you charge.

And almost nobody is treating it that way yet.

The "obvious" model was once a bet

Remember when you paid for a movie one at a time — plus a fine if you returned it late?

Netflix changed nothing about the product. Same DVDs, same envelopes. It just changed the unit it charged for — from "a rental" to "a month of unlimited access." Blockbuster kept the old unit. We know how that ended.

The model we now treat as obvious was a contested bet at the time. Someone had to choose the unit — and choosing it well was the whole game.

The same decision is being fought over right now — and nobody's won yet

That bet is back on the table. AI broke the assumption underneath SaaS pricing, and the entire industry is now experimenting in public to find the new unit.

The tell that it's unsolved: the best companies in the world are arriving at completely different answers.

Intercom charges per outcome. Its AI agent, Fin, bills $0.99 per outcome — and an outcome only counts when Fin actually delivers one: either it resolves the customer's issue, or it successfully completes a workflow that hands off to a human. You don't pay for attempts that go nowhere. Their stated principle is the interesting part: pricing should track value delivered, and the vendor should carry the risk when the product doesn't perform. The unit isn't a seat or a message. It's a result.

Notion changed its model twice in roughly a year. It started selling AI as a separate per-seat add-on, then in 2025 killed the add-on and bundled AI into its core paid tiers, then layered a usage-based credit system on top for heavier agent workflows. Add-on, then bundled, then bundled-plus-consumption — three structures in quick succession, on the same product.

Zendesk had to legally define a word. It bills per "automated resolution" — but to do that, it had to define exactly what counts as resolved: a ticket the AI closes without human help, confirmed after 72 hours of no further activity. Read that again. They had to write a 72-hour rule to make the word "resolved" billable. That's how hard defining the unit actually is.

And the reversals tell you nobody's confident: Slack and Loom both launched AI as a per-user add-on, then later folded it into their core pricing — the same retreat Notion made.

Four serious companies. Per-outcome, bundled, per-defined-resolution, add-ons-then-not. Multiple public reversals. No consensus winner.

Nobody has agreed yet — which is precisely why this is a live product problem and not a settled best practice you can copy.

So what can you do about it?

For most of SaaS history, pricing landed on a PM's desk as a given. Finance or the founders set the model; you built, sometimes tested, and shipped it. The model was the weather — something you worked under, not something you touched.

That's over. When the unit of value is unsettled, choosing it is a product decision — and it belongs to the person closest to how customers actually get value. That's you.

So here's the work. Three steps, in order.

Step 1: Run the anchor test before anything else

Ask one question: if a person did this task manually, what would that cost per month?

That number is your ceiling, not your token cost. Intercom did this out loud — a human support agent costs a company somewhere between $5 and $15 to resolve one ticket. Fin charges $0.99 per resolution. Roughly a tenth of the human cost, and still a healthy margin.

Most AI features should land between 10% and 50% of the human cost of the same work. If you can't name that human cost, you don't yet know what you're selling — and no pricing model will fix that.

Step 2: Pressure-test your candidate unit with four questions

You probably already have a hunch about the unit. A resolution. A generated report. A processed document. Before you build anything around it, put it through four checks. A unit that fails any one of them will break in production.

Consistent — does one unit mean roughly the same amount of work every time? If "one report" sometimes means three paragraphs and sometimes means forty pages, you're charging one price for two products.

Attributable — did the AI actually produce the result, or did a human finish it? This is the trap most teams walk into. If a person still reviews, edits, or clicks send, the customer is still paying for that person's time, and they will never agree to pay you for the outcome as well. If your AI gets the job 90% of the way there, you don't have an outcome to sell. You have an assist.

Measurable — can you define the unit tightly enough to put it on an invoice? Zendesk bills per automated resolution, and to make that word billable they had to write a rule: the AI closes the ticket without human help, confirmed after 72 hours with no further activity. They needed a 72-hour rule to make "resolved" mean something in a contract. That's the bar.

Predictable — can the customer forecast the bill before the month starts? A finance team that can't budget for you will cap you, and a capped account stops growing.

A unit that fails one of these four isn't a pricing problem you can solve with a better price point. It's the wrong unit.

Step 3: Ship the structure, then change it

Notion changed its AI structure twice in about a year — separate per-seat add-on, then bundled into paid tiers, then a usage-based credit layer on top for heavier agent workflows. Same product, three structures. That's not indecision. That's a team treating the model as a thing you iterate.

The version most teams land on is a hybrid: a flat base fee that gives finance a predictable number, plus a usage layer for the expensive AI work. Include some volume in every tier so people actually try the feature — the moment a user feels a meter running on every click, they stop exploring, and they never reach the point where the value is obvious.

Then watch one number: gross margin per interaction. Not per account, per interaction. If your top 5% of users can wipe out the margin from the other 95%, the model is broken, and it will keep looking fine on the dashboard right up until it doesn't.

Your next real release might not be a feature. It might be the way you charge for the ones you already shipped.

What's the most interesting pricing-model shift you've seen a company pull off lately? Always keen to collect good examples.

Next time I'll get specific about the harder half of this: how to actually test a pricing model without blowing up your revenue or your customer trust. If that's useful to you, follow along — that one's coming.

This is one piece of a longer framework I teach in Chapter 5 of Product Strategy in the AI Era — including the full unit-economics model, tier architecture, and how to structure hybrid pricing so a power user can't wipe out your margin.