AI Product Pricing: Why Per-Seat Is Dead

Early this year the financial press coined a word for a market event it had never needed before: the SaaSpocalypse. In a single 48-hour window, roughly 285 billion dollars of software company value evaporated. Not because the products stopped working. Because investors finally did the math on what AI agents do to a pricing model built on counting human logins.

Here is the number that math turns on. Per-seat pricing, the model almost every software company has run on for twenty years, fell from about 21 percent of SaaS to about 15 percent in twelve months. That is not a soft decline. That is a model being repriced in real time, and the repricing is not finished.

The reason is simple enough to state in one line, and uncomfortable enough that most founders look away from it. You were never really selling a seat. You were selling the value a person produced while sitting in that seat, and the seat was just an easy thing to count. AI severs those two. When one agent closes a thousand support tickets without ever logging in, the login you used to bill for is now attached to almost none of the value being delivered.

So the question every builder shipping AI now has to answer is not whether per-seat dies. It is what the billable unit becomes when the seat stops standing in for value, and how you move your price there without burning down the revenue you already have. I have re-priced products through exactly this shift, and I have watched founders copy the hot new pricing model off a competitor’s landing page and quietly destroy their margin doing it. This is the field guide I wish I had handed them first.

Why the seat was always a proxy

Per-seat pricing worked for twenty years because of a quiet assumption that was true the whole time and is true no longer. The assumption was that work scales with people. If a company wanted twice the output from your software, it needed roughly twice the humans using it, so charging per human tracked the value the customer received. The seat was a proxy, and it was a good one, because headcount and output moved together.

Agents break the link. The whole point of an AI agent is to produce output that used to need a person, without needing the person. A support team that handled a thousand tickets with twenty agents now handles ten thousand with the same twenty humans supervising software that never logs in. The value the customer gets went up tenfold. The seat count did not move. If your price is stapled to seats, you just gave away ninety percent of the value you created, and you taught your smartest customers that the way to save money on your product is to use it less, by humans, and more, by agents.

This is the trap underneath the trap. Per-seat does not just leave money on the table in the AI era. It inverts the incentive. Your best customers, the ones automating hardest, are the ones who pay you least relative to the value they extract. The customer doing the most with your product becomes your worst account. No pricing model survives that for long, which is why the market repriced it so violently.

And the seat was never the only proxy you could have picked. It was the easiest one to meter in a world before AI, where counting people was cheap and counting work was hard. That world is gone. We can now meter the actual work, the actual actions, and in some cases the actual outcomes, cleanly and in real time. Once you can count the thing the customer actually buys, billing for a stand-in becomes a choice, and a bad one. If you want the wider menu of models this opens up, I wrote the full set in Revenue Models for AI Products. This piece is about the one move underneath all of them: getting your billable unit closer to value.

The Pricing Staircase: the framework

Picture pricing as a staircase. At the top sits the seat, the furthest abstraction from value, a proxy for a proxy. At the bottom sits the outcome, the actual result the customer wanted. Every step down ties your price closer to the value the customer receives, and AI is pushing every software company down the stairs whether the founders want to move or not.

The four steps are not four flavors of the same thing. They are four different answers to one question: what unit will I count and charge for. Read them as a descent.

Seat. You bill per human who can log in. The unit is access. It is the easiest thing in the world to count and forecast, which is why it ruled for two decades, and it is now the furthest unit from the value an agent-heavy customer receives.

Usage. You bill per unit consumed, tokens, API calls, messages, compute. The unit is volume. This was the first reflex when AI costs showed up, because your own cost is metered this way, so passing it through feels fair. It aligns your price with your cost, which is real progress, but it aligns your price with the customer’s value only loosely, and it introduces a new problem we will get to.

Action. You bill per unit of work the software completes: a record updated, a case summarized, a ticket triaged, a document generated. The unit is work done. This is a real step closer to value, because the customer is now paying for things happening, not for permission to make things happen.

Outcome. You bill per result the customer actually wanted: a resolved conversation, a qualified lead, a booked meeting, a collected invoice. The unit is the result. This is the closest you can get to value, because you are charging for the exact thing the customer hired the software to produce. It is also, as we will see, the most dangerous step to stand on if you are not ready for it.

Pricing gravity: why the unit falls

The reason every AI company drifts down the staircase has a clean mechanism behind it, and naming the mechanism makes the strategy obvious instead of trendy. Call it pricing gravity. As the cost of executing a unit of work falls toward zero, the customer stops being willing to pay for execution, and the only thing left worth paying for is the decision about what work to do and the result it produces. Price gets pulled toward the result the same way water finds the lowest point.

For twenty years, execution was expensive, so charging for access to people who could execute made sense. The bottleneck was the human in the seat. Now a model produces a competent draft of almost anything in seconds, and the bottleneck moves. A customer paying you for the ability to do work, when the work itself is nearly free, feels the mismatch immediately. They start asking why they pay forty seats when six humans plus your agents do the job. The gravity is them, repricing you in their own heads before they ever email procurement.

This is the same force I traced from the cost side in AI Gross Margins. There the story was that your costs stopped being fixed and started scaling with usage. Here it is the mirror image on the revenue side: your price has to stop being fixed per seat and start scaling with the value you deliver, or the two halves of your unit economics drift apart until the gap swallows the business. Margin and pricing are the same problem seen from two ends, and you cannot fix one while ignoring the other.

Pricing gravity does not mean everyone belongs at the bottom step tomorrow. It means the equilibrium is moving down, and standing still is itself a decision to be repriced by the market on the market’s terms rather than yours. The founders who win this are the ones who choose their step deliberately, knowing exactly why they are not one rung higher or lower.

The Margin Inversion most founders miss

Before choosing a step, you have to see the trap that makes per-seat actively dangerous in AI, not merely outdated. In classic software, your price was fixed per seat and your cost of serving that seat was close to nothing. One more user cost you a rounding error in storage. That is the whole reason SaaS posted 80 to 90 percent gross margins and venture investors loved it. Price flat, cost near zero, margin enormous.

AI breaks the second half. Every action your agent takes burns real compute, and that compute is a real bill from a model provider that arrives whether or not you priced for it. So if you keep charging a flat per-seat fee while your customer’s agents do more and more work under that seat, your revenue per account stays flat while your cost per account climbs with every task. The margin does not just shrink. It can invert, where your best, most active customers cost you more to serve than they pay.

Look at the middle row, because that is where most AI products sit right now without realizing it. They bolted an AI feature onto a per-seat plan and shipped it, and the compute bill is quietly eating the plan. The fix is not to ban usage. It is to make sure that when usage rises, revenue rises with it, so the line you are charging and the line you are paying never diverge. Even the AI labs sit at 50 to 60 percent gross margins rather than the old 80 to 90, because compute is a permanent cost of goods now, and pretending otherwise is how a growing AI company posts more revenue every quarter and less profit.

The Value-Metric Test: three gates every price must pass

So which step do you stand on. The answer is not always the bottom one, and the founders who assume it is are the ones who get hurt. A billable unit is only worth using if it passes three gates at once. I run every pricing decision through these three questions, and they kill more bad ideas than any spreadsheet.

Gate one, Aligned. Does the unit move with the value the customer receives. Per-seat fails this gate hard in the AI era, because value and seat count have come unstuck. If a customer can triple their output without adding a seat, your seat-based price is not aligned, full stop. This is the gate the whole market just failed per-seat on.

Gate two, Deliverable. Can you reliably hit the unit at a cost you control. This is the gate everyone forgets, and it is the one that quietly bankrupts people. If you charge per resolved ticket but your agent only resolves the ticket 70 percent of the time and burns three times the expected compute on the hard ones, you have priced a promise you cannot keep at a cost you cannot predict. Pure outcome pricing fails this gate whenever the agent behind it is not reliable enough yet.

Gate three, Predictable. Can the customer forecast the bill before they sign and again before it arrives. Pure usage-based pricing fails this gate constantly. A bill that swings from 15 dollars to 150 with no warning creates anxiety, slows the buying decision, and breeds churn. A power user can rack up a surprise bill in the thousands without noticing, and one such surprise can cost you the trust you spent a year building.

Here is what the test reveals once you hold all three gates up at once. No single pure model passes all three. Per-seat fails Aligned. Pure usage fails Predictable. Pure outcome fails Deliverable until your agent is genuinely reliable. The structure that can pass all three at the same time is a hybrid: a base fee that keeps the bill predictable and covers your fixed cost, plus a metered value unit on top that keeps you aligned, plus caps that keep the meter from ever shocking the customer. That is not a compromise. It is the only shape that satisfies the constraints.

The receipts: how the winners actually price

None of this is theory anymore. The companies furthest down the staircase have already shown what each step does in the wild, and their receipts are the best teacher available.

Intercom moved to the outcome step and it worked. Their Fin agent charges 0.99 dollars per resolved conversation. Not per seat. Not per month. Per result. That one pricing move is widely credited with taking the product from around one million to over a hundred million in annual recurring revenue, because the unit they bill is the exact thing a support leader wants to buy, a closed ticket, and the buyer can calculate the return in their head. HubSpot followed the same logic and priced resolution even lower. When you can deliver the outcome reliably, charging for it is the most honest and most expansive thing you can do.

Salesforce shows what happens when you have not figured out your step yet. Agentforce has shipped three different pricing models in roughly eighteen months. It launched at 2 dollars per conversation. Then it moved to a credit system, around 100,000 credits for 500 dollars, where each action burns about 20 credits, which works out to roughly 10 cents an action. Then it added per-user licenses starting around 125 dollars a month. Three models, three steps of the staircase, in a year and a half. That is not incompetence. It is the largest software company on earth searching for its billable unit in public because the ground moved under it too. The multi-model approach reportedly works better than the single per-conversation price ever did, and the product crossed half a billion in annual recurring revenue while still reaching only a small slice of the customer base. The lesson is not which price they landed on. It is that even Salesforce had to climb down the stairs by trial, and they had the brand to survive the thrash. You may not.

The usage step has a body count too. More than one AI coding tool priced purely on consumption, watched users burn a month of quota in a handful of heavy prompts, and faced a backlash loud enough that the chief executive had to publish a public apology and issue refunds. The product was good. The pricing was unpredictable, and unpredictable pricing reads to a customer as a broken promise even when the software works perfectly. That is gate three failing in real life, and it is expensive.

The pattern across all of them is the same. The companies that win are not the ones that picked the trendiest model. They are the ones whose billable unit cleared all three gates for their specific product, and who wrapped a base and a cap around the meter so the customer never got surprised. The migration toward value is real, but the ones who survived it did it with guardrails.

The Re-Pricing Path: how to migrate without nuking revenue

Knowing the right step is the easy part. Moving an existing product onto it without detonating your revenue and infuriating your installed base is the hard part, and it is where most re-pricing efforts die. You cannot just email your customers on a Monday and tell them their per-seat plan is now per-outcome. Here is the path I use, and the order matters.

Keep a base fee. Do not jump straight to a pure meter. A base platform fee gives you a predictable revenue floor, covers your fixed costs, and pays for the work the agent attempts but does not finish, which a pure outcome model would force you to eat for free. The base is what makes the whole structure safe for both sides.

Meter the real value unit on top. Add a per-action or per-outcome charge above the base, attached to whichever unit cleared all three gates for your product. This is the part that keeps your price aligned as customers automate more, so your revenue grows with the value you deliver instead of staying frozen at the seat line.

Cap it. Ship spend caps, usage alerts, and in-product visibility on day one, not after the first angry email. Bill shock is a product problem, not a billing problem, and the fix lives in the product: let customers set a ceiling, warn them as they approach it, and never let a surprise bill land. This single step is what separates the usage models that grew from the ones that triggered public apologies.

Price the meter off cost per correct task. Set your unit price against what it costs to produce a correct result, not a raw API call. A cheap call that returns a wrong answer the customer has to redo is not cheap, it is negative. I unpack this exact metric in Cost Per Correct Task, and it is the number your meter should be built on, because it is the only one that ties your price to value actually delivered rather than motion.

Roll out on expansion first. Do not re-price your whole base in one move. Introduce the new structure on new logos and at renewal, grandfather your existing seats for a defined window, and let the new model prove itself on fresh accounts before you ask loyal customers to switch. You get the data and the testimonials before you take the risk, and your installed base feels respected instead of ambushed.

The contrarian take: outcome pricing is downstream of reliability

Here is the part the pricing threads get wrong, and it is the most important thing in this piece. Outcome-based pricing has become the thing every founder thinks they are supposed to want, the bottom of the staircase treated as the obvious destination. Intercom did it, it printed money, so everyone copies the 0.99-per-result move onto their own product and waits for the same result. Most of them are about to learn an expensive lesson.

You cannot charge for an outcome you cannot reliably deliver at a cost you control. Pricing power is downstream of reliability. If your agent resolves the ticket 95 percent of the time at a predictable compute cost, outcome pricing is a gift, because you are charging for a result you can actually produce. If your agent resolves it 70 percent of the time and burns wildly different amounts of compute on the hard cases, outcome pricing is a loaded gun pointed at your own margin. You are now contractually paid only when you succeed, while you pay the compute bill every time you try, including the times you fail. The math only works if the success rate is high and the cost per attempt is stable. Both of those are reliability properties, not pricing properties.

This is why the real sequence runs in the opposite order from the one founders rush. It is not pricing first. It is reliability, then measurability, then outcome pricing. First you make the agent reliable enough that the outcome happens almost every time, which is the hard engineering work I wrote about in Loop Engineering and The AI Eval Budget. Then you make the outcome cleanly measurable, so neither side argues about whether it happened. Only then have you earned the right to price on it. Skip the first two and outcome pricing does not align your incentives, it transfers all the execution risk onto you while the customer pays only on the upside.

So the contrarian move is to be honest about which step your product has actually earned. Plenty of strong AI products belong on the action step, not the outcome step, because their work is real and countable but the final outcome still depends on things outside their control. There is no shame in that. Standing one rung too high on the staircase, charging for outcomes you cannot yet guarantee, is how you end up paying your customers to use your unreliable agent. Earn the bottom step. Do not just claim it.

What to do Monday morning

Enough theory. Here is the work, in order, that you can start this week.

Name your current value metric out loud. Write down the single unit you charge for today, and then write down the unit of value your customer actually receives. If those two are not the same word, you have a pricing gap, and the size of the gap is the size of your risk. For most per-seat AI products, this exercise is uncomfortable, which is the point.

Run the three gates on three candidate units. Take seat, action, and outcome for your product and put each through Aligned, Deliverable, Predictable. Be brutally honest on gate two for the outcome unit. Do you actually deliver the result reliably, at a cost you can predict, today. Not next quarter. The honest answer usually picks your step for you.

Design the hybrid, not the pure model. Sketch a base fee that covers your fixed cost and your unfinished work, plus a meter on your chosen value unit, plus a cap. Resist the urge to go pure-anything. The pure models each fail a gate. The hybrid is the only shape that holds.

Build the meter on cost per correct task. Before you set a unit price, calculate what it actually costs you to produce one correct result, including the failed attempts. Price above that with room to spare. If you cannot calculate it, you are not ready to meter yet, and that is a reliability project before it is a pricing project.

Pilot on new accounts. Put the new structure in front of your next ten new logos and your next renewal cohort, keep your existing base grandfathered, and watch the numbers for sixty days before you touch anyone else. Let the new model earn its place with evidence, the same audience-first discipline I argue for everywhere, including in the AI-Native Founder Playbook. Pricing is a product you ship to your customers, and like any product, you launch it small, measure it, and expand what works.

The seat is not coming back. The companies that thrive through this are not the ones with the cleverest new price. They are the ones who understood that the seat was always a stand-in for value, found the unit that actually represents value for their product, and moved their price there with a base under it and a cap over it before the market forced them to. One agent now does the work of many, a shift I traced in The One-Person Company, and your pricing has to answer for that reality whether or not you choose to look at it. Look at it.

FAQ

What is AI product pricing and why is it different from SaaS pricing?

AI product pricing is how you charge for software where AI agents do work that humans used to do. It differs from classic SaaS pricing because the old model, charging per human seat, assumed that value scaled with the number of people using the tool. AI breaks that assumption by producing output without a person in the seat, so the seat stops representing value. On top of that, every AI action carries a real compute cost, so your price has to account for a cost of goods that classic SaaS never had.

Why is per-seat pricing dying?

Because agents have severed the link between a human login and the value delivered. When one agent does the work that used to need ten people, a per-seat price collects almost none of the value created, and it punishes your most automated, most valuable customers by charging them the least. The market repriced this fast: per-seat fell from about 21 percent of SaaS to about 15 percent in a year, and roughly 285 billion dollars of software value was wiped in a 48-hour window once investors did the math.

What is outcome-based pricing?

Outcome-based pricing charges per result the customer actually wanted, like a resolved support ticket, a qualified lead, or a collected invoice, rather than per seat or per unit of usage. Intercom’s Fin agent is the best known example at 0.99 dollars per resolved conversation. It delivers the tightest alignment between price and value, but it puts all the cost risk on you, because you pay to attempt every result and only get paid when you succeed.

Is usage-based pricing better than per-seat for AI?

It is more aligned with your costs, since your compute bill is metered the same way, but it has its own failure mode: unpredictable bills. A charge that swings widely month to month creates buying friction and churn, and a single surprise bill can cost you a customer’s trust. Usage works best as one component of a hybrid, paired with a base fee and hard spend caps, rather than as a pure model on its own.

How do I keep gross margins healthy with AI pricing?

Make sure your revenue rises in step with your cost. The danger with per-seat in AI is that your price stays flat while compute cost climbs with every agent action, which erodes and can even invert your margin on heavy accounts. Meter the value unit so that more usage means more revenue, price each unit off your cost per correct task rather than a raw call, and expect AI gross margins to sit closer to 50 to 60 percent than the old SaaS 80 to 90, because compute is a permanent cost of goods.

Should an early-stage startup use outcome-based pricing?

Usually not yet, and this is the most common expensive mistake. Outcome pricing only works when your agent delivers the outcome reliably at a cost you can predict, and most early products are not there. Pricing power is downstream of reliability. Earn the action step first, charge for work the software clearly does, and move to the outcome step only once your success rate is high and your cost per attempt is stable. Standing one rung too high means paying to run an unreliable agent while you only collect on the wins.

How do I migrate off per-seat without losing revenue?

Move in order and do not re-price everyone at once. Keep a base platform fee for a predictable floor, add a meter on your real value unit on top, cap the meter with spend limits and alerts to prevent bill shock, price the meter off cost per correct task, and roll the new structure out on new logos and renewals first while grandfathering existing seats for a set window. You collect data and testimonials on fresh accounts before you ask loyal customers to switch.

Is per-seat pricing completely dead, or does it still work anywhere?

It still works where value genuinely tracks headcount and humans, not agents, do the core work, which is a shrinking set of cases. For any product where AI does meaningful work under the hood, pure per-seat is structurally mismatched with value and will be repriced by the market if you do not reprice it yourself. The realistic destination for most products is a hybrid that keeps a base fee, which can even look seat-like, and adds a value meter on top, rather than a return to pure per-seat.