How to price an AI product

If you price your AI product like software, you will leave most of your value on the table. And you'll do it politely, quietly, and with a smile, because the old SaaS playbook feels safe. It isn't. It's the single most expensive mistake founders are making right now.

Here's why.

The labor budget is ten times the software budget

Every wave of technology unlocks a new monetization model. SaaS gave us subscriptions. Cloud gave us pay-as-you-go. AI's shift is more fundamental than either. For the first time, you are not selling software for a human to use. You are selling software that does the work a human used to do.

That changes which budget you're drawing from.

Software and IT budgets are a rounding error compared to labor budgets. Labor is roughly 10x larger. When your product resolves a support ticket, drafts a contract, qualifies a lead, or reconciles an invoice without a human in the loop, you are competing with the fully loaded cost of the person who used to do that work, not with the Salesforce license sitting next to it.

If you price against the software line item, you have anchored yourself in the smaller budget. You have made yourself easy to buy and impossible to monetize. Founders who understand this are capturing 25 to 50 percent of the value they create. Founders who don't are capturing the polite SaaS fraction of 10 to 15 percent and calling it a win.

Per-seat pricing is catching a falling knife

There is one diagnostic question every AI founder should ask before touching a pricing page:

Is my product making people superhuman, or is it replacing them?

If it makes people superhuman (a copilot, an assistant, an accelerator), seat-based pricing can still work, because more value flows through more humans. Charge per seat and you scale with adoption.

But if your product replaces labor, per-seat pricing is a contract you've signed against your customer's success. The better your product gets, the fewer seats they need. You have built a monetization model that shrinks as you deliver more value. That is not a business. That is a countdown.

Intercom saw this clearly with Fin. They leapt from the most-resented pricing model in SaaS, per-seat, to $0.99 per AI-resolved ticket, and zero when a human has to step in. The contrast tells the whole story: I win when you win. That is a sentence no seat-based vendor can honestly say in the AI era.

The autonomy-and-attribution map

So how do you decide what model to use? I use a simple 2x2. Two axes, four quadrants, four different pricing answers.

The vertical axis is autonomy. How much of the work does the product do without a human in the loop? Low autonomy means a human is driving and the AI is assisting. High autonomy means the AI completes the task end to end and the human is, at most, reviewing.

The horizontal axis is attribution. Can the value delivered be measured and shown clearly? Low attribution means the impact is diffuse: productivity gains you have to argue for. High attribution means you can point at a dashboard and say: that ticket got resolved, that lead got qualified, that hour got saved, that dollar got recovered.

The four quadrants:

Low autonomy, low attribution: per-seat. A human is doing the work and the AI is helping in ways that are hard to isolate. Classic copilot territory. Per-seat still fits here, because you're genuinely renting a productivity tool. But be honest about whether you're really here, or whether you're just afraid to leave.

Low autonomy, high attribution: hybrid (seat plus usage). Humans still drive, but you can measure specific units of value: documents processed, queries answered, analyses run. Charge a platform fee to clear the procurement bar, then layer usage on top to capture the value delivered. This is the razor-and-blade of the AI era.

High autonomy, low attribution: usage-based. The agent does the work end to end but the business outcome is hard to isolate cleanly. Charge by consumption: tasks completed, runs executed. Usage scales with value delivered even when you can't point at the dollar it produced.

High autonomy, high attribution: outcome-based. The holy grail. The product acts autonomously and the result is undeniable. Charge per resolved ticket, per recovered dollar, per booked meeting. This is where you can defensibly capture 25 to 50 percent of the value you create, and it's the quadrant incumbents structurally cannot follow you into, because outcome-based pricing destroys the seat-based margins their business depends on. The monetization model itself becomes the moat.

Wherever you sit on that map today, one job is the same: build attribution into the product now, before you need it. Dashboards that show tickets resolved, hours saved, revenue influenced. Not because your current pricing depends on it, but because your next pricing model does. You cannot migrate customers from seats to outcomes without proof. And proof is a product decision, not a pricing decision.

Now let me price three actual readers.

Reader one: the AI-native founder

A startup sells an AI support agent that autonomously resolves about 60 percent of a brand's tickets, no human in the loop on those. They charge $1,500 a month flat, plus $50 per seat for the dashboard. Twelve customers, mostly converted pilots, but new deals stall in procurement, and one buyer literally asked: “Why am I paying per seat if the point is fewer agents?”

Place them on the map first: high autonomy, high attribution. A resolved ticket is one of the cleanest attribution units in all of AI. It either closed without a human or it didn't. Binary. Auditable. They sit squarely in the outcome-based quadrant, and they are pricing themselves in the per-seat quadrant. That gap is the entire problem. That buyer wasn't being difficult. They were telling the founder the truth: the model contradicts the product.

Three things are going wrong at once. The $1,500 anchors them in the software budget, so procurement evaluates them against Zendesk add-ons instead of the headcount they replace. The seat fee sends a signal that directly contradicts the pitch; the buyer's brain rejects it before procurement ever sees the contract. And they are capturing a rounding error: a resolved support ticket costs a company somewhere between $4 and $12 fully loaded, which means at any real volume they're creating tens of thousands of dollars of monthly value and capturing 3 to 5 percent of it.

The rebuild:

Model: outcome-based, priced per autonomously resolved ticket. A human touches it, you don't charge.

Metric: AI-resolved ticket, defined tightly in the contract: closed without human intervention, satisfaction not negative, no reopens within X days. Tight definitions protect both sides and make renewals trivial.

Number: somewhere in the $1.00 to $2.50 range, set from the customer's own cost-per-ticket baseline. If their fully loaded human-resolved ticket costs $8, charging $2 captures 25 percent of the value and saves them 75. That is the ratio that makes CFOs sign. Don't set one universal price; set a share of their attributed savings, then translate it into a per-ticket number.

Keep a small platform fee ($500 to $1,000 a month) to cover onboarding and clear the “is this a real vendor” bar. Kill the per-seat line entirely.

Sanity check: a customer doing 20,000 tickets a month at 60 percent autonomous resolution is 12,000 resolved tickets. At $1.50 each, that's $18,000 a month, versus $1,500 plus seat fees today. Over 10x the revenue per customer, while still saving the customer 75 percent against their human cost. Both sides win. That's the test.

And for the next pilot, four rules, non-negotiable:

1. Charge for the pilot. Free pilots attract tire-kickers and train the buyer to expect zero.
2. Decouple pilot pricing from commercial pricing in writing. One sentence in the agreement: “Pilot pricing reflects setup and validation and is not indicative of production pricing.” That sentence is worth millions in aggregate.
3. Co-build the business case with the customer from day one. Jointly define current cost per ticket, volume, what an autonomous resolution saves, and what you'll call success. When their fingerprints are on the assumptions, they can't challenge the conclusions at procurement.
4. Build the attribution dashboard before the pilot starts. Not a slide: a live view inside the product. The dashboard is the pricing strategy.

When the founder walks into the next procurement conversation, the answer to the price question isn't a number. It's this: “You're currently spending roughly $X per ticket with your human team. Our agent resolves 60 percent of them autonomously. We charge a fraction of your current cost, only when we resolve, and nothing when we don't.” That sentence moves the deal out of the software budget and into the labor budget, and the seat question disappears, because there are no seats left to argue about.

Reader two: the SaaS leader bolting AI on

A project-management tool at $12 per seat per month, 40,000 paid seats. They shipped an AI copilot: meeting summaries, task drafting, weekly status reports. Competitors are giving similar features away free. The board is split between “charge $8 per seat as an add-on” and “bundle it free to defend the base.”

The board is having the wrong argument. That binary skips the only question that matters: what is the AI actually doing, and can you prove it?

Place it on the map honestly. Meeting summaries and task drafting are a copilot. A human still runs the project. That's low autonomy. And if you could show hours saved per user but don't instrument it today, you're low attribution too, whether you like it or not. That's the per-seat quadrant, possibly evolving toward hybrid. It is not the outcome quadrant, and any pricing that pretends otherwise will collapse.

There are three doors, not two:

AI belongs in the base when its job is to defend the seat, not to sell itself: competitors are commoditizing the feature, it touches every user's daily workflow, and switching costs are your real business model. That's exactly this case. If a competitor's free copilot is good enough and yours is behind an $8 paywall, you're not monetizing AI; you're handing them a wedge into your $5.76M-a-year seat engine. But bundling free while holding the base price flat is the amateur move. Bundling is your justification to raise the base. At the next renewal cycle, $12 becomes $14 or $15, framed as the AI-native version of the platform. On 40,000 seats, every dollar of base raise is $480K a year, so a $3 raise is $1.44M: far more than an $8 add-on at a realistic 15 percent attach would generate (about $576K). And the story is clean: here's what's now included, here's the new price.

AI should be an add-on when it's genuinely differentiated, only a subset of users want it, and attribution is clear: an advanced analytics module, a compliance copilot. Summaries and drafting are not that. They're horizontal, and competitors are proving they can't sustain a standalone price. Save this door for the next capability, the one that isn't commoditized yet.

AI needs its own meter when it does discrete units of work with clear per-unit value and wildly varying consumption. Not true of today's copilot. It becomes true the moment they ship something autonomous: an agent that builds project plans from a brief or reconciles status across fifty projects. The meter isn't for today. It's for what they ship next quarter, and that capability should be designed as metered from day one and never bundled by default.

The trap to name explicitly: the board compromises at $4 per seat, gets 20 percent attach, books roughly $384K a year, and calls it a win. Meanwhile free copilots become table stakes, the base erodes, and in eighteen months the add-on collapses to zero. They will have monetized AI badly and failed to defend the core. Both losses at once.

Before the exec team leaves the room, one question for the CEO: “In three years, what percentage of our revenue comes from something other than seats?” If the honest answer is “we don't know,” they haven't built an AI business. They've built an AI feature.

Reader three: the solo builder

A one-person company: an AI tool that turns long videos into short clips, $19 a month unlimited. Power users hammer it. Inference costs eat 70 percent of revenue on the heaviest accounts while most subscribers barely use it.

Unlimited pricing on an AI product with real inference costs is a bet that usage will average out. It never does. Usage on any AI tool is not a bell curve; it's a power law. Price a flat fee against a power law and you get exactly this: heavy users subsidized by light users, light users subsidized by the founder's savings account. You cannot fix it with a price increase. Raise $19 to $29 and the light users churn while the heavy users celebrate the bargain. The model has to change.

The fix is hybrid: a low base plus metered usage on the unit that drives both cost and value. Here, minutes of source video processed. Concretely:

• Starter: $9 a month, 60 minutes included. Overage at $0.20 a minute.
• Creator: $29 a month, 300 minutes. Overage at $0.15.
• Pro: $79 a month, 1,000 minutes. Overage at $0.12.

This kills the unlimited-abuse problem, preserves a cheap entry point for the light users who were the healthy part of the book, and introduces good-better-best, where the middle tier becomes the anchor and Pro exists partly to make Creator look reasonable. Grandfather existing customers for 6 to 12 months and frame the change as “plans that scale with how you actually use the product,” not as a price increase. And cap your free tier in minutes, not days: time-based trials on AI products bleed money, and usage caps train customers from day one that value comes in units.

For the version of this reader who hasn't launched yet: you don't get to choose whether you'll have the pricing conversation with your customers. Only when. Have it before you write the pricing page. Fifteen to twenty conversations, with actual target users, not friends (friends will encourage you; strangers will price you). Mix three profiles: people who pay for a substitute today, people who do it manually, and people who considered the problem and gave up. Pay $30 to $50 for the call; paid calls filter for seriousness on both sides.

Then the questions, in order, with half your follow-ups being “why”:

1. “Walk me through the last time you had this problem. What did you actually do?” Start in their world. Their words become your landing page.
2. “What did that cost you, in time or money?” This is your value baseline, and it's usually labor, not software.
3. Describe the product in one sentence, then: “How often would you use it, and for what?” Frequency tells you flat versus metered.
4. The three price questions: “At what price is this an easy yes?” (acceptable), “At what price is it expensive but still worth it?” (anchor here: this is where they respect what you built), “At what price would you laugh and close the tab?” (your ceiling). After each number: why? The number is meaningless. The reasoning is everything.
5. “If I charged per unit of work instead of a monthly fee, how would you feel?” Test the model, not just the number, before you build the billing system.
6. “What would have to be true for you to pay double?” Sometimes the answer is “nothing.” Sometimes it's your Pro tier, discovered in one sentence.
7. “Who else is involved in this decision?” It tells you whether you're pricing to a user or a buyer.

Never ask “would you pay $19 for this?” Yes/no questions on absolute prices produce garbage. People are absolutely meaningless and relatively super smart: always ask ranges and comparisons. And if creators say $19 while agencies say $99, you do not have a $59 product. You have two products with two prices, and one pricing page will fail both. Averages always lie.

The traps that survive contact with all of this

Founders read everything above, nod along, and still blow it. The ones I actually see:

The free pilot that becomes the anchor. The tell: month five of a 90-day free pilot, and the customer opens the contract conversation with “well, we've been using it for free, so...” The fix: charge for every pilot, even $10K on a future $500K deal, and put the decoupling sentence in writing. The purpose of a POC is not to prove the technology. It's to co-build the business case that makes your champion unbeatable at contract time.

Pricing on tokens instead of outcomes. The tell: “tokens,” “credits,” or “compute units” on your pricing page, and customers doing math to guess their bill. Tokens are your cost unit, not their value unit, and every efficiency gain you ship becomes a revenue cut. The fix: price on the unit the customer already measures (resolved tickets, documents analyzed, minutes processed), even when the underlying math is identical. Present “per document” next to “per 10,000 tokens” and watch: customers are never indifferent.

Underpricing out of commodity anxiety. The tell: “we're basically a wrapper on GPT, so we can't charge much.” The customer is not buying the model. They're buying the work delivered. Intel didn't stop Dell from charging for computers; AWS didn't stop Snowflake. If your product saves $50K a month in analyst time and you charge $2K because you fear API price cuts, you are not being humble. You're being irresponsible with the value you created. Your reluctance to charge is internal and emotional, not external and logical.

The minivation. The tell: great product, raving customers, strong retention, and revenue per customer that's a fraction of value delivered, justified as “we'll optimize pricing after product-market fit.” Charging $50 a month for something that saves $5,000 feels safe. It is a slow-motion catastrophe: every reference call and review is training the market on an anchor you'll never raise. The gap between what customers would pay and what you charge is almost always 3 to 10x. Close half of it this quarter.

Feature shock, and the hidden gem. Two of my four product failure types, both amplified by AI. Because the model can do everything cheaply, founders ship everything: a 47-feature Swiss Army knife nobody can buy. Run the killer-feature test: anything valued by under 20 percent of target customers and actively unwanted by more than 20 percent gets cut or sold separately. And once a quarter, ask your top ten customers: “if we removed one capability, which one makes you cancel?” That answer is your hidden gem. Sometimes it should be its own product. Sometimes it should be the anchor of the whole company.

Land and pray. The tell: “we're land-and-expand,” followed by a long pause when asked what the expansion mechanic is. Ninety percent of companies claiming land-and-expand are only landing. If you price low to land, you need a mechanical path up: usage components that grow with adoption, modules that unlock at higher tiers, contractual step-ups. If none exist, you're not land-and-expand. You're just underpriced.

Optimizing the number while the model is wrong. The deadliest one, because it feels like work. Weeks debating $49 versus $59 while the product is priced per seat and should be priced per outcome. How you charge matters more than how much. Michelin couldn't charge 20 percent more for a tire that lasted 20 percent longer, so it charged by the mile and captured the value instantly. That was a model change, not a number change. Revisit price points often; revisit the model rarely but decisively.

The trap underneath all of these is treating pricing as a downstream task, something you handle after the product is built and the launch has “real data.” In the AI era, pricing is a product design input, not a finance output. The founders who understand this capture 25 to 50 percent of the value they create. The ones who don't build extraordinary products and hand most of the value to their customers as charity. Neither of those is wrong. But only one of them is a business.

The Monday-morning plan

Five days. No pricing consultant, no six-month project. Just you and the discipline to actually do it.

Monday: place yourself on the map. Two hours, alone. Does your product make people superhuman or replace their work? Can you point at a dashboard and show what got done, saved, or earned? Plot yourself. Write down the model your quadrant calls for and the model you use today. If they match, skip to Friday. If they don't, you now know exactly what you're solving for.

Tuesday: talk to five customers. Your best, your worst, your newest, one who nearly churned, and one prospect stalled in procurement. Thirty minutes each, four questions: how do you actually use this in a normal week; what would you have paid someone to do this work before us; if we charged per outcome instead of per seat, how would you feel; and the three price questions. Follow every number with “why.” The numbers will lie. The reasons will not.

Wednesday: do the value math. For your top three segments: value created (hours saved times fully loaded labor cost, using Tuesday's numbers, not your marketing deck's) versus value captured (what they pay you). Divide. Under 15 percent: you're a minivation with real room to move. Over 40 percent: you have pricing power most founders would kill for; protect it and don't apologize. Three hours, and it reveals more than a quarter of pricing analysis.

Thursday: design the new model on one page. The metric you'll charge on (the unit of value the customer already measures; not tokens, not seats if you replace labor). Three tiers, not one, so sales has pivots instead of a hill to die on. The number, anchored in Tuesday's “expensive but worth it” zone, never the “easy yes” zone: you can always come down, you almost never go up. And the fence between tiers: if a customer can't articulate the difference in one sentence, the fence isn't real and everyone lands on cheap. If it doesn't fit on a page, it's too complicated to sell.

Friday: ship something. The Friday mistake is a Notion doc titled “Q2 Pricing Strategy” that never ships. Pick one: update the pricing page for new customers only, quote your next pipeline deal on the new model, or structure the upcoming pilot with the new metric, a paid fee, and the decoupling clause. You don't need to migrate everyone this week. You need to start. The compounding cost of selling the wrong model every day is larger than the risk of shipping an imperfect new one.

Then two calendar reminders: six months out, revisit the model; twelve months out, revisit the numbers. That cadence is not optional.

The one thing to remember a year from now

In the AI era, you are no longer selling software. You are selling work delivered. Price accordingly, or someone else will.

This piece was produced in working sessions with Corner's Madhavan Ramanujam advisor, an AI modeled on his published work on monetization. It was not written or reviewed by Madhavan Ramanujam. To run these frameworks against your own numbers, open a session: pick Price your AI product, or start with Price check if you just want a verdict on your current price.