There is no universal AI pricing strategy, and the search for one is the mistake. AI products run at 25 to 60% gross margins where classic SaaS ran 70 to 90%, so playbooks built for flat-rate software break on contact. PricingSaaS tracked 8,394 pricing changes across 498 companies in 2025; nobody has settled. The durable method: diagnose your starting position, pick a value layer you can defend, and price against the margin you can see per customer.
The SaaS pricing era is over for AI
The stable-margin SaaS pricing era is over, and the frantic hunt for "the" AI model to copy is the wrong response. For two decades, software pricing was a solved problem: charge per seat, hold gross margin in the 70 to 90% range, and grow by adding logos. That worked because an extra user on a paid-for server cost almost nothing.
AI deleted that assumption. Every inference has a real, variable cost that lands the moment a customer hits your product, and that cost moves every time you change models.
The market is reacting by changing pricing constantly. PricingSaaS counted 8,394 pricing and packaging events across 498 companies in 2025, with packaging changes up 21% year over year (PricingSaaS Q1 2026 Trends Report). That is not a market converging on a best practice. That is a market that has not found one. If you are building real-time monetization for AI products, the takeaway is simple: the playbook is not coming, and you need a method instead.
A pricing model is a bet on your unit economics. In AI, that bet is wrong more often than it is right, and you usually cannot tell until the invoice or the renewal arrives.
Why AI pricing strategy has no settled answer
AI pricing has no settled answer because the cost structure underneath it is unstable in ways SaaS pricing never had to handle. The variables that decide your margin (model choice, token volume, task complexity, customer behavior) all move independently, and the result per customer often stays invisible until it is too late to act on it.
Why SaaS pricing playbooks break on AI
SaaS playbooks assume high, stable gross margins, and AI does not provide them. Bessemer puts AI-native gross margins at 50 to 60%, against 80 to 90% for traditional SaaS, and pegs the fastest-growing "Supernova" cohort near 25% in its first stage (Bessemer Venture Partners). ICONIQ's panel of roughly 300 software executives reported AI gross margins around 45% in 2025, climbing toward a projected 52% in 2026 (ICONIQ).
The consequence is structural, not temporary. The SaaS CFO estimates that an AI company needs roughly six times the revenue of a traditional SaaS company to produce the same EBITDA (The SaaS CFO). Every pricing decision that ignores variable cost of goods sold is borrowing against a margin that may not exist.
Three value layers, three different risks
There are three layers you can price on, and each one moves the risk to a different place. Most "AI pricing models" advice stops at naming them. The decision that matters is who absorbs the risk when reality diverges from the plan.
| Value layer | You charge for | Margin signal | Who carries the risk | Where it fits |
|---|---|---|---|---|
| Consumption | Tokens, calls, compute time | Cleanest cost recovery | Vendor carries demand risk; buyers fear unpredictable bills | Infrastructure and API products (Twilio, Snowflake, ElevenLabs) |
| Workflow / effort | A discrete unit of work done | Decent, if you can cost the work | Vendor, when the per-workflow cost metric is wrong | Vertical AI with bounded tasks (Cursor, Replit) |
| Outcome | A verified result | Highest value alignment | Vendor carries performance risk | Measurable, repeatable results (Intercom Fin, Zendesk, Salesforce Agentforce) |
Consumption pricing gives the cleanest cost signal, but it transfers demand risk to you and unsettles buyers who cannot predict their bill. This is why pure consumption-based pricing fits infrastructure products. Twilio drew 73% of its revenue from usage-based fees in its Q1 2025 10-Q, and Snowflake runs almost entirely on consumption.
Outcome pricing sits at the other end. Intercom's Fin agent charges $0.99 per resolved ticket (Intercom, Get Paid Podcast), Zendesk prices automated resolutions around $1.50 (Zendesk), and Salesforce Agentforce charges about $0.10 per action (Salesforce). Outcome pricing aligns price with value better than anything else, but it hands the vendor the performance risk, which few products outside support automation are ready to carry.
The dominant move in 2025 was neither extreme. It was hybrid and credits. Kyle Poyar's data shows flat-fee pricing falling from 29% to 22% of companies and hybrid pricing rising from 27% to 41% in a single year (Growth Unhinged).
Credit models grew 126% year over year as both AI-native companies and legacy players like Figma and HubSpot adopted them (PricingSaaS Q1 2026). Credits are a useful wrapper while you calibrate the value layer underneath. They are not a destination.
Your starting position changes everything
Your starting position decides your strategy more than the model menu does, because three companies in three situations need three different answers. The pre-revenue team, the team with revenue that feels wrong, and the established business bolting AI onto a seat-based product are not solving the same problem.
A pre-revenue product giving away a free beta risks training customers to expect free compute forever, and doing it with no cost signal is how projects die quietly. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, citing unclear ROI and inadequate cost controls (Gartner, June 2025).
A product with revenue that feels wrong has data but a broken model. Replit's gross margin swung from 36% to negative 14% as its agent consumed more model cost than the pricing covered (Aakash Gupta, February 2026).
The established SaaS company adding an AI tier faces the 2026 renewal cliff. Tropic found AI driving 20 to 37% price increases at renewal, while fewer than a third of companies can tie AI spend to measurable P&L impact (Tropic, December 2025). Zylo's 2026 SaaS Management Index reports 79% of IT leaders hit renewal price increases, 78% saw unexpected AI-related charges, and 61% cut projects over unplanned costs (Zylo, June 2026). The same patterns show up in broader AI monetization strategies for startups, where the right move depends entirely on where the company is starting from.
The traps that look like strategy
The most expensive pricing mistakes look like strategy from the inside. Four show up repeatedly.
Cost-plus pricing sets the price off your GPU bill rather than the customer's value, anchoring you to your costs instead of their willingness to pay. The soft-ROI copilot charges for interactions or seats while the AI only advises, never closing a loop a CFO can measure. Stranded credits and runaway complexity erode trust: Cursor's June 2025 credit changes drew a public backlash when opaque balances produced surprise charges (Cursor), and price restructuring made up 26.7% of all 2025 pricing events (PricingSaaS Q1 2026).
The fourth trap is the one we see most. When we talk with AI teams hitting their first margin crisis, the pattern is almost always the same: the customer they were proudest of, the one with the highest engagement, turned out to be the one losing them the most money.
Wildfire Labs worked the math and found that a power user at 2,000 to 3,000 queries a month can carry a contribution margin between zero and negative $20 (Wildfire Labs, March 2026). Intercom has said a single Fin customer's monthly bill can range from $50 to $30,000 on one flat per-resolution rate (Aakash Gupta, February 2026). Average revenue looks healthy while specific customers bleed, and aggregate dashboards never show it.
Stop picking a model. Start defending a margin.
The reframe that fixes this is to judge a pricing model by one question: which margin can I defend per customer, and can I even see it? A pricing model is a hypothesis about your unit economics. AI is the first software category where that hypothesis is routinely wrong, in a direction that compounds, and invisible until renewal. The model you pick matters less than your ability to measure whether it is working.
Pricing in AI is margin defense, not revenue optimization. Ridgeway Financial Services frames it bluntly for finance teams: pricing is a compute risk management tool. Intercom's CFO Dan Griggs reached the same conclusion from the product side, describing the goal as connecting how the company monetizes to how customers realize value (Get Paid Podcast).
That conversation is impossible without per-customer cost data, and most teams do not have it. Only about 43% of companies can attribute AI cost to a specific customer, and only 22% track it per transaction (CloudZero, State of AI Costs 2025). A pricing strategy you cannot measure at the customer level is a guess you renew every quarter. This is also why mature products converge on hybrid pricing for AI products: a base fee covers the floor cost while usage or outcome charges scale with the margin you can actually observe.
When flat pricing is still the right call
Flat or seat-based pricing is the right call in specific, defensible cases. The honest version of this argument has to name them, or it is just a pitch.
The first case is low variance. If your AI feature consumes roughly the same compute for every user, so the 90th-percentile customer costs about what the median costs, a flat or seat price stays honest. Intercom has kept Fin at $0.99 per resolution since launch despite rising conversation complexity, choosing simplicity over differential rates.
The second case is genuine infrastructure. For products where consumption is the value, like Twilio, Snowflake, or Cloudflare, pure consumption pricing is the honest model, and layering outcome pricing on top would add attribution complexity nobody needs.
The third case is the deliberate free beta. Giving usage away to learn can be right, as long as you watch the cost signal while you do it. Early-stage teams have a structural edge here, because early-stage teams have a pricing-iteration advantage before legacy contracts and entrenched expectations set.
The line is variance, not ideology. If your heaviest customer costs twice your median, flat pricing is fine. If they cost fifty times, it is a slow leak.
An AI pricing strategy is a method, not a playbook
The method is what survives, even when the answer keeps changing. Start from value rather than cost. Pick a value layer you can defend given your variance and your ability to measure outcomes. Measure margin per customer, where the aggregate revenue line hides the truth. And keep pricing changeable by configuration rather than by a code release, because you will get it wrong, fix it, and probably get it wrong again.
That last point is where strategies quietly fail. A pricing model you cannot change without an engineering sprint is a model you will not change, even when the margin data says you should.
The teams that price well can see gross margin per customer as events happen and adjust pricing without shipping code. The model they started with matters less than the feedback loop they built around it. That is the layer Credyt provides: real-time usage-based billing, per-customer cost and margin attribution, and pricing models you can version in a dashboard instead of re-architecting. The method above only becomes practical when the margin signal is live and the pricing is soft enough to move.
Pricing for AI will stay a moving target. The goal is to see the miss early enough that it costs you a configuration change instead of a customer. See how Credyt works.