The paradox at the center of AI business models
An AI customer support startup I spoke with last quarter has the kind of product-market fit founders dream about. Enterprise customers love them. NPS in the 70s. Churn near zero. Every demo converts.
Their revenue per customer is shrinking.
The reason: their AI works. A customer who used to need 30 support agents now needs 15. The same customer, same product, deeper adoption, half the seat licenses. The startup is being penalized for delivering exactly what they promised.
This is the paradox at the center of AI business models in 2025. The pricing logic that built SaaS — pay per seat, pay per user, pay per human at a keyboard — collapses when the AI is doing the work instead. Every AI founder shipping today is staring at some version of this problem. Most haven't priced it yet.
Why this matters and where the thinking comes from
Andreessen Horowitz, Bessemer Venture Partners, and a handful of operator-investors have spent the last two years writing about this. Their core argument: AI products require fundamentally different pricing models than traditional SaaS, and the companies that figure this out first will be worth dramatically more than the ones still selling seats.
This is no longer theoretical. Investors are now stress-testing AI pricing in diligence. The question "how does your pricing scale when your AI gets better at the customer's work?" is the new "what's your CAC payback?" If you can't answer it cleanly, the round gets harder.
1. The deflation paradox
The first and most obvious problem: better AI means fewer humans, which means fewer seats, which means less revenue from existing customers.
A traditional SaaS company gets stronger as customers grow. More employees means more seats, more revenue, more expansion. The product's value compounds with the customer's headcount.
An AI product breaks this equation. The better the AI performs, the more headcount the customer can shed — or never hire in the first place. A customer support AI that's 50% effective triggers no seat reductions. A customer support AI that's 95% effective triggers significant ones. Success destroys revenue.
What this means for you: if you're selling AI by the seat, you've structured your business to be punished for getting better at your customer's job. That's the wrong incentive at every level — product, sales, and the customer relationship.
2. The cost-revenue mismatch
The second problem is operational. AI inference costs scale with usage. Every query, every workflow, every generated output costs you something. Seat pricing decouples revenue from usage entirely.
The result: you lose money on heavy users and overcharge light ones. Your enterprise customer running 500,000 queries a month pays the same as the customer running 5,000. Your gross margins look fine in aggregate but the underlying cohort economics are chaotic.
The early ChatGPT Enterprise pricing model surfaced this clearly. OpenAI quickly moved toward consumption-aware tiers because seat pricing couldn't reconcile with the cost reality of inference at scale.
What this means for you: if your AI product has any inference cost variability across customers, seat pricing is misaligning your unit economics. You need either consumption pricing or a hybrid model that captures usage above a baseline.
Customers don't measure value in seats. They measure it in tickets resolved, deals closed, code shipped. Price what they care about.
3. Value capture is harder
The third problem is strategic. Customers measure AI value in outcomes — dollars saved, tickets resolved, leads qualified, deals closed. They don't measure it in seats. When AI delivers outsized value, seat pricing leaves money on the table because it has no relationship to the value being delivered.
A customer support AI that saves a company $5 million annually in agent costs while priced at $200 per seat for 50 seats — $120,000 in ARR — is capturing 2.4% of the value it's delivering. That gap is invisible to the customer in the short term, but it becomes visible the moment a competitor prices on outcomes.
The AI companies pricing on outcomes right now — Intercom's Fin charges per resolved conversation, Klarna's customer service AI is priced against deflected human interactions, Harvey charges enterprise legal firms based on workflows automated — are capturing meaningfully more of the value they create than seat-priced competitors.
The three models replacing seats
Three pricing models are emerging as the alternatives to seats for AI products:
Usage-based pricing — pay per query, per workflow, per generated output. Works well when the unit of work is clearly defined and customers can easily forecast their usage. Best for developer tools, generation products, and infrastructure plays. Risk: customers fear unpredictable bills and may underutilize the product to control spend.
Outcome-based pricing — pay per resolved ticket, per qualified lead, per closed deal, per workflow completed. Works well when the outcome is measurable and attributable to the AI. Captures the most value but requires sophisticated attribution and trust. Best for customer support, sales, and recruiting AI.
Hybrid platform pricing — a base subscription that includes a usage allowance, with consumption overage above the baseline. Looks more familiar to procurement teams. Captures both predictable platform value and variable usage. Best for enterprise AI products where buyers need budget predictability.
The right model depends on your product. Most AI companies in 2025 are converging on hybrid platform pricing as the safest path, with outcome-based pricing as the aspirational endpoint.
What to do about it
Three concrete actions if you're building an AI product:
Stop selling seats unless you have a specific reason to. The default for any AI product launching in 2025 should be usage-based or hybrid. Seats are appropriate when your AI augments specific human roles in ways that scale with team size. They are inappropriate when your AI replaces or reduces human work.
Instrument outcomes from day one. Even if you're not pricing on outcomes today, build the measurement infrastructure. Track tickets resolved, deals closed, workflows automated, whatever your customer's measurable output is. You will need this data when you migrate pricing models in 12-18 months.
Talk to your customers about value, not seats. If you can't tell your customer how many dollars you saved them or how much work you did for them, your pricing conversation defaults to seats by gravity. The reframe starts with the value conversation.
The takeaway
The seat is dying as the primary unit of SaaS pricing for AI products. It's not dying everywhere — collaboration tools, project management, and software where humans are still doing most of the work will continue to use seats. But for AI-native products, the seat is a vestigial pricing model from the era when software was a tool and humans were the engine.
In the era where the AI is the engine, you need pricing that captures the engine's output. Usage, outcomes, or hybrid platforms. Pick one. Build the infrastructure to support it. Start now, because by the time your competitor figures this out it will be too late to catch up.
Want to understand the unit economics behind AI businesses? Try the AI Business Models pack on Gargiulo — five scenarios on usage pricing, outcome pricing, and the structural shifts reshaping software economics. Sterling has notes.
Sources: Andreessen Horowitz, Bessemer Venture Partners, public pricing data from Intercom, Klarna, Harvey, and OpenAI.