AI broke pricing, and most founders haven't caught up

SaaS pricing was hard but at least it was predictable. You had seats, tiers, and maybe usage overages. The unit economics were stable. You could model your gross margins twelve months out and be roughly right. AI products have broken that model in ways most founders are still pretending not to notice.

The problem isn't that AI is expensive. It's that the cost structure is fundamentally different, and the pricing models founders reach for by default were designed for a different cost structure.

Why the old models don't fit

Seat pricing assumes flat marginal cost

In traditional SaaS, adding a user costs you almost nothing extra. The marginal cost of one more seat is roughly zero. That's why per-seat pricing works so well: you capture value per user and your costs don't scale with them. With AI, every query, every document processed, every agent action has a real cost. Power users can cost you ten times what light users cost. Flat per-seat pricing turns your best customers into your worst margin customers.

Usage pricing creates the wrong incentives

The obvious fix is usage-based pricing: charge per token, per query, per action. But this creates a different problem. Users become anxious about costs and under-use the product. The product that delivers the most value is also the one that costs the most to use, so the most engaged users feel punished. Twilio and AWS can get away with pure consumption pricing because their customers are developers who are comfortable reasoning about usage. Most end users are not.

Outcome pricing is compelling but hard to measure

The cleanest model is to charge for outcomes: a percentage of the value created, a fee per successful action, a share of time saved. This aligns incentives perfectly in theory. In practice, attributing outcomes to AI is genuinely difficult. Did the AI close the deal, or did the human? Did it save two hours or twenty minutes? Outcome pricing also requires customers to trust your measurement, which takes time to establish.

What's actually working

The founders who have figured this out tend to use hybrid models that combine a base subscription with some form of usage cap or credits system. The subscription covers light-to-moderate use and creates predictable revenue. Credits or overage pricing handles heavy users without subsidising them.

The key insight is that you need pricing that lets customers predict their bill. Unpredictable costs are the single biggest reason AI products lose enterprise deals. A customer who can't answer "what will this cost us next quarter?" will not sign a contract.

The model cost curve is on your side

There is one structural advantage that makes AI pricing easier over time: inference costs keep falling. What costs a dollar today will cost ten cents in eighteen months. This means your unit economics improve automatically as the underlying models get cheaper, even if you hold prices flat.

The implication is that locking in annual contracts at today's prices is often a good deal for founders, because your costs will fall while your revenue stays fixed. The founders who understand this are structuring deals accordingly. The ones who don't are leaving margin on the table or worse, signing deals they'll regret when usage scales.

Pricing is always a proxy for your understanding of how customers get value from your product. If your pricing feels wrong, the first question to ask isn't "what model should we use?" It's "do we actually understand where the value lands?"