AI is fun to sell at the moment. The product team has more to show, demos move, and there's a new story to tell. 

But the cost side is where things get awkward. 

Every AI feature comes with a real per-query expense, and on flat-fee pricing that expense compounds in the background. 

By the time it shows up in the margin report, it's usually too late to fix without a full repricing.

The money is in the structure, not the price point.
_{- Ulrik Lehrskov-Schmidt}

 The real question, then, is how to price AI in a way that captures the value it creates - and that doesn't lose money on the heaviest users along the way.

AI pricing has become a large part of what we do at WillingnessToPay - half of our active engagements involved a large AI component at one point. The methods below have emerged from those projects (see also: our AI Products practice).

Why AI is hard to price

Two things move at once, and both move fast: what AI can do, and what it costs to do it.

Token costs are the clearest example: 

Input token prices fell roughly 97% through 2024 - one leading model dropped from $5 per million tokens to about 15 cents. 

But as models got cheaper they also got better, which led customers to ask them harder, more token-hungry questions, and new deployment patterns like on-device edge AI emerged at the same time. 

The per-unit cost was halving even as the total compute per customer kept climbing.

That double uncertainty is what makes CFOs uneasy. 

Push hard to capture market share now, and the business might overpay for customers who would have come anyway. Stay prudent and conservative, and a competitor could build the same thing twice as fast for half the price by Q3. 

Every concrete pricing decision inherits that uncertainty, which is why the strategic position has to be decided first.

The strategic bet: value vs. commodity

Two opposing bets sit underneath AI pricing. They're really two ends of a spectrum rather than a hard either-or, but every business needs to know where on that spectrum it stands.

1. The commodity bet

AI drives the price of everything down toward the cost of electricity - a view Sam Altman, CEO of OpenAI, has gestured at publicly. 

If the commodity bet turns out to be right, the play is to:

• Make money now rather than overpay to acquire customers
• Hold back on heavy product investment
• Keep dry powder so the business has cash to move when the moment comes. 

The pricing structure here looks like infrastructure: charge for inputs, and run like a utility.

2. The value bet

There's real, durable value to capture in AI products. If the value bet is right, the play is to win as many customers as quickly as possible (even at a loss in the early years) and monetize them later, once switching costs are high enough to keep them. 

The pricing structure prices for value: charge for the outcome.

Many businesses sit somewhere between the two ends, and a hybrid approach is common - a value-priced core with input-priced overage. 

Neither end is automatically right; they're bets on where AI is heading as a business model. 

But picking a position is the first decision, because it sets everything downstream: whether the meter sits on inputs or outputs, and whether the business is optimizing for breadth or capture.

Whichever bet a business places, the same question follows: do the economics still work? For traditional SaaS, the answer is built into the model. For AI, it isn't.

The margin trap: where AI strains SaaS economics

Traditional SaaS runs on fixed cost. Build the product once, host it, and each additional customer adds almost nothing to serve - which is how SaaS gets to ~80% gross margins.

AI doesn't fit that model. 

Every query has a real per-unit cost, and the more a customer uses the feature, the more it costs to serve them.

Price AI like traditional SaaS - flat fee, unlimited usage - and the structure starts working against the business. Heavy users erode margins while light users subsidize them. 

The result is what we like to call an adverse selection machine: the customers paying the most become the ones the business can least afford to keep.

OpenAI publicly struggled with heavy users on ChatGPT Pro as real consumption ran past projections, and Atlassian tied a 10% price increase directly to rising AI infrastructure cost. 

Both happened in public, in the last 18 months, in companies that know what they're doing.

So, if the CFO can't explain the AI cost curve - which customers cost what to serve, by feature and by use case - the business isn't ready to scale AI monetization. Every new deal could be making the company poorer, which is the question to answer before adding another tier.

Three elements of margin-safe AI pricing

Avoiding the trap takes more than awareness of it. There's a structural answer - three things, working together, that protect margin as AI scales.

Inputs vs. outcomes: tokens or value

The central design choice in AI pricing is what to count - and there are essentially two families to choose from. The distinction matters more here than in almost any other pricing question.

• Inputs - tokens, API calls, compute. 

These are resources consumed in the customer's value-creation process. They're a cost, and they're generic, so one input meter can run across forty different use cases. That makes input pricing a strong fit for a horizontal infrastructure play, and for the early learning phase when use cases are multiplying faster than they can be packaged.

AWS prices the plumbing this way because it can't realistically value-price.

If the offering is infrastructure, or a horizontal tool where the vendor genuinely can't tell what each customer uses the AI for, the inputs are the right meter.

• Outcomes - leads generated, tickets resolved, work completed.
  
  These are what the customer actually wants. Outcome pricing captures value and commands a higher price - the cost is commitment to a specific use case that can be measured.

Our recommendation at WillingnessToPay: start at inputs to learn, then shift to outcomes once a valuable value chain locks in. Token pricing is a strong learning model precisely because it's generic. Once the reliably-created outcome is clear, pricing the output is where the real value pricing lives.

Nobody actually wants to buy AI - or tokens, or hours.

Take a €250M services company that billed by the minute and suddenly had AI that could halve the hours needed to serve clients. 

The instinct was to charge for "AI hours" to recoup the lost €125M, but that misses the point. The actual question is what job those hours were doing for the customer in the first place - and how to price that outcome directly, leaning on AI to deliver it, rather than re-pricing a unit of input nobody wanted to pay for.

What customers are buying is the job, and the technology behind it is just the means.

The inputs-to-outcomes shift isn't theoretical. You can watch it playing out across SaaS right now, in a three-phase arc that most AI products move through whether they plan to or not.

The three phases of AI monetization

There's a visible arc in how SaaS companies have monetized AI over the last two years. The first wave launched it as a paid add-on or premium line item, while the current wave is folding it back into the core product and raising base prices instead. The rebundling is a pricing signal, and it tends to follow three phases.

The three-phase arc is the macro pattern. For a company deciding today where AI fits in its existing packaging, the question is more immediate: which route do you actually take?

How to introduce AI into an existing product

If a company already has packaging - tiers, modules, SKUs - there are several routes for adding AI. Two stand out because they're the most common (and the easiest to get wrong.)

AI as a premium feature in the top tier (the upsell motion)

Putting AI only in the most expensive tier is the intuitive move. If it works, the company gets paid a lot and AI pulls buyers up into the larger product.

For AI, this almost always struggles - and it struggles in a predictable way. 

The feature stays hidden from most of the customer base, while the cost concentrates in exactly the customers most likely to use it heavily. And a fast-moving capability ends up tied to the slowest-moving tier - the one customers commit to with annual contracts and change last.

Giving AI away free

The other route is to not charge for AI directly and let the existing metric do the work. A payroll tool charging per payslip processes more payslips and lowers churn - the AI never has to be billed separately, and that can be the right call. 

The catch, however, is the compute bill: without a cap on usage volume, the CFO is right to worry about writing customers a blank check, which is why a fair-usage limit usually belongs alongside it.

The principle behind the menu is the same as the rest of this page: decide the value-vs-commodity bet, protect margin with visibility and bands, and push pricing toward the outcome as the use case proves out. 

Where AI sits in the packaging [link to SaaS Packaging hub when live] is one decision, and what to charge for is another. The two have to agree, or the whole structure strains at the worst possible moment - when usage scales.