This document is a pricing checklist for a SaaS team adding AI to its product. It takes what about 300 executives at software companies building AI products told ICONIQ in a survey published in July 2026, on revenue, pricing, cost and quality, and turns it into ten questions to answer with your own numbers before you change how you charge.
1. AI stopped being a feature and became a revenue line
ICONIQ Venture & Growth, an investor in software companies, published its 2026 State of AI report on July 9, 2026. It rests on a survey of about 300 executives at software companies that build AI products, run in the second quarter of 2026 and compared with a similar survey from the fourth quarter of 2025. The respondents are chief executives and heads of engineering, AI and product. 87% are in North America. Their companies range from 1 to 5 million dollars in revenue to more than 1 billion.
On average, the executives report, AI products went from 32% of revenue in 2025 to a projected 42% in 2026. Gross margin on those products went from 45% to a projected 53%. ICONIQ puts the shift in one line: "A year ago, boards wanted to hear the AI strategy. Now they want AI unit economics."
These figures come from companies with revenue of up to more than 1 billion dollars, all of them already selling AI products, and the 2026 numbers are what their executives expect, not results.
2. The price is moving from access to usage and results
Subscriptions are still the most common way to charge. What grew between the two surveys is everything else. Usage-based pricing, where the customer pays per task, document or request, went from 35% to 42% of companies. Outcome-based pricing, where the customer pays for a result, went from 18% to 23%. Companies blend 1.7 ways of charging on average.
ICONIQ's reading is that pricing is being rebuilt to reflect actual usage and outcomes. The cost data explains why. As an AI product scales, the share of its cost that goes to people falls and the share that goes to inference, the cost of running the AI model on every request, rises. A seat price was designed for software whose cost per user barely moved. An AI feature costs more every time it is used.
| How you charge | What the customer pays for | The number you need before offering it | Where it breaks |
|---|---|---|---|
| Subscription or platform fee | Access to the product | What your heaviest accounts cost you per month, inference included | A few heavy accounts eat the margin of everyone else |
| Seat | People with access | How much of the work the AI does per seat | When the product does the work, value stops growing with headcount |
| Usage | Tasks, documents, requests | Cost per task on a bad day, when the AI retries and corrects itself several times | The customer cannot predict the bill, and finance pushes back |
| Outcome | A result both sides can see | A written definition of the result that both sides can audit | Disputes over what counts as done, and every failure is your cost |
3. The cost moves faster than the plan
Two thirds of the companies report better economics per request to the AI. Managing inference costs helps, so does routing requests across several AI models (they run about 3.3 on average), and so does revenue growth that spreads fixed costs.
The surprises come from agents, AI that takes several steps on its own to finish a task. ICONIQ reports one case from a company's internal tools, not its product: a workflow budgeted at ten cents a run drifted past a dollar fifty because the agents kept retrying and correcting their own work. Token spend, what the AI provider charges for the text the model reads and writes, was one of the biggest surprises in that internal spending. The same drift can happen inside a product that uses agents, and there the price has to absorb it.
4. Quality is the other half of the price
Quality control at these companies is still mostly reactive. 67% find problems through user feedback and 63% by watching the AI while customers use it. Adversarial testing, where a team tries on purpose to make the product fail before customers do, is used by 21%.
Data-protection guardrails are nearly universal. Defenses against risks specific to AI lag behind: 44% detect prompt injection, instructions hidden in the input that take control of the AI, and 38% have controls that stop data from leaking out through the model.
In their internal tools, close to half of the companies still need a person to step in on 30% or more of agent tasks, and the most common failure is a workflow of several steps that stops partway through. That figure is about internal tools, not products. It shows how often agents still need a person.
This matters for pricing because an outcome price turns every failure into your cost. A product whose quality is measured by customer complaints cannot safely charge by result. A product that is tested before release, with a known rate of human intervention, can.
5. Ten questions before you change how you charge
Answer them with your own numbers. A question you cannot answer is the first piece of work.
| # | Question | What a good answer looks like |
|---|---|---|
| 1 | What does one task cost you on a bad day, with retries? | The cost of your most expensive tasks, measured on what customers actually do, not on a test |
| 2 | Who absorbs a spike in usage: you, the customer or a cap? | Written in the contract and visible in the product |
| 3 | Can the customer predict next quarter's bill? | A range their finance team can defend |
| 4 | If you charge by outcome, is the outcome defined in writing? | Both sides can audit it on the same data |
| 5 | How do you find a wrong answer before the customer does? | Tests run before each release, not only feedback after |
| 6 | Have you tried to break the product on purpose? | Prompt injection and data leaks tested, results recorded |
| 7 | What share of tasks needs a person, and who pays for that time? | A measured rate, priced in or charged separately |
| 8 | Does the price hold if a customer's usage doubles? | Margin checked at twice the current usage |
| 9 | Can you switch to another AI model if its price or quality changes? | More than one AI model tested on your own tasks |
| 10 | Which margin are you aiming for, and by when? | A target on the AI product, reviewed each quarter |
A price set without these answers ends up set by the first cost nobody saw coming.
Sources
- ICONIQ Venture & Growth, 2026 State of AI Report: The Builder's Economy, July 9, 2026 (surveys fielded in the fourth quarter of 2025 and the second quarter of 2026). https://www.iconiq.com/growth/reports/state-of-ai-2026
Analysis and conclusions by yaab, based on the published reports listed above.