AI is changing your market or your pricing: what to decide, and what buyers will check
AI hits a software business twice: it shrinks the work customers pay for, and it adds a variable cost of delivery. Seats, usage and outcome pricing with real price lists, gross margin under inference costs, what buyers count as recurring, the new contract clauses, and the roads open to a founder.
The short answer. AI hits a software or service business in two separate places, and founders often look at only one. On the revenue side, it reduces the work customers pay for: fewer seats, fewer hours, fewer tickets handled by people, and competitors who price by the task instead of by the user. On the cost side, it adds a variable cost of delivery that traditional software barely had: model inference. A pricing change that answers the first shock without modelling the second can grow revenue and shrink gross margin at the same time. Buyers and investors now read both, and they read the definitions behind the numbers: what counts as recurring, what sits in cost of revenue, and what a model provider's next release would replace.
1. What the market is pricing now
Three pricing logics now coexist, often inside one vendor's price list.
- Per seat. Predictable for both sides, and exposed when AI lets the customer do the same work with fewer people.
- Per unit of consumption. Credits, tokens, actions or conversations. Salesforce lists its Agentforce at $2 per conversation, or as Flex Credits at $500 per 100,000 credits, alongside per-user licenses.
- Per outcome. Intercom prices its Fin AI agent from $0.99 per outcome, charges for one outcome per conversation however many actions the agent takes, and counts a resolution when the customer confirms the answer helped or leaves without asking for more help. It also lets customers set usage reminders and hard limits.
The details of those two price lists carry the main lessons. Outcome pricing is only as good as its definition of an outcome, and the definition is a commercial term the customer will negotiate and audit. Consumption pricing needs controls, because finance teams will not sign an open-ended meter. And both vendors keep a per-user component: pure models are rarer than headlines suggest.
2. Choosing a model: the mechanics behind each
Seats
Keep seats where the value really follows people: collaboration, administration, accountability. Where AI replaces part of the user's work, seat revenue erodes from the customer's side at renewal, usually without warning. A buyer will model that erosion in its base case; you should model it first.
Consumption
Consumption aligns revenue with your own variable cost, which protects margin. It also makes revenue harder to forecast, makes net revenue retention swing with customers' activity, and moves part of the business from contracted to variable revenue. Committed spend with drawdown, included volumes with overage, and caps that customers control are the usual ways to keep it bankable.
Outcomes
Outcome pricing captures value best when the outcome is frequent, measurable inside your own system, attributable to your product and hard to dispute. It shifts risk to you: failed attempts cost inference and earn nothing, and model errors become revenue questions. Write the definition, the measurement, the audit right and the dispute process before the first invoice.
Seat erosion, in numbers
A hypothetical support customer pays for 200 seats at $50 a month, $120,000 a year. AI lets it handle the same volume with 30 percent fewer agents. At renewal it asks for 140 seats: $84,000, a 30 percent cut with no change in how much it uses the product. Multiply that by the share of your revenue in roles AI is absorbing, and you have the number a buyer's model will contain, whether or not yours does.
Outcome pricing, in numbers
The economics of an outcome price depend on how often the attempt succeeds. In a hypothetical case, each conversation costs $0.12 in inference and the agent resolves 60 percent of them; each billed outcome then carries $0.20 of inference, plus, say, $0.10 of hosting and support. At $0.99 per outcome, that leaves about a 70 percent gross margin. If the resolution rate falls to 40 percent and the model cost rises to $0.20 per conversation, each billed outcome carries $0.60 of cost and the margin falls to about 39 percent, at the same price.
| Hypothetical outcome economics | Gross margin |
|---|---|
| $0.12 per conversation, 60% resolved, $0.99 per outcome | ~70% |
| $0.20 per conversation, 40% resolved, $0.99 per outcome | ~39% |
Resolution rate is therefore a margin metric as much as a quality metric, and it belongs in the board pack next to revenue.
Hybrids and the migration itself
Most companies end with a platform fee plus usage or outcomes. The harder part is the migration of existing customers. It depends on contracts: price change clauses, renewal dates, notice periods and most-favored-customer terms decide how fast a new model can apply. Grandfathering protects retention and delays revenue. Measure price realization and discounting by cohort, because a new list price that sales discounts back to the old one has changed nothing except the paperwork.
3. Gross margin is where AI pricing succeeds or fails
Bessemer's State of AI 2025 describes two kinds of fast-growing AI companies. The fastest, which it calls Supernovas, average about 25 percent gross margin, often negative, trading margin for distribution. The steadier Shooting Stars average about 60 percent, slightly below software peers because of model costs. The PitchBook and NVCA Venture Monitor for the third quarter of 2026 adds that limited inference supply keeps costs high and leaves businesses whose unit economics depend on GPU availability and pricing exposed.
For an established software company adding AI, the practical levers are known:
- Routing. Send simple tasks to small, cheap models and keep frontier models for the hard ones.
- Context discipline. Cost scales with what you send the model; retrieval that sends less, caching of repeated work and batching of non-urgent jobs cut it.
- Build or rent. Hosting open-weight models can lower unit cost at volume and adds infrastructure and staff cost; the break-even is a calculation, not a belief.
- Vendor terms. Committed-use discounts, price protection and rate limits in your contracts with model providers are part of your gross margin.
- Pass-through. Usage or outcome pricing that moves with your own cost protects margin when model prices change in either direction.
One accounting point matters in any later deal: inference and the hosting that serves customers belong in cost of revenue. A company that reports them under research and development shows a better gross margin until a buyer's quality-of-earnings review moves them back, and then the multiple is applied to the corrected margin. Our note on the technology behind SaaS gross margin shows how buyers trace margin to architecture.
4. How buyers and investors read AI-era revenue
- What counts as recurring. Contracted subscription and committed spend are read as recurring. Usage above commitment is often discounted or excluded from ARR, and a run rate built from a strong month is challenged. Define ARR in writing and apply it consistently.
- Retention under the new model. Net revenue retention measured on consumption moves with customers' activity. Buyers look at cohorts, gross retention and the share of revenue under commitment, not only the headline percentage.
- Defensibility. The Venture Monitor notes that many investors doubt the durability of application-layer AI products as model providers extend their own. Expect the question in its plainest form: which part of the product would a model provider's next release replace, and what remains yours? Our note on an AI product or a wrapper sets out how buyers test it.
- Concentration on one provider. Dependence on one model vendor, its terms of service, its rate limits and its pricing is a supplier risk, priced like any other.
- Rights in data. Whether customer data may be used as the product uses it, including for training or evaluation, and whether customers were told.
Prepaid credits and revenue recognition
Credits sold in advance are deferred revenue until they are used. Credits that expire unused raise the question of breakage: when, and on what evidence, the company recognizes revenue for credits it expects never to be used. Under US accounting rules that estimate has to be supportable, and a buyer's accountants will test it. Aggressive breakage assumptions inflate current revenue and become an adjustment in diligence.
5. The customer contract has new clauses
AI features change what customers ask for and what buyers review: how customer data is used and whether it trains models, with an opt-out where required; who owns outputs; disclaimers on accuracy and the human review the customer is expected to do; liability caps that cover automated actions; the list of sub-processors, which now includes model providers; and, in some markets, indemnities against third-party claims arising from outputs. A buyer's counsel will read a sample of customer contracts for exactly these points. Inconsistent terms across customers, written ad hoc by sales, are a finding in themselves.
6. The options, side by side
- Reprice the existing product. Move to a hybrid at renewals, with a clear outcome or usage definition and margin modelled per customer.
- Repackage. AI as a paid add-on keeps the core price stable and measures demand; AI in every plan defends against competitors and costs margin.
- Own a layer. Proprietary data, workflow depth, integrations and distribution are what stay yours when models improve. Investment there is the strongest answer to the defensibility question.
- Partner. A model provider, a platform or a larger vendor in your market can bring distribution, and sometimes becomes the buyer.
- Sell, and when. Buyers pay for a transition that is proven in the numbers, not one that is promised. Selling before the shift starts captures the old model at its old multiple, if a buyer still believes in it; selling after a proven migration captures the new one. Selling in the middle, when seat revenue is falling and the new revenue is not yet visible, usually captures neither.
- Not yet. If customers are not moving and competitors are not pricing differently in your segment, measured preparation can be the right answer: cost per task, margin per customer, contract terms, and a migration plan ready for when they do.
7. What the technical review will ask
Pricing claims are tested in the system. A technical review of an AI-enabled product asks for cost per task and its trend, the share of requests by model, how evaluations measure quality before and after a model change, what happens when a provider is down or changes its terms, where customer data flows, and how much of the code was written with AI tools and reviewed by whom. Our notes on AI-written code in due diligence and AI coding agents in due diligence describe the code side.
8. What to measure from now on
A company that will be read by buyers in an AI market needs a small set of numbers it reports every month, defined the same way every month:
- cost per task or per conversation, split into inference, hosting and human review;
- success or resolution rate, and its trend after each model change;
- model mix: the share of requests served by each model and provider;
- gross margin per customer and per plan, not only in total;
- seats or users per customer over time, to see erosion before renewal;
- the share of revenue under commitment, and realized price after discounts.
These numbers answer most of a buyer's AI questions before they are asked. Their absence is itself an answer: it tells the buyer the company does not yet know its own economics.
Questions to answer before you change your pricing
- Which part of your revenue depends on the number of people your customers employ, and how fast could it shrink?
- What is your cost per task or per customer today, and how sensitive is it to model prices?
- Which outcome could you price, measure inside your own system and defend in a dispute?
- What do your contracts allow you to change, and when do the largest customers renew?
- Where are inference and hosting reported in your accounts, and would a buyer agree?
- Which part of the product would a model provider's next release replace, and what remains yours?
- If a buyer looked at you during the transition, what would it see in the numbers?
Where we fit
When AI changes the market, the decision is rarely only about the price list. It is about which business you will own in two years, and whether to transform it, partner, sell or wait. Our Second Opinion lays those roads side by side: two private calls and a written report on how buyers and investors would see the company, the options compared, including "not yet", and what to do first. Its fee is the same whatever you choose. You can download a sample report to see the form. When the question is technical, a review of the product's AI architecture, costs and code answers it with evidence.
We do not broker sales, introduce buyers or investors, or take referral or success fees, and nothing here is legal, tax, accounting or investment advice. Revenue recognition, contract terms and pricing law belong with your accountants and counsel.
Sources
- PitchBook and NVCA, Venture Monitor, Q3 2026: inference costs and investor doubts about application-layer defensibility.
- Bessemer Venture Partners, The State of AI 2025: gross margins of fast-growing AI companies.
- Salesforce, Agentforce pricing: per conversation, Flex Credits and user licenses.
- Intercom, pricing: Fin priced per outcome alongside seats.
- Intercom, Fin outcomes and resolutions: how an outcome is defined and counted, usage limits.
Questions
Should a SaaS company move from seat pricing to usage or outcome pricing because of AI?
Not automatically. Seats still fit where value follows people. Where AI replaces part of the user's work, seat revenue erodes at renewal, and usage or outcome pricing can follow value and protect margin. Most companies land on a hybrid: a platform fee plus usage or outcomes, migrated at renewals under what existing contracts allow.
How do AI vendors price outcomes today?
Intercom prices its Fin agent from $0.99 per outcome, charges one outcome per conversation, and counts a resolution when the customer confirms the answer helped or leaves without asking for more help. Salesforce lists Agentforce at $2 per conversation or Flex Credits at $500 per 100,000 credits, alongside user licenses. The definition of the outcome is the key commercial term.
What gross margin do AI companies have?
It varies widely. Bessemer's State of AI 2025 reports that the fastest-growing AI startups average about 25 percent gross margin, often negative, while steadier fast growers average about 60 percent, slightly below software peers. Inference costs, model routing, caching, vendor terms and pricing that passes cost through decide where a company lands.
Do buyers count usage revenue as ARR?
Often only in part. Contracted subscription and committed spend are read as recurring; usage above commitment is frequently discounted or excluded, and run rates built from a strong month are challenged. Define ARR in writing, apply it consistently, and show retention by cohort and the share of revenue under commitment.
Where should inference costs sit in the accounts?
In cost of revenue, together with the hosting that serves customers. Reporting them under research and development flatters gross margin until a buyer's quality-of-earnings review moves them back, and the valuation multiple is then applied to the corrected margin.
Is it better to sell before or after an AI pricing transition?
Buyers pay for a transition proven in the numbers, not a promised one. Selling before the shift captures the old model if buyers still believe in it; selling after a proven migration captures the new one. The middle, with falling seat revenue and new revenue not yet visible, usually captures neither.
