The market is not short of advice on pricing AI. It is short of a way to decide.
Go outcome-based, look at Intercom charging 0.99 dollars per resolved ticket.1 Per seat is dead. Usage is the future. Each claim comes with a real case attached, and each one is defensible in isolation.
The advice is not wrong. It answers a question you are not asking. Whether outcome-based pricing fits your AI product depends on variables the playbooks never examine: whether the outcome can be measured and attributed to it, whether a human or a customer-operated agent triggers the consumption, whether your buyer signs by credit card or through procurement, whether you are defending an installed base or attacking one. Change one and the answer changes with it. Which is why more than three quarters of AI providers say they are unsure how to price their agentic offerings.2
"It depends" is not an evasion. It is a decision problem, and decision problems can be structured.
The timing is not incidental. Across portfolios, AI capabilities are reaching the point where they must carry revenue instead of goodwill, and the market is mid-turn: 92 percent of vendors run subscription models today, but asked what they would choose for new AI features, usage-based intent jumps from 37 to 69 percent while the classic licence model collapses to 7 percent.3 The window in which AI pricing can be designed, rather than inherited from precedents set in a hurry, is now.
The unit of analysis: the product, not the vendor
The same company will correctly price an embedded copilot, an open MCP interface and an autonomous agent in three different ways. Any framework that hands a software company one answer for all its AI is answering the wrong question.
This is also the fastest way to end an internal deadlock. Most arguments about AI pricing are two people describing two different products.
Part 1
AI-based software products
By AI-based software products we mean products or features whose functionality rests on some form of AI technology, from large language models to classic machine learning. The pricing logic does not care about the architecture. It cares about how the capability meets the customer, and on that criterion three types dominate B2B software.
Why the typology matters
Each type carries a characteristic profile on the dimensions of Part 2, and a portfolio typically contains more than one. That is the first concrete reason why the unit of analysis has to be the product: the same vendor, facing the same market, will correctly price these three things in three different ways.
Two numbers give the types their commercial weight. Basic AI is being absorbed into the expected baseline fast, with 65 percent of DACH vendors that have an AI feature already including it in their smallest package.3 And agent economics are un-SaaS-like, with gross margins 10 to 20 percentage points below classic software, which is why an agent priced per seat quietly cannibalizes the seats it replaces.
Each type with its typical instances, from the chat function to the predictive-maintenance module of an MES vendor, and the profile it carries into the assessment: who the user is, whether compute costs are marginal, how attributable the outcome is.
Part 2
The strategic dimensions
Before any pricing question can be answered, the situation has to be measured. The dimensions fall into two classes, and the split carries real explanatory weight. Vendor-level dimensions are constant across everything a company sells: what kind of company it is, whom it sells to, what it wants to achieve. Product-level dimensions vary with every feature: who uses it, what it costs to run, how measurable its value is. The vendor-level factors stay fixed while the product-level factors move, and the pricing answer moves with them.
Eleven dimensions make up the assessment. The strategic objective is deliberately the odd one out: it is ranked rather than picked, because it is the tiebreaker. When the factual dimensions pull in opposite directions, the top-ranked objective resolves the tie. A vendor maximizing volume accepts margin dilution that one optimizing contribution margin must refuse, on identical facts. If the four cannot be ranked for a product, the pricing problem is a strategy problem, and it should be solved as one first.
| Dimension | How to fill it | States |
|---|---|---|
| Vendor-level dimensions · constant across the portfolio | ||
| Vendor type | Select one | AI-native · AI-first transformer · Established with AI features |
| Buying context | Select one | Self-service · Sales-assisted SMB · Enterprise procurement · Tender grade |
| Budget certainty need | Select one | Variance-tolerant · Corridor-plannable · Fixed-only |
| Strategic objective | Rank all four | Contribution margin · Revenue · Volume · Retention and lock-in |
| Product-level dimensions · assessed per feature | ||
| User of the feature | Select all that apply | A human · A vendor-controlled agent · A customer-operated agent |
| Compute cost share | Select one | Up to 20 percent · 20 to 40 percent · Above 40 percent |
| Primary value source | Select one, plus an optional second | Data intelligence · Automation · Interface · Accessibility and interoperability |
| Value uplift | Select one | Marginal · Relevant · Significant |
| Demand | Select one | Active pull · Latent pull · Push |
| Attribution | Select the furthest yes | Not quantifiable · No single accepted metric · Metric, not yet measurable · Measurable today |
| Value variance | Select one | Low · Medium · High |
One column of ticks per AI product or feature. Where a capability is reachable through several access paths, for instance a user interface for humans and an MCP endpoint for customer agents, fill it once per path.
Two warnings
- The dimensions are not independent.
Demand, value uplift and margin often move together. That is not noise to be corrected; the patterns in the whitepaper exploit exactly these correlations, which is why they can resolve four decisions at once.
- They do not map one-to-one onto pricing answers.
The assessment locates your product; it does not decide for it. What it gives you is the input: which options are gated, which trade-offs are genuinely live, and which direction the weight of your situation points to.
All eleven dimensions in full: the exact states with their definitions, and for each a note on what it steers and what it rules out. Among them the three attribution questions, which decide what a vendor can monetize; preference does not enter into it.
Part 3
The four decision levels
Part 2 measured the situation. Four things then have to be decided, and they are coupled rather than independent: bundle a feature into the host product, and its metric, its model freedom and its differentiation move with that one choice.
3.1Packaging: where does the product live commercially?
- Bundled into the existing lineup. Part of a package the customer already buys. In the base package the feature drives adoption and retention, in higher tiers it drives upsell.
- New bundle or package. Anchor of a new edition of the product line.
- Standalone product or add-on. Own SKU, own price, bought separately.
The core question: can the value be sold in isolation, or is it only credible in the context of the host product?
3.2Price model: how is the price charged over time?
- Subscription. A recurring fee for a defined scope, whether that scope is access, a package or a volume allowance.
- Pay-per-use. The price follows consumption. Fairness at the cost of predictability.
- Hybrid. Base plus allowance plus overage, or credit budgets. Predictability inside a corridor, participation beyond it.
Largely a question of risk allocation. Subscription places usage risk on the vendor, pure usage flips it onto the customer, hybrid splits it. Which is why hybrid is becoming the default: 59 percent of DACH vendors already combine more than one pricing metric.3
3.3Metric: what unit is the price attached to?
- Proxy metrics. Seat and its siblings, devices, sites, data volume, or a flat fee. They price the size of the deployment rather than its use: easy to forecast, blind to value.
- Resource and activity metrics. Tokens and compute units, then documents processed, calls, minutes. Meterable and understandable, but variable in what they are worth.
- Output and outcome metrics. Workflows completed and drafts accepted, then cases resolved and savings achieved. The strongest commercial logic, and the hardest preconditions.
Two things decide how far up a product may climb. Every candidate has to correlate with the value and be forecastable by the customer. And attribution sets the ceiling: a product climbs only as high as its furthest yes allows. Buyer preference already leans upward, with 86 percent preferring usage- or outcome-linked models over seat-based structures for AI,2 but preference does not move the ceiling.
3.4Differentiation: does everyone pay the same?
- Uniform price. One price for everyone. Simple, and it leaves the spread uncaptured.
- Structural differentiation. Fences on observable traits such as segment, size or use case, or self-selection through tiers, allowances and usage components.
- Customer-specific differentiation. Through the sales mechanism rather than the structure: negotiated prices, discount frameworks, deal desks. Flexible, but it needs governance and floors.
Worth its complexity only where value variance is high and the uplift is worth capturing. Enterprise and tender-grade contexts cap what is tolerable: a scheme procurement cannot audit will be negotiated away.
For each of the four levels, the dimensions that steer it and how each one tilts the choice: why AI rarely justifies a fifth package, how budget certainty sets the admissible corridor, the full metric ladder with the candidates each value source generates, and when a fence works rather than self-selection.
Part 4
How to decide
Let us be precise about what this part does not do. It does not derive a price from the assessment. There is no such derivation: the dimensions are correlated, the four levels are coupled, and real situations contain genuine conflicts that analysis alone cannot resolve. What can be structured is the path to the judgment.
Every conflict runs between two goals the vendor holds at once. What makes them workable is that each has a deterministic core, something that must happen regardless of preference, and a set of tactics that satisfy it.
- Value linkage versus provability.
The hard gate. The furthest attribution question answered with yes caps the metric ladder. No tactic overrides it, only sequencing: instrument for the target metric, carry the price on output or activity meanwhile, and migrate when the evidence lands.
- Predictability versus value linkage.
With fixed-only customers the variable share must be bounded. Tactics: base plus allowance plus overage, credit budgets, or time-shifted settlement so the agreed budget of the current period holds.
- Metering simplicity versus value capture.
With customer-operated agents even the weak anchor disappears: automated volume is an artefact of the customer's implementation, unbounded and gameable through batching. The value must move to the permission layer.
- Adoption versus margin.
With dominant compute cost, no construct with uncapped included usage survives. Not the flat price, not the unlimited allowance, not the uncapped base bundle, which is a flat price in disguise.
- Competitiveness versus monetization.
Once competitors ship a capability without surcharge and customers assume it is included, that layer is not separately chargeable, whatever it cost to build. Basic AI features have crossed this line.
- Simplicity versus capturing the spread.
In enterprise and tender-grade contexts the complexity cap binds: a scheme procurement cannot audit will be benchmarked against its most favourable cell and negotiated away.
- Monetize now versus learn first.
A sequencing decision, and it only works if the trigger is defined upfront: which evidence moves you to the target model. Without it, the interim state quietly becomes permanent.
Where a conflict survives its tactics, the top-ranked strategic objective decides. That is not a weakness of the framework. It is the honest residue of the problem.
Each conflict as a card with its deterministic impact and its mitigation tactics as separable variants, plus a summary table. Then six patterns, recurring product-in-context constellations from the integrated feature to the open interface, each resolving packaging, price model, metric and differentiation at once, with the variant that applies when one factor points against the card. And three anonymized client cases along one template: starting situation, pricing challenge, pricing solution design, rationale, each naming what was consciously given up and the trigger at which the decision is reopened.
Running the system
In practice the path is short. Fill the assessment for one product, or twice if two access paths exist. Walk the conflicts, note which ones your profile activates and what their deterministic cores already fix. Find your pattern and read the resolution. Where a factor points against it, apply the named conflict's tactics, and let the top-ranked objective decide the residue.
One step is skipped more often than any other, and it is the reason so many interim prices become permanent: write down what you decided and, for every sequenced decision, the trigger that reopens it.
Wrap-up
Everything here reduces to a change of question. Vendors arrive asking which pricing model is right for AI, as if the market were withholding a single correct answer. It is not. The useful question is narrower and harder: for this one feature, in this one situation, which options has reality already closed, and among the ones that remain, which trade-off am I willing to accept?
That reframing does real work. It explains why the same vendor lands on different answers for different products, and should. It explains why copying another vendor so often disappoints: their answer is correct for their attribution, their cost structure, their buyers, and yours differ.
Two disciplines matter more than any single choice. Decide per feature, not per company: a portfolio priced with one logic leaves money and adoption on the table in equal measure. And write down not only what you decided but the trigger that reopens it, because the most expensive pricing decisions are not the wrong ones, they are the interim ones that quietly became permanent because nobody defined when to revisit them.
None of this makes pricing AI easy. It makes it decidable.
A closing note
The honest way to hold this framework in mind is not "which model do I want" but "which death do I die for this feature, and which ones has reality already spared me". Every pricing choice gives something up: margin, adoption, simplicity, or the value left uncaptured at the top. The system does not remove the sacrifice. It makes the sacrifice conscious, and it stops you paying for one reality has already taken off the table.
- Intercom, Fin pricing; billed at 0.99 dollars per resolution.
- Simon-Kucher, "Why AI agents break traditional SaaS pricing models, and what leaders must do next" (2026), and "How to choose the right pricing model for AI agents" (2026).
- hy and OMR Reviews, "SaaS & AI Pricing Report 2026": survey of around 180 software companies and analysis of 4,197 OMR Reviews profiles.
One of Europe's leading experts on the monetization of software products, with more than 150 pricing projects behind him.