AI product cost comes from tokens, tool calls, search, storage, retries, and human review. That does not mean every internal cost should appear directly on the customer’s invoice.

A strong pricing unit is not a copy of the vendor’s bill. It must be predictable, controllable, and auditable for the customer while moving with customer value and product expansion.

The Four Layers of Pricing Design

Value metric: What increases when customer value increases?
Meter: How is that unit measured accurately?
Price curve: How do rate, tiers, minimums, and caps apply?
Guardrail: How are budget, disputes, refunds, failure, and abuse handled?

Stripe’s usage-based pricing guide explains that a usage metric should align with customer value, be predictable before purchase, and be measured reliably. Internal tokens may map to vendor cost while mapping poorly to customer value and predictability.

Compare Four Models

Model	Best Conditions	Advantage	Risk	Required Control
Seat-based	Repeated per-person use, collaboration, admin	Predictable and familiar to procurement	Agent use separates from headcount; low seat utilization	Role and active-seat definitions, true-up
Usage-based	API calls, documents, minutes, or work volume align with value	Low entry barrier and natural expansion	Bill anxiety, internal units, runaway loops	Real-time meter, alerts, cap, estimator
Outcome-based	Resolution, qualification, or recovery is verifiable	Strong value alignment, less payment for failure	Definition, attribution, disputes, gaming	Outcome contract, audit log, exception and cancellation rules
Hybrid	Base value plus variable usage or outcome	Balances predictability and expansion	Invoice complexity	Simple included allowance, overage rate, cap

No model always wins. Seats may fit collaboration, usage may fit high-volume APIs, and outcomes may fit work with a verifiable result.

Six Questions for Choosing the Unit

Score each candidate unit from one to five.

Criterion	Question
Value alignment	Does customer value rise when the unit rises?
Predictability	Can a buyer estimate the monthly range before purchase?
Controllability	Can the customer directly control use and budget?
Measurement and audit	Can both parties reproduce and verify the same result?
Margin stability	Can the price absorb cost variance, failure, and retries?
Resistance and gaming	Does minimizing the unit damage product value or quality?

If one unit cannot meet these conditions, consider a hybrid: base subscription plus included usage and overage, or an outcome fee.

Contract Grammar for Outcome Pricing

“Successful outcome” requires at least eight clauses:

Start event and end event
Success, partial success, failure, and human handoff
Prevention of duplicate charges within the same conversation or task
Judgment evidence from the customer or a third-party system
Cancellation, refund, and post-completion failure treatment
Dispute window and log retention
Factors outside either party’s control
Protection against increasing outcome count by reducing quality

Intercom’s explanation of Fin outcomes is one concrete example distinguishing service resolution, procedure handoff, and qualification. It does not prove that outcome pricing is universally superior. It demonstrates why definitions, duplication, failure, and reporting rules must be explicit.

The Cost Floor and Value Ceiling

Fully Loaded Cost per Verified Success

Model + tools and APIs + infrastructure + retries + monitoring + human review + support + recovery + security and administration, divided by verified successful cases

Customer Value Ceiling

Verifiable annual value from time saved, cost avoided, gross profit gained, and risk reduced

Price must stay above the cost floor and below the customer value ceiling. The FinOps Foundation’s Unit Economics capability likewise connects technology consumption with units of business value.

Illustrative Pricing Structures

Product	Value Event	Starting Structure	Reason
Team AI document tool	Active user and shared document	Base seat plus included expensive generation	Combines collaboration value and variable cost
High-volume extraction API	Page or document processed	Monthly minimum plus usage tiers	Clear meter and volume discount
Customer-support agent	Verified resolution	Platform fee plus successful outcome	Aligns value while recovering fixed integration cost
Autonomous research agent	Verified successful task	Subscription plus task credits and cost cap	Covers execution variance and review

These are structural examples, not price recommendations.

Run Pricing Experiments in Order

Interview 10–20 buyers and users about alternatives, value, and budget cycles.
Use historical or pilot logs to inspect the unit distribution and outliers.
Show low-, medium-, and high-usage invoice examples to test predictability.
Combine cost, customer value, and alternatives to create a price range.
Test only on new cohorts and compare activation, expansion, gross margin, support, and disputes.
Plan protection, contracts, and communication for existing customers before changing price.

Operating Dashboard

Activation and paid conversion
Usage distribution and overage share by account
Gross margin and cost per verified success
Bill shock and budget-cap frequency
Billing inquiries, disputes, credits, and refunds
Expansion, downgrade, and churn
Failure, human handoff, and outcome-judgment rate

It is easy to add a margin to token cost and publish the result. If the customer cannot predict or control that unit, the supplier is mostly transferring its own cost volatility.

A unit the customer cannot predict is rarely a good value metric.