The strong rivalSimplicity and predictability have value

Buyers often prefer a seat, fixed subscription or capacity commitment because it is budgetable, comparable and legible to procurement. Sellers benefit from simple quoting, billing and revenue recognition. An elaborate value-sharing model can cost more to measure, negotiate and dispute than it improves alignment. Complexity should earn its place.

Use the simplest metric that keeps value, marginal cost and risk within tolerable ranges for the cohort. A low-variance drafting assistant may suit a subscription with a fair-use boundary. A predictable batch service may suit reserved capacity. A high-volume automation may need metering and a spend cap. Outcome sharing belongs only where both parties can observe the outcome, agree a baseline, isolate contribution and wait for the relevant period.

Simplicity is not concealment. A flat plan still needs an internal cost envelope, tail assumption, quality class and change rule. Customers need a forecastable maximum and a plain invoice. Providers need protection against adverse selection and runaway use. When a metric cannot be explained at the buying table or reproduced from event records, it is not simple in practice.

Original constructThe seven-layer AI Price Contract Stack

The AI Price Contract Stack is this article's synthesis. Each layer answers a different economic question, and lower layers may force a change above them. The result is a contract architecture, not a maturity score.

Metric distanceChoose proximity to value without losing auditability

Seats are easy to forecast when human access bounds use. They fail when one seat orchestrates machine volume or when value varies radically by role. Compute and tokens track a provider input and can support pass-through economics, but concise valuable work may use less than verbose low-value work. Billing raw implementation detail can also discourage efficient product design.

Actions are measurable but can reward splitting, merging or retry behaviour. Tasks move closer to customer work if completion has a stable boundary. Cases can align with a business process when duplicates, reopenings, appeals and partial completion are defined. Capacity sells availability or throughput and fits customers that value entitlement more than exact consumption.

Outcomes are nearest to value but hardest to attribute. They become unsuitable when the provider cannot observe the result, customer actions dominate, the lag is long, the baseline is disputable or optimisation could harm affected people. A hybrid can place an auditable operational unit under a small value-linked true-up while preserving a predictable base.

Score candidates qualitatively against causal proximity, auditability, forecastability, gaming, measurement cost, lag and risk. Do not average these dimensions into one number. A metric can be excellent on measurability and unacceptable on incentives. The contract should document the chosen trade and the evidence that would reopen it.

Economic envelopeModel distributions before setting the price

Cost is a floor and constraint, not proof of willingness to pay. Build a provider cost distribution by unit and cohort. Include routing, context, output, tool calls, retrieval, storage, evaluation, review, support, failures, security response, compliance and shared capacity. Separate costs the product controls from volatile upstream rates, and parameterise official rate cards rather than freezing today's figures into a durable contract.

Build the customer value distribution independently. Start from an alternative, not from the seller's cost. Estimate time, capacity, revenue, loss avoidance or decision improvement only where the causal route is evidenced. Apply adoption and capture: unused capability creates no realised value, and value retained by another participant cannot be charged twice. Show a range and identify who receives surplus.

Then stress normal and tail cases. Concurrency can matter more than total monthly use. Seasonality can make a plan profitable annually but operationally impossible at peaks. Human review and incidents create discontinuities. A rare high-consequence case can dominate expected loss. Margin should be shown as a range under volume, routing, quality and control assumptions, not a precise percentage that hides variance.

Cohorts must remain distinct when their economics differ structurally. A low-risk internal assistant and a consequential external agent may use the same model while requiring different evidence, latency, review, insurance and remedy. One list price can conceal cross-subsidy and adverse selection. Segment by value route, usage shape and risk class before discounting.

Contract mechanicsTurn the ranges into a governable commercial promise

Begin with a cohort sheet rather than a universal price list. Describe the customer's problem, current alternative, affected workflow, users, machine actors and people who bear consequences. State the value range under low, expected and high adoption. Name the customer actions required to realise it. If the result depends on process redesign, clean data or local sales effort, those contributions belong in the baseline and attribution rule rather than being credited entirely to the product.

Next define the billable event as an operational state transition. A completed case might require validated input, one authorised attempt, a result inside the quality envelope and delivery to the named system. Decide how retries, duplicates, reopened cases, partial work, customer cancellation, provider failure and abstention are treated. A sentence such as one case equals one request is not enough when software can issue requests automatically or when a single business case creates several technical calls.

Measurement rights should be symmetrical enough for a dispute. The provider needs event records to create the invoice; the customer needs a readable statement and the ability to challenge classification. Preserve identifiers, timestamps, version, service class and exclusion reason without collecting unnecessary content. Define the evidence hierarchy when records disagree, the time allowed for a challenge and the person authorised to resolve it. An opaque proprietary meter turns every commercial conversation into a trust negotiation.

Separate entitlement from consumption. Reserved capacity pays for readiness, concurrency and resources held whether or not the customer uses them. An included allowance can reduce invoice noise. Metered overage pays for variable operational work. A quality adjustment addresses a defined service failure. A value true-up shares a separately observed outcome. Keeping these functions distinct lets each line answer a customer question and prevents one metric from carrying incompatible meanings.

Design the cap as an operating control, not only a finance control. A hard stop can create harm if it interrupts a consequential workflow. The contract may instead route excess work to a slower model, queue it, request customer approval, fall back to a non-AI process or preserve essential actions while pausing discretionary ones. State who receives alerts, how much forecast window exists and which use can continue. The cap must protect budget without creating an undisclosed denial of service.

Commitments should purchase something identifiable: capacity, price stability, implementation support or a lower unit rate supported by predictable volume. A minimum that merely transfers demand risk to the customer can destroy the value argument. Allow ramp periods where adoption is uncertain, then review actual distribution before renewal. If the customer must redesign a workflow to capture value, match commitment timing to the transformation rather than to the seller's quarter.

Tiers work when they correspond to real cost, quality or value classes. A lower tier might use delayed batch execution; a higher tier might reserve concurrency and defined response. Do not create artificial degradation solely to force upgrade, and do not imply that safety, privacy or required remedy are optional premium features. Every tier needs a coherent quality boundary, and the customer should know when routing changes the billed unit or evidence.

Reopeners manage changes neither party should price indefinitely at signature. Trigger them with observable bands: usage distribution, concurrency, upstream cost index, service scope, model family, regulatory classification, verification burden or loss experience. Define the evidence period and negotiation consequence. A reopener is not unilateral discretion to reset price whenever margin disappoints. Until the trigger is met and the process completed, the existing contract remains the reference.

Discount review should restate the stack after every concession. If unit price falls, ask whether commitment rises, scope narrows, support changes, risk class falls or expected efficiency is evidenced. Never finance a discount by silently removing evaluation, human review, security response or remedy. Record which assumption funds the reduction and how the contract changes if that assumption fails. This makes sales authority compatible with the product's control obligations.

Procurement comprehension is itself a design test. Ask a buyer unfamiliar with the build architecture to forecast three usage scenarios, identify the maximum spend, explain the quality promise and locate the dispute and exit terms. Ask provider operations to reproduce the same invoices from event data and show margin ranges. If either side cannot do so without the deal team translating hidden rules, simplify the unit or surface the missing term.

Learning belongs in the commercial mechanism because an AI product changes. A cheaper routing policy may reduce cost while changing latency, evidence or abstention. A stronger model may improve quality while raising variable cost. Contract review should compare the new version inside the promised task population, update the cost distribution and decide whether the change fits the existing band. Model novelty alone is not a price event.

Exit tests the honesty of the value claim. Define export formats, migration assistance, retained evidence, deletion, unfinished cases, open incidents and outstanding disputes. A customer should not need to abandon records or remedies to leave. Where switching work creates real provider cost, price it transparently rather than hiding it in lock-in. The final invoice must remain reproducible after service ends, and affected-party obligations must survive termination where applicable.

Failure economicsAllocate risk without disguising obligation

Risk allocation starts with the action path. Identify error, delay, rework, third-party outage, security event, privacy failure and harmful action; affected parties; reversibility; detection and remedy. Price only exposures the contract can validly allocate. Regulatory or other legal obligations do not disappear because a schedule labels them customer risk.

Seller absorption may fit failures inside the provider's control. Buyer responsibility may fit misuse outside documented conditions, provided the boundary is clear and enforceable. Shared mechanisms can address attribution or operational interdependence. Some uses remain unacceptable because loss cannot be bounded, controlled or insured. The commercial answer is then to narrow the product or make no offer.

Service credits compensate a limited service failure under agreed terms; they are not automatically full remedy for business loss or affected-person harm. Liability limits, exclusions and insurance are separate. Quality bands should connect price to task population, latency, evidence, abstention and control, not imply that a premium tier may purchase unsafe behaviour.

A risk-blind discount can win a high-consequence workflow while excluding verification, incident, support and liability cost. The apparent margin depends on not performing the controls the product promise requires. Commercial approval should therefore link discount authority to the quality and risk class, and show which assumption or service scope changes when price falls.

Commercial failure modesWhere one metric breaks

The negative control is a no-offer decision. If the provider cannot observe the value event, reproduce an invoice, forecast customer exposure or bound tail loss, a paid discovery may be honest. A price invented to close the deal creates an economic promise that operations cannot fulfil.

Worked contractA hybrid invoice with an evidence thread

Consider a service that completes auditable customer cases. The contract starts with reserved capacity for predictable availability and the provider's standing control cost. It includes an allowance of completed cases defined by a shared event schema. Overage uses the same case unit and a published tier. A spend cap and routing rule protect the customer's budget without silently abandoning high-consequence work.

Quality is a gate, not a vague discount. The case qualifies only inside the stated task population, evidence and latency band. Abstentions routed as designed are measured separately from failed completions. A breach can produce a limited service credit, operational correction and, where applicable, other remedies. The credit is not described as extinguishing every consequence.

A small outcome true-up applies only to a cohort with an agreed baseline, observation period and attribution rule. Customer-controlled changes, unrelated campaigns and excluded conditions are recorded. The true-up has a floor and cap so neither party carries an unlimited model of the other's business. If evidence becomes contested, the base and case invoice remain reproducible while the outcome component pauses.

The invoice thread shows capacity reservation, eligible cases, included cases, overage, exclusions, quality adjustments, outcome evidence, credits and cap. It links each line to versions and dispute records. A reopener activates when model routing, upstream rates, service scope, risk class or usage distribution moves outside the agreed band. Exit terms cover migration, data portability and unfinished remedies.

Interactive instrumentAI Price Contract Designer

Set the cohort's actual economics. The Designer returns one of five routes with an architecture and unresolved exposure. It shows ranges conceptually and never claims one correct price.

Five routesWhat each commercial decision means

Use a simple subscription or reserved-capacity contract

Choose this for predictable, low-risk use where access or entitlement carries value and tail cost is bounded. Publish allowance, service class and reopener even if the invoice has one line.

Use metered usage with caps and quality guardrails

Choose an action, task or case unit when event boundaries are reproducible and volume drives cost or value. Add forecasts, alerts, routing, cap and quality gates. Do not expose raw tokens as value merely because they are easy to count.

Use attributable outcome or value sharing

Use this only with a credible baseline, observable outcome, accepted attribution, governable lag and capped risk. Preserve customer surplus and a dispute route. Pause the outcome component when evidence is contested.

Use a hybrid base, usage and risk contract

Combine reserved capacity, an auditable operational unit, quality band and bounded true-up when no single metric carries the economics. The additional complexity must solve named variance and remain invoiceable.

Do not offer this price until value or tail loss is bounded

Choose paid discovery, narrow scope or no offer when value is invented, invoices are opaque, severe risk lacks allocation, gaming is harmful, or tail cost cannot be survived. Commercial urgency does not make an incoherent contract valid.

Price-to-outcome ledgerEvidence across the contract life

Maintain the price-to-outcome ledger by cohort: value event and baseline; metric and measurement rights; provider cost distribution and shared allocation; customer usage and value distribution; quality and risk class; gross-margin range under normal and tail cases; attribution, lag and exclusions; invoice evidence; dispute and audit; reopener, cap and exit. Keep list price, invoice metric, value, cost, allocated risk and legal obligation distinct.

Before launch, reconstruct sample invoices from raw events and let customer operations challenge them. Test duplicate work, retries, partial cases, cancellations, abstentions, outages and time-zone boundaries. Ask whether the least expensive behaviour damages quality or reclassifies work. A metric that requires repeated manual interpretation will create dispute cost and sales friction.

During operation, monitor distributions rather than one average. Show included and overage use, concurrency, routing mix, review burden, incident cost, credits, cohort margin range and cap proximity. Do not rank customers as good or bad from margin alone; investigate whether the product design, sales promise or risk class is mismatched.

Use current provider rate cards only as dated parameters inside scenarios. Model changes, caching, batch service, reserved throughput and tool charges can alter upstream cost. The customer contract needs a reopener or routing strategy rather than an unqualified promise that today's input cost persists.

Price review should state whose value increased, which cost moved, what risk changed and whether the metric still shapes desirable behaviour. A higher list price without stronger customer surplus evidence is not value-based pricing. A lower price achieved by removing verification or remedy is not efficiency. A new fee tied to an unobservable event is not alignment.

At exit, preserve invoice and dispute evidence for the agreed period, complete unresolved remedies, support migration and apply portability and deletion duties as applicable. Do not use lock-in to compensate for a weak value route. A customer should be able to understand how the final invoice was produced and what happens to dependent workflows.

Range reviewDecide with scenarios, not a ceremonial midpoint

Build three internally coherent scenarios for every proposed cohort: cautious adoption, expected operation and stressed tail. Each uses the same value-event definition but different adoption, usage, routing, review and incident assumptions. Show customer spend, captured-value range, provider cost range and unresolved loss exposure together. The midpoint is not the answer; the shape and failure boundary guide the contract.

Run sensitivities one mechanism at a time before combining them. Change concurrency while holding monthly volume, then routing mix, outcome lag, verification burden, support demand and upstream rate. This reveals which uncertainty needs a cap, reservation, buffer or reopener. A dense model that changes every assumption simultaneously may look rigorous while hiding the contractual lever that matters.

Compare the proposed contract with the customer's alternative under the same conditions. The alternative may include labour, delay, error, existing software, outsourcing, foregone capacity and transition cost. Do not count sunk spending as new product value or assume every saved minute becomes cash. State whether value appears as released capacity, avoided loss, better service or an option to handle more work, and who can actually capture it.

Test margin after the promised controls, not before them. Include evaluation, abstention handling, review, monitoring, security response, support and remedy required by the quality and risk class. Show what happens when controls activate more often than expected. If positive economics require controls never to trigger, the price is financing a weaker service than the product claims.

Finally, rehearse disagreement. Give finance, product, sales, operations, risk and a customer representative the same receipt and ask which event is billable, which outcome is attributable, who bears a failed action and what happens at the cap. Resolve incompatible answers in the contract language and event schema. A signed document does not create alignment if operating interpretations still differ.

Review the range after live use without converting the first cohort into a permanent benchmark. Early customers may have unusual integration support, higher motivation or workflows selected for feasibility. Separate implementation learning from steady-state economics and preserve acquisition concessions as commercial decisions rather than product cost. When later cohorts differ, revise segmentation before averaging them together. A price that works only because one customer supplies expert labour, clean data or unusually patient reviewers is not portable to a cohort without those contributions.

Record the decision even when no price changes. Stable evidence may justify retaining a simple contract; growing uncertainty may justify holding the offer while more data arrives. The review owner states which scenario remains credible, which exposure is unresolved and the next observable trigger. This prevents routine renewal from silently accepting tail loss or charging for value that the product no longer creates. Test that record against the next invoice and operating review.

Evidence boundaryPricing research, market examples and scoped duties

Value-pricing research supports combining customer value, alternatives and cost while recognising organisational and judgement barriers. Software-pricing work supplies subscription, usage and two-part structures. Performance-based agreement research shows attribution, evidence, transaction-cost and risk-sharing constraints in its studied context. These sources do not establish willingness to pay for a particular AI product.

Official OpenAI, Anthropic, Google Cloud and AWS pages illustrate that upstream charging can vary by model, input and output, modality, caching, batch, service or reservation. Their live prices change and are not copied here. FinOps contributes allocation, forecasting, anomaly and unit-economic practice. Provider cost remains a parameter, not customer value.

NIST sources inform quality, risk, human-role and accountability costs, not a pricing prescription. EU AI Act duties depend on system classification, role, provision and application timing. ISO/IEC 42001 concerns management systems rather than a price formula. OECD principles were updated in 2024 and constrain acceptable value capture at principle level. Applicable law and contract advice remain fact- and jurisdiction-specific.

The seven-layer stack, price-to-outcome ledger and Designer are this article's synthesis. Test them through evidence-backed segments, scenario ranges, reproducible invoices, tail drills, dispute exercises and exit rehearsals. A commercial contract is coherent only when the priced service can be delivered, governed and explained under the conditions customers will actually create.