Answer first: AI SaaS pricing should start with the customer outcome and the full workflow cost, not a model token price. Include input and output use, repeated context, retrieval, tools, storage, failed jobs, support and provider variability; then choose a usage measure customers can predict and set margin guardrails that do not turn legitimate use into a surprise bill.

Key takeaways

  • Define the billable customer outcome before selecting a meter.
  • Measure full workflow cost, including failed and repeated work.
  • Use caching, routing and product limits as economic controls with quality guardrails.
  • Stress expansion, provider-price and heavy-user scenarios before launch.

Decision framework

Decision area Evidence to require Risk it controls
Model Input, cached input and output units Variable provider bill
Orchestration Retrieval, tools, reruns and validation Hidden workflow cost
Infrastructure Compute, storage, queues and observability Understated cost
Failure Aborted, low-quality and customer-repeated work Unbillable consumption
Service Support, review and contractual commitments Margin erosion

An AI software company should price against the cost and value of a completed customer workflow, not the cost of a single model call. Heavy usage can be commercially attractive, but only if revenue grows faster than inference, retrieval, human review, retries and support.

A flat subscription is not inherently wrong. It becomes dangerous when management cannot explain which customers consume resources, which tasks require repeated work and how the contract limits unprofitable usage.

Define the billable outcome

Choose a unit customers understand: a processed document, a resolved case or a completed analysis. Then define completion carefully. A generated draft requiring extensive correction may not provide the same value as a usable result.

Avoid making the billing unit identical to an internal technical measurement merely because that measurement is convenient. Customers may find tokens difficult to forecast, while a simple per-task price can still require a sophisticated internal cost model.

Provider pricing illustrates why costs vary. OpenAI distinguishes categories such as input, cached input and output, alongside other tool or processing charges where relevant. 1 Check current model and deployment terms rather than embedding a short-lived price table in the commercial strategy.

Calculate cost across the workflow

Measure retrieval, model calls, retries, tool use, storage, monitoring and human intervention.

Attribute costs to customers and task types. Include failed tasks that generate expense without a billable result.

In an illustrative service, a £2 task incurs £0.20 of automated processing, £0.60 of review and £0.15 of support allocation. Contribution before fixed costs is £1.05. If difficult tasks require three reviews rather than one, the average can deteriorate even when model prices fall.

Track the distribution. Averages conceal customers uploading unusually long files or repeatedly requesting regeneration. The busiest account may be the best customer by revenue and the worst by contribution.

Separate product friction from legitimate consumption

A customer retrying because the output is poor should not automatically be treated as valuable engagement. Record why tasks repeat: new work, user exploration, system error or correction.

Introduce quality measures alongside usage. A higher number of requests per completed outcome can indicate declining efficiency. If the company rewards teams solely for activity, it can encourage a product that consumes more resources without producing more value.

Reconcile product events with billed events. Record failed and retried tasks so finance can distinguish cost from a valid charge. This matters when a customer buys one completed outcome but the company incurs several calls to produce it.

Choose an understandable pricing structure

Possible structures include a platform fee with included usage, a per-outcome charge or a committed volume with overages. Test the structure against predictable customer budgets and the company's cost exposure.

Set limits transparently. Specify file size, concurrency, permitted automation, fair use and what happens when the customer exceeds the allowance. Hidden throttling can create support costs and damage enterprise confidence.

Commitment discounts should buy something valuable, such as predictable revenue or reduced selling cost. Do not discount heavily in return for a usage forecast that the customer is free to abandon without consequence.

Stress-test expansion revenue

Consumption businesses can grow within an account, but consumption can also fall when customers optimise workloads. Snowflake's annual report provides a public example of a company explaining consumption-sensitive economics and related risks. 2 Its experience is context, not a benchmark for an early-stage AI product.

Model a high-usage customer, a low-usage customer and a customer using an expensive workflow. Change model costs, quality, support and collected revenue separately. Show where the contract becomes unprofitable.

Separate economic experiments from customer surprises

Pricing tests need clear terms and appropriate communication. A customer should understand the unit, allowance and overage before usage creates a bill.

For a hypothetical document product, compare a per-page offer with a per-completed-document offer using the same task population. Long documents, failed extraction and repeated review may affect both customer understanding and cost.

Measure support questions and disputed bills alongside conversion. A price that looks more profitable only because customers misunderstand the meter is a weak foundation for retention.

Audit the unbilled workload

Internal demonstrations, free trials, debugging and customer-success work can consume substantial resources. Attribute them separately so paying-customer margins do not absorb every experimental cost without explanation.

Then examine whether free usage predicts paid adoption. A generous trial can be sensible when it produces suitable customers, but unlimited automated access may create expense without a credible route to revenue.

Set controls for abnormal consumption and review false alarms. A legitimate customer processing a backlog should have a workable way to explain it. A product that suddenly throttles an important workflow without notice can create an enterprise incident even when the financial limit is reasonable.

The pricing model therefore needs a communication model: forecastable usage, understandable alerts and an agreed route when needs change. Predictable bills and controlled costs support the same relationship, rather than being opposing objectives.

What should the pricing meeting decide?

Agree the outcome customers buy, the cost to deliver it and the protection needed against unbounded expense. Then choose the simplest pricing structure that reflects those facts.

Trace the heaviest customer's work from request through accepted result to collected contribution. That account is the best test of the pricing model. Growing demand creates value when the company can deliver it profitably and explain the bill.

A practical 90-day implementation plan

  1. Instrument cost and customer outcome at workflow and account level.
  2. Create heavy, median and light usage cohorts plus failure and support overlays.
  3. Model provider-price, model-routing, caching and growth scenarios against contribution margin.
  4. Pilot price presentation and spend controls with customers before changing the commercial contract.
  5. Review the ten most expensive workflows manually, separating valuable complexity from poor prompts, avoidable reruns, retrieval defects and product-generated retries.
  6. Give customers a usage forecast, threshold alerts and an escalation route, then measure disputed bills and support demand before wider commercial rollout.

Frequently asked questions

Should an AI SaaS product charge per token?

Only if tokens are understandable and controllable for the buyer. Many products need a customer-facing unit such as document, minute, case or completed workflow.

What belongs in AI contribution margin?

Include model and tool calls, infrastructure, failed work, support and other variable service costs attributable to delivering the outcome.

How can margin improve without a price rise?

Reduce repeated context, route work appropriately, cache where supported, prevent unnecessary reruns and improve first-pass success while monitoring quality.

Reporting by Shoreditch Talk