Time horizon and confidence
Test the billable unit now; revisit pricing after 12–24 months of customer and cost evidence.
Confidence: Medium on the need for verifiable units; low on any single model becoming dominant. These are editorial judgements, not statistical probabilities.
Billing infrastructure is evidence of feasibility
Stripe's documentation distinguishes Billing Meters usage recording from its Metronome integration path. These are provider-specific implementation options, not evidence that a pricing model will succeed commercially. [A1]
Stripe also documents prepaid and promotional billing credits for eligible services. They have restrictions and are not general-purpose stored value or third-party payment balances. [A2]
Three units allocate uncertainty differently
Zopio's interpretation is that usage pricing fits products whose customers understand the consumed resource and can relate it to value. It can expose buyers to cost uncertainty when the same task requires different amounts of computation.
Charging per completed task can be clearer, provided completion is agreed. A hypothetical document-processing service could charge for a document accepted under defined quality criteria. Retries, rejected results and human corrections need an explicit treatment rather than becoming hidden invoice items.
Outcome pricing fits only where the outcome is observable and attribution is defensible. A collection assistant cannot assume that every payment following its message was caused by that message. Baselines, customer actions and negotiated terms matter. A completed task is not automatically an incremental business result.
A credible unit must survive an invoice dispute
Define the event that creates a charge, its customer and time period, and the evidence available to explain it. Agree how to handle duplicates, partial completion, cancellations and later corrections. The billing system needs a reproducible record rather than a summary generated after the dispute.
Show customers the connection between the service delivered and the invoice. A technical meter may support that record without being the unit shown to the buyer. Avoid promising outcome alignment if the contract ultimately charges for attempts regardless of result.
Keep commercial credits, cash receipts and recognised revenue conceptually separate. Their accounting treatment depends on the contract and applicable rules; issuing credits does not by itself establish that revenue has been earned.
Protect the economics before scaling
Track contribution after model and tool costs, retries, human review and support. A flat task price can become unattractive if the distribution of task difficulty changes. Set scope boundaries and budget controls before encouraging unlimited autonomous work.
Pilot one customer segment and one unit with a clear acceptance rule. Compare customers' ability to understand charges with dispute frequency, repeat use and contribution. A hybrid model may be reasonable, but adding several pricing components can also make an understandable product harder to buy.
Who should change pricing and who can wait
AI providers with unpredictable delivery costs or repeated billing disputes should reassess their unit now. Providers whose subscription already matches value and sustains margins need evidence before replacing it.
We would favour outcome pricing more strongly when attribution and settlement are inexpensive and accepted by customers. We would favour simpler usage or task pricing where verification costs consume the proposed benefit. A provider's billing feature does not resolve that choice.
Choose a billable unit both parties can verify.
Distinguish completed work from attributable business outcomes.
Define retries, corrections and acceptance before charging.
Measure contribution after delivery, review and support costs.