The useful unit in an AI medical simulation business is not a token. It is a completed learning encounter with feedback that the user can inspect and use. Token prices matter, but a model that ignores speech, abandoned sessions, retries, content review and support can make an expensive service look deceptively cheap.
The worked model below is an original hypothetical exercise constructed on 19 September 2026. None of its usage or cost assumptions are iatroX operating data, model-provider prices or estimates of a competitor's margins. Its purpose is to show how the economics change when the denominator is defined properly.
Define what counts as completion
An encounter can start when a user opens a case, speaks the first sentence or reaches the first clinical exchange. Those are different events. Completion might mean reaching the case endpoint, receiving feedback or reviewing that feedback. Choose a definition appropriate to the product and retain the intermediate events.
For this exercise, a completed case means an encounter that reaches its planned endpoint and successfully delivers feedback. A case abandoned after several minutes still incurs costs, but it does not enter the completion denominator. A completed case that requires regeneration of unusable feedback counts once, while all associated processing costs remain included.
This distinction prevents a favourable denominator from concealing friction. Counting every start as a completed learning experience would lower the reported unit cost without improving either the product or its economics.
Build the cost map from the actual workflow
Separate the stages that consume resources: audio capture and transcription, generated speech where used, model inference, case-state management, source retrieval, feedback generation and storage. Then add retries, support, payment processing and human work at the appropriate level.
Not every cost is proportional to encounter length. Some feedback work happens once per completed case. Some infrastructure remains available regardless of use. Case authoring and clinical review may support many future encounters, but updates and corrections still require resources. Keep these distinctions visible rather than forcing everything into a single per-minute number.
A vendor invoice can also combine several products. Allocate shared costs consistently, documenting the method. Otherwise, a simulation business may appear profitable simply because its inference or support costs have been assigned to another part of the platform.
A deliberately transparent base case
Assume 1,000 starts produce 800 completed cases. Across all starts, including abandoned attempts, users generate 12,000 audio minutes. The illustrative variable-cost ledger is:
| Hypothetical cost component | Assumption | Cost |
|---|---|---|
| Speech processing | 12,000 minutes at an assumed £0.015 per minute | £180 |
| Inference and orchestration | Assumed total across all starts | £240 |
| Feedback | 800 completed cases at an assumed £0.10 each | £80 |
| Retries and attributable support | Assumed total | £120 |
| Total | All four components | £620 |
Cost per start is £0.62. Cost per completed case is £620 divided by 800, or £0.775, approximately £0.78. Both figures are mathematically correct; only the second answers the stated unit question.
Now assume completion falls to 600 while the same total cost is incurred. The cost per completed case becomes approximately £1.03. Nothing about the token price changed. More expenditure was consumed by activity that did not reach the defined learning endpoint.
Test the expensive behaviours, not only the average user
Long encounters, repeated attempts and detailed feedback can all be valuable learning behaviours. They should be modelled rather than treated as misuse because they cost more. The product question is whether the learning experience remains useful; the commercial question is whether the subscription can support the distribution of use.
Run separate scenarios for a short text encounter, a longer voice encounter, a case requiring feedback regeneration and a user practising intensively before an examination. Do not assume the average across all registered accounts describes the cost of serving an active subscriber.
Also distinguish repeated meaningful practice from technical duplication. A user retrying because audio failed is different from a user deliberately practising an improved consultation. Both generate expenditure, but they call for different product responses.
Keep contribution margin separate from full profitability
Suppose a hypothetical content-review programme costs £2,400 and is allocated across 12,000 completed cases. That adds £0.20 per completion under that allocation. If only 6,000 cases occur, the allocation doubles. This is an accounting assumption, not a cash cost that literally appears each time a case is completed.
Payment costs, taxes where applicable, refunds, acquisition, general engineering and organisational overhead require their own treatment. A positive margin after model inference is not the same as positive contribution after support, and neither is the same as operating profit.
For an annual subscription, compare costs over the entitlement period rather than matching the full upfront payment against one quiet month. For monthly access, examine whether acquisition and onboarding costs are recovered before an examination-related cancellation. Retain the difference between cash collection and revenue recognition.
What a bundled price does and does not tell you
The September 2026 iatroX product brief states £99 paid upfront for a year, equivalent to £8.25 a month, or £29 monthly, covering questions, Socratic Tutor, planning, simulations and CPD together. Those published prices describe the customer offer. They do not reveal the underlying cost per case or the proportion of usage in each feature.
This article is published by iatroX and applies the same scrutiny to its own proposition. A bundled product should be evaluated against the several learning methods one customer genuinely uses for a relevant goal, not against an imagined customer sitting unrelated examinations simultaneously.
Before making an investment or pricing conclusion, request defined cohorts, actual cost attribution and sensitivity analysis. The hypothetical ledger here demonstrates the questions to ask; it cannot establish whether any named supplier has attractive or unattractive margins.
Frequently asked questions
Why not report cost per token as the main unit?
It excludes other resources and does not indicate whether the user completed a useful encounter. Keep token costs inside a wider workflow model.
Should abandoned cases be excluded from costs?
No. Include their costs even when they are excluded from the completed-case denominator.
Does iatroX's subscription price reveal its simulation margin?
No. Margin requires actual usage, costs and a consistent allocation method, none of which are published as operating data in this exercise.