If every competitor could use the same underlying AI model, a clinical product would still need to justify its content, workflow, evaluation, distribution and operational reliability. Those can be meaningful advantages, but none should be accepted merely because a company calls it a moat. Ask what evidence supports the advantage and what would disprove it.
Remove the easiest claim first
Imagine an original hypothetical supplier whose pitch centres on using a particularly capable model. For the thought experiment, assume a rival gains access to the same model under equivalent technical conditions.
What would the customer still miss if they switched? The answer might involve a well-designed workflow, legitimate specialist content access, reliable support or an established integration. It might also be less substantial than the pitch suggested.
This exercise does not predict that all models will become identical or that model development is unimportant. It isolates the parts of the business that should be evaluated separately from access to a component another company may also obtain.
Turn each claim into an evidence request
Use the following original framework for a product discussion:
| Claim | Evidence to request | Evidence that would weaken it |
|---|---|---|
| Distinctive content access | Current rights, scope and usable coverage | Rights exclude the claimed use or can disappear without a replacement |
| Better workflow | Observed task completion and user research | Users routinely export work elsewhere to finish the task |
| Strong clinical evaluation | Relevant methods, versions and limitations | Results concern a different task or obsolete configuration |
| Durable distribution | Retained use and renewable relationships | Access is broad but meaningful activity remains limited |
| Reliable operation | Incident handling, monitoring and recovery evidence | Important failures cannot be detected or reproduced |
The table is a due-diligence aid, not a validated investment score. A missing document is a question to resolve, not proof that a competitor has no capability.
Rights are useful only when they support the product's actual use
A partnership logo or a journal subscription does not establish permission for every form of retrieval, display or model development. Examine what the agreement allows, its duration and what happens if access ends.
MedlinePlus's permissions guidance, checked on 19 September 2026, provides a concrete reminder that even one website can contain material with different reuse conditions. The general lesson is to evaluate rights at the relevant content and use level, not to infer a blanket licence from access.
Content rights may support a valuable product, but exclusivity is not the only useful question. Currency, coverage, update processes and the ability to show the user the supporting source may also affect the actual experience.
Evaluate the workflow at its difficult point
A product demonstration often shows a clean input and a satisfying output. Ask to see the point where work becomes awkward: a missing document, a conflicting source, an interrupted session or a user who needs to correct an assumption.
For a learning platform, the important moment might be after a wrong answer. Can the learner understand the misconception and practise again? For a clinical workflow product, it might be reviewing and correcting a draft before it reaches the record.
A broad feature list does not establish that those transitions work well. Nor does a specialist interface necessarily establish durable advantage if users find an equivalent route elsewhere. Observe the complete task and ask customers what they would lose, not just which screens they like.
Separate evaluation capacity from a favourable result
A company should be able to explain how it tests the intended function, how versions are identified and how failures affect development. One impressive benchmark does not establish all of those capabilities.
For clinical applications, check whether the evaluation concerns the same population, task and product configuration as the claim. For education, distinguish supported performance, independent learning, retention and actual examination outcomes.
The joint medical-device change-control principles published through the FDA, checked on 19 September 2026, illustrate the importance of bounded changes and evidence in their specified device context. They are not a universal certification checklist for every learning application.
The analytical point is broader: an evaluation process that can detect a regression may remain valuable even when the underlying model improves. A result with no reproducible method offers less assurance about the next release.
Distribution is not the same as availability
A supplier can be available to a large professional audience without becoming part of anyone's routine. Examine the path from access to repeated completion of the intended task.
Ask which relationships would survive a change of model supplier. A contract, a teaching community and a search-driven audience each have different strengths and dependencies. None should be called durable without evidence of retention, renewal or repeated choice.
Also ask how concentrated the route is. A product that depends on one distribution partner may grow quickly while carrying a material dependency. That is not automatically a poor strategy; it is a condition the business should understand and manage.
Apply the same test to iatroX
This article is published by iatroX and includes its own proposition in the analysis. The iatroX methodology checked on 19 September 2026 describes source retrieval and checking, while the public learning offer combines questions, tutoring, planning, simulations and CPD.
Those are documented design and product features, not proof of defensibility or superior outcomes. A meaningful assessment would ask whether the combination helps a defined user complete a relevant task and whether that usefulness is sustained.
For a candidate preparing for one examination, the potentially valuable combination is explanation, practice and feedback around that goal. The number of unrelated examinations in the catalogue is not, by itself, a reason the candidate could not switch.
A conclusion by buyer question
A clinician should ask whether the product remains useful when the first answer needs checking or the learning task becomes difficult. An institution should ask whether implementation, evidence and continuity are dependable. An investor should ask whether those benefits support a sustainable purchasing relationship.
The same model can sit inside very different products. The defensible claim is not that the surrounding work is automatically valuable, but that its value can be demonstrated under conditions where model access alone cannot explain the result.
Frequently asked questions
Does using a third-party model mean a clinical AI company has no defensibility?
No: content, workflows, evaluation and customer relationships may add value. Each claimed advantage still needs evidence beyond the model's reputation.
Is a large user count enough to demonstrate a distribution advantage?
Not by itself: inspect definitions, activity, retention and the route through which users arrived. Availability and cumulative registration do not establish sustained use.
Does this framework identify an investment winner?
No: it organises evidence requests and possible weaknesses. An investment conclusion requires company-specific information and broader due diligence.
Discuss clinical AI evidence and product evaluation with iatroX Insights →