Updated 21 October 2026.
The first question for a clinical AI startup is not whether CE marking is cheaper than UKCA. It is what product the company is actually building, for whom, and with what claimed medical purpose. Choosing a route before defining those points can produce an apparently efficient regulatory plan for a product the company does not intend to sell.
Britain's announced move to indefinite CE recognition changes the market-access calculation. It does not eliminate the need to qualify the software, classify it correctly or generate evidence appropriate to its function. The following framework is intended to make that commercial discussion more disciplined, not to assign a classification to an unspecified product.
Define the output before discussing the mark
A software interface might retrieve documents, summarise an encounter, suggest a code or recommend a patient-specific action. These outputs can look similar on a screen while carrying materially different intended purposes and risks.
Write a precise description of the user, input, output, clinical context and limitations. Then compare that description with the product design, instructions and promotional claims. A marketing page promising diagnosis cannot be reconciled simply by placing "education only" in a footer.
The MHRA's software guidance is the appropriate starting point for qualification in Great Britain. For ambient products, its specific AVT guidance makes the boundary between administrative and medical functions particularly relevant. Avoid classifying an entire company when the actual issue is one defined function.
Do not import EU Rule 11 into Great Britain by shorthand
The European and British frameworks should be assessed on their own terms. The updated European software guidance explains EU qualification and classification. Rule 11 generally places qualifying software providing information for diagnostic or therapeutic decisions in Class IIa, with higher classes where the consequences specified in the rule justify them.
That is not the same as saying all healthcare software is Class IIa, or that every product with a clinician reviewing the output is Class I. Human involvement does not by itself determine qualification or class. Equally, adding a chatbot interface does not automatically make a static educational resource a medical device.
A startup should document the reasoning behind its position and obtain appropriate specialist input where the boundary is uncertain. That work is more valuable before the product promise has become embedded in sales contracts and investor materials.
Build the market map around actual customers
A business targeting several EU countries and Great Britain may find a maintained EU conformity route commercially coherent. A narrowly scoped Great Britain business may have different priorities. Neither should assume that the geographic reach of a route equals accessible revenue.
List the first credible buyers, their procurement requirements, the integrations they use and the evidence they expect. An additional country is not a sales opportunity until the product can function in that country's clinical and organisational environment.
The government's decision reduces uncertainty around future recognition. It does not translate the product, secure a local implementation partner or resolve national reimbursement differences. Treat regulatory access as one line in the market-entry plan rather than as the entire plan.
Compare lifecycle costs, not application invoices
A useful cost model includes evidence generation, quality management, conformity assessment where required, technical documentation, surveillance, change control and internal staff time. It should also include the opportunity cost of delays and the cost of maintaining more than one product configuration.
There is no defensible universal price for "CE marking an AI app". The scope, class, evidence gap and assessment arrangement can change the cost substantially. A precise-looking estimate that ignores those variables can be less useful than a range with transparent assumptions.
Ask advisers and assessment organisations to separate one-off work, recurring fees and tasks retained by the manufacturer. Compare the same scope across quotes. A low external bill may simply mean more work remains with the startup, not that the route is intrinsically cheaper.
Plan for what happens after the first release
Clinical AI companies often think about model updates as product improvements. The regulatory and assurance perspective asks whether a change affects the intended function, risks or performance that were evaluated.
A practical development plan should identify controlled versions of the model, prompts, retrieval corpus, interfaces and important configuration. It should also define what triggers additional testing and how an organisation can understand a change before it reaches clinicians.
Great Britain's surveillance requirements reinforce the importance of lifecycle responsibilities. The aim is not to freeze a product indefinitely. It is to make improvement traceable and to ensure that the evidence still describes what users are receiving.
Keep registration and local assurance in the budget
Recognition of a CE route does not remove MHRA registration or the relevant representation arrangements. Those obligations belong in the launch checklist even when the conformity assessment has been completed elsewhere.
For NHS-facing products, DTAC and applicable clinical safety standards introduce additional assurance work. A founder should identify who produces each document and who assesses the deployed workflow. Assigning everything to the hospital's digital team is not an implementation strategy.
A pilot also needs boundaries. Specify the setting, permitted tasks, data handling, review process, incident escalation and success criteria. A carefully bounded pilot can produce useful evidence; a vague "trial" may simply transfer uncertainty from the supplier to clinicians.
Three route-planning scenarios
For an EU-conforming supplier entering Great Britain, the immediate priorities are implementation of the recognition policy, current domestic obligations and local validation. Repeating all technical work should not be assumed necessary, but neither should local assurance be assumed unnecessary.
For a British startup planning international growth, compare the evidence and quality foundations required by each intended market early. Build reusable documentation where appropriate, while preserving jurisdiction-specific distinctions. Delaying that work can create avoidable redesign later.
For a reference or education business, begin with a clear intended-purpose boundary and reassess it when features change. Moving from explaining a guideline to recommending an action for an identified patient can change the analysis. A new commercial claim should trigger a product-governance review, not just a website update.
Regulation should support a focused product strategy
The strongest startup does not necessarily offer the broadest list of AI functions. It offers a coherent, useful product whose evidence and delivery match the promise. A narrower task can be commercially attractive when the buyer values depth and reliability.
iatroX's combination of reference and learning illustrates why product strategy should be described by use case rather than by the generic label "medical AI". The same clarity helps competitors, advisers and clinicians understand what evidence is relevant.
After the CE decision, the opportunity is to remove avoidable duplication from the business plan while strengthening the work that matters. A route is successful when it supports a maintainable product and a credible customer proposition, not merely when a certificate arrives.
Frequently asked questions
Should every UK medical AI startup now choose EU MDR?
No. The appropriate strategy depends on intended purpose, classification, markets, evidence and resources. Indefinite recognition changes incentives but does not produce a universal answer.
Does clinician review keep software outside medical-device regulation?
Not automatically. Qualification and classification depend on the product's intended purpose and functionality under the applicable framework. Human review is relevant to design and risk management, not a blanket exemption.
Can a startup avoid MHRA registration by using CE recognition?
No. Applicable Great Britain registration and other domestic obligations remain separate from recognition of the conformity route.
See how focused clinical workflows can fit together
Explore how iatroX works across clinical reference, reasoning practice and professional learning.