AI bot Kredo Telegram channel
Please leave your request and we will contact you
Required field
Required field
Неверный ввод
Согласие обязательно
Написать в WhatsApp Проверить досье в Telegram-боте
Bot vs. Expert: Responsibility in Device Registration
27.09.2026

Bot vs. Expert: Responsibility in Device Registration

In 2025, medical device registration in Russia entered a new stage: automation powered by artificial intelligence (AI) has stopped being just a convenient add-on and become an effective tool for a managed, accelerated path to a registration certificate.

Source: "Medvestnik" portal (original publication, in Russian)

Zones for automation: routine checks, data collection, formatting

The procedure for issuing a medical device registration certificate is governed by the "Rules for State Registration of Medical Devices," approved by Russian Government Resolution No. 1684 of November 30, 2024, which took effect on March 1, 2025. Roszdravnadzor is the government authority responsible for medical device registration. The new Rules substantially tightened requirements for the structure and electronic format of the registration dossier, which directly increased the role of automated checks at the preparation stage.

Algorithmic tools built on large language models are effective at tasks with formalized requirements, a rigid structure, and a direct tie to the regulatory framework. These include checking that a dossier is complete, that it's correctly formatted, that the information in it is current, and that the required presentation formats are followed. The formalization of requirements for electronic documents — including the requirement to submit PDF files with a searchable text layer — has further increased the value of automated quality control for the registration dossier.

In practice, this looks as follows:

First, algorithms perform routine completeness checks: they compare the actual set of files against the mandatory dossier structure for the specific device and risk class. This reduces the likelihood of errors caused by missing documents or sections, inconsistent wording, or the use of outdated versions of foundational materials.

Second, automated systems collect data and perform end-to-end cross-checks. They extract key parameters — intended purpose, area of application, composition, configuration, versions, and identifiers — and verify that the wording is consistent across the technical documentation, the operating materials, and the technical documentation extract.

Third, algorithms check formatting and how materials are presented. In 2025, Roszdravnadzor repeatedly published clarifications for applicants on the correct way to submit photographic images and operating documentation when amending a registration dossier, including when adding new product variants. The recurrence of such clarifications points to persistent, systemic formal inconsistencies in this part of the dossier.

Real-world examples illustrate this division of labor between algorithm and expert clearly. When registering class 2a diagnostic devices, algorithms flag in advance any missing mandatory sections of the technical documentation and inconsistencies in the list of accessories, then run a repeat checklist validation and cross-check parameters across dossier sections. At the next stage, the expert steps in, reviews the entire documentation, confirms the inconsistencies, and decides which elements need correcting and how the product variant should correctly be described.

For medical software, algorithms often flag mismatches between a device's stated intended purpose and the description of its functions across different documents — typically caused by using different versions of the descriptions or mixing marketing language with regulatory wording. The expert brings such wording into line with current regulatory requirements.

A similar approach applies when amending a registration certificate: the algorithm checks that the package of images and operating documentation for a new product variant is complete and produces a checklist of inconsistencies, which the expert then uses to validate that the submitted materials meet the regulator's requirements.

Zones for expertise: interpreting ambiguous rules, resolving contradictions

Despite the existence of formalized regulatory requirements, medical device registration practice still involves a substantial number of situations where regulation is a framework rather than an exhaustive rulebook. Regulations set the overall architecture of requirements but don't describe every possible design solution, use scenario, or clinical context for a specific device in detail. As a result, parameters such as clinical efficacy, sufficiency of the evidence base, the acceptability of claimed equivalence, and the accuracy of the stated intended use cannot be determined by a purely formal document check.

In such cases, analyzing the context of the device's use becomes decisive. The same device can present a different risk profile depending on the clinical setting, patient category, mode of operation, and user qualifications. An algorithmic check can flag formal inconsistencies or logical gaps between sections of the dossier, but it cannot assess whether the clinical data presented is actually relevant to the stated intended use, whether it covers the critical use scenarios, or whether it reflects real-world conditions of use in medical practice.

This becomes especially important when confirming equivalence and sufficiency of evidence. A formal similarity in technical characteristics, or matching functional parameters, doesn't in itself mean equivalence in terms of clinical effect and safety. Assessing this requires professional judgment based on an analysis of clinical studies, post-market experience, and established regulatory practice. This is precisely where the line between automated checking and expert judgment lies: an algorithm can structure information and highlight potential risk areas, but the final conclusion on whether a device matches its stated intended use and is fit to be placed on the market remains the expert's prerogative.

Algorithmic processing of a registration dossier is effective at catching internal inconsistencies between its sections. For example, in some cases the stated intended use and the clinical effects described in the assessment don't fully match the device's technical characteristics or its instructions for use. Such discrepancies often arise from different teams working in parallel, the use of outdated document versions, or materials being adapted for different markets. An algorithm can flag the existence of a conflict, but determining its regulatory significance requires expert professional judgment.

The expert analyzes whether the discrepancies found are merely editorial or descriptive in nature, or whether they touch on critical questions of safety, efficacy, or device classification. This takes into account both the formal requirements of the regulations and Roszdravnadzor's established enforcement practice, including its typical requests and the reasons behind previous refusals or suspensions of registration. The expert also assesses strategic regulatory risk: in some cases, revising the clinical section requires additional studies; in others, changing the technical description or instructions for use may affect the device's stated intended use and risk class. An algorithm can highlight possible scenarios, but choosing the optimal strategy remains the expert's job.

The regulatory framework for AI-enabled medical devices deserves particular attention. In 2025, the regulator moved from a declarative approach to practical enforcement mechanisms, approving a procedure for automated reporting of such devices' real-world performance data. Developers and rights holders are now required to regularly submit data on usage frequency, malfunctions, algorithm errors, and possible incidents through dedicated information systems. According to Roszdravnadzor, 48 AI-enabled medical devices were on the state register in 2025, most of them developed by Russian companies.

This approach substantially changes how responsibility is understood across every stage of a device's life cycle. Registration is no longer treated as a final step, but as the starting point for ongoing monitoring of the device's real-world performance. This is especially important for AI-enabled devices, since their algorithms can be updated, retrained, and show a shifting risk profile as usage data accumulates. Under these conditions, the expert's responsibilities extend beyond preparing the registration dossier to include assessing the device's post-market management system, monitoring software updates, and evaluating the developer's readiness to handle incidents.

It's in this context that, by 2025, the human-in-the-loop model — keeping a human in the decision-making chain — has become the standard approach in medical device registration, particularly in software- and AI-related fields. Under this model, the algorithm performs preliminary analysis: it checks that the dossier is complete, cross-checks key parameters, and produces a structured report on any inconsistencies found. The expert interprets the results, separating formal issues from substantive ones, and decides how to revise the dossier and shape regulatory strategy. Final responsibility for the content of the documents, their interpretation, and the consequences of the decisions made rests entirely with the expert and cannot be delegated to an algorithm.

According to KREDO's project data, algorithmic preparation delivers a registration dossier that is up to 75% ready, while expert review typically identifies anywhere from 10% to 20% of items needing correction, depending on the device's complexity. These corrections include both technical errors and substantive edits related to clarifying intended use, the evidence base, and terminology. The boundary between the two zones of responsibility therefore stays clear: the algorithm improves efficiency and reduces the likelihood of formal errors, while the expert makes the management decisions and bears responsibility for the final version of the dossier submitted to Roszdravnadzor.

Conclusion

Algorithms can substantially cut the time needed to assemble a registration dossier and minimize the likelihood of formal errors, but removing the expert from the registration process is not possible. Automation is a tool for improving professional efficiency, while responsibility for interpreting requirements, resolving contradictions, shaping the evidentiary strategy, and final validation remains with the human expert.

Andrey Antonov, Founder and CEO of KREDO LLC

Frequently asked questions

Who is responsible for registering a medical device in Russia?
Roszdravnadzor is the government authority responsible for medical device registration. The procedure is governed by the "Rules for State Registration of Medical Devices," approved by Russian Government Resolution No. 1684 of November 30, 2024, which took effect on March 1, 2025.
What work can algorithms actually take over?
Automation is most effective where the requirements are formalized: routine dossier completeness checks against the mandatory structure for the specific device and risk class, data collection with end-to-end cross-checks, and checks of formatting and document presentation.
Why did automated dossier checks become more important in 2025?
The new Rules substantially tightened requirements for the structure and electronic format of the registration dossier, including the need to submit PDF files with a searchable text layer. That directly increased the value of automated quality control at the preparation stage.
Where does the expert still make the call?
Regulation sets a framework rather than an exhaustive rulebook, so clinical efficacy, sufficiency of the evidence base, acceptability of claimed equivalence and accuracy of the stated intended use cannot be settled by a purely formal document check. The final conclusion on whether a device matches its stated intended use and is fit to be placed on the market remains the expert's prerogative.
How does this division of labour look in practice?
When registering class 2a diagnostic devices, algorithms flag missing mandatory sections and inconsistencies in the list of accessories, and the expert then reviews the documentation and decides what needs correcting. For medical software, algorithms typically catch mismatches between the stated intended purpose and the description of functions across documents.

See also

Expert insights on medical device registration
Regulatory updates, case studies and checklists — in our Telegram channel.
Subscribe