An attempt at the truth
How working with Cobra actually unfolds.
A description for regulatory affairs professionals — written without marketing language, because the truth is more useful than a pitch.
Tendai
Your guide
An attempt at the truth
A description for regulatory affairs professionals — written without marketing language, because the truth is more useful than a pitch.
Tendai
Your guide
Most descriptions of AI tools in regulatory affairs promise the same thing: time saved, work reduced, productivity multiplied. We have all read those pages. We have probably written some of them. They are not entirely false, but they are not the truth either — and a regulatory professional knows the difference.
This page is an attempt at the truth. What does working with Cobra actually look like, what is it for, what is it not for, and what does it ask of the people who use it.
All project examples on this page are illustrative: device details, study references and figures have been altered and anonymised.
Clinical evaluation under EU MDR is harder than it was a decade ago. The evidence base for any device class is larger. The methodological expectations from Notified Bodies are higher. The volume of literature published every year keeps growing. A single experienced reviewer can still produce a competent CER for one device — but doing it well, doing it reproducibly, doing it for several devices in parallel, doing it again two years later for the update, eventually exceeds what a single reviewer can do without losing methodological consistency.
This is not a problem AI can solve by replacing the reviewer. It is a problem AI can address by changing the structure of the work, so that the reviewer's expertise gets concentrated where it actually matters.
That is what Cobra is for.
First the constants
Before describing what changes, it is worth saying clearly what does not.
Regulatory responsibility does not change. The reviewer who signs off on the CER carries the same legal and methodological responsibility as before. No tool removes that.
The need for clinical and regulatory expertise does not change. A CER cannot be written by someone who does not understand the device, its intended purpose, its predicate landscape, and the regulatory framework. AI does not replace that understanding.
The methodology itself — MEDDEV 2.7/1 rev 4, Article 61 equivalence reasoning, the structure of an audit-defensible argument — does not change. AI does not invent new regulatory standards. It works within the existing ones.
Knowing what does not change matters. Many tools in this space implicitly suggest they are doing the regulatory thinking. They are not. What they are doing is reorganising the surface of the work. A reviewer who relies on a tool to do the thinking will eventually be caught — by the Notified Body, by an audit finding, by a difficult question they cannot answer.
What does change
Each layer has its own character. Some are entirely human. Some are platform-driven with human validation. Together they describe what cooperation between regulatory expertise and AI-assisted infrastructure actually looks like.
Before a Device Summary is written and before any platform work begins, there are conversations between the consultant and the device team. These conversations are where the foundational regulatory decisions are made — and they cannot be automated.
What is the intended purpose, and which indications belong inside it? What clinical and pre-clinical data exists in-house? Where, given this specific device, are the foreseeable methodological tensions the search will eventually surface? Which competitor devices, predicate references, transfer technologies will be examined?
These questions are answered in dialogue, not in a workflow. The output is not a document — it is alignment.
This layer is the most undervalued part of the process from the outside, because there is nothing to show. No screen, no extraction, no document. But it is where the project is actually shaped.
The decisions from Layer zero are translated into a structured Device Summary. This is not a free-form document. The Device Summary follows a defined schema — sections for intended purpose, indications, contraindications, patient population, device components, mechanism of action, predicate landscape, transfer technology references — because the platform expects specific information in specific places to function well.
Building this Device Summary is the second place where regulatory experience matters. Knowing what level of detail the platform needs in each field, knowing how to phrase an indication so that the search will catch the right literature, knowing which device descriptors will steer the classification rubric correctly — this is judgement work informed by familiarity with how the platform behaves.
It is a translation step: from the open dialogue of Layer zero into the structured input the platform requires.
From the Device Summary, the platform generates a Search Plan. The Search Plan contains the search queries themselves, the filters, the database choices — and also a regulatory framing that explains the methodological logic underlying the search. Transparency by design.
The generated Search Plan is then validated by the consultant and the CER author together. For a standard project, the validation is light. For a complex case, the regulatory section is substantially edited. That edit is human judgement, applied to platform output.
Once validated, the search itself runs on the platform. The relevant gain is not speed — it is reproducibility. The same logic can be re-run a year from now and produce the same baseline plus the new evidence that has appeared since. The methodology persists across time.
Each retrieved abstract is classified against a device-specific rubric. The rubric is built per project — it is not a generic template. The criteria reflect MEDDEV principles: indication fit, technology fit, outcome relevance, population match, evidence level, technical similarity, clinical setting similarity.
The platform applies this rubric to every abstract and produces a classification: pivotal candidate, supportive evidence, background context, exclude, or unclear. Every classification is accompanied by structured reasoning, traceable to the rubric.
This is the layer where AI does the most visible work — but the classification is a proposal, not a decision. A reviewer goes through the classifications, validates, overrides, annotates. The reviewer's judgement is captured alongside the AI's classification, and the resulting record is auditable in a way that purely manual screening rarely is.
For the studies that matter — the pivotal candidates, the studies that will carry argumentative weight in the CER — there is a fourth layer that no tool can fully automate. Each of these studies needs a methodological annotation: what does this study mean for the CER's argument, in which chapter does it belong, what limitations must be addressed honestly, how does it integrate with the equivalence strategy.
This is where the consultant's regulatory judgement is most visible. Reviewer notes, written per study, capture the methodological function each study serves. Critical limitations are flagged with concrete framing for the CER author.
The output is what the CER author actually uses while writing. It is not a list of studies. It is a structured working document with concrete recommendations.
Methodological function: Recent meta-analysis showing the 3D-printed material has no outcome advantage over standard implants.
How to address in CER: Lead with this finding honestly. Do not bury. Use as anchor for the three-pillar equivalence argument: material safety (established) + porosity mechanism (validated by transfer evidence) + PSI approach (landmark cohort).
Argumentation template: "The equivalence argumentation does not rest on a claimed outcome advantage of the new material, but on the combination of established material safety, a mechanism innovation validated through transfer evidence, and a patient-specific approach with documented advantage." (illustrative wording)
What this means for the reviewer's work
The honest answer is that working with Cobra is usually less work overall — but the saving is not uniform, and the most important changes are qualitative, not quantitative.
In a manual workflow, several costs accumulate that are not always visible until afterwards. The search strategy itself is built through trial and error: first attempt too narrow, then too broad, refined again, often with several iterations before it stabilises. Building a comprehensive search strategy is a specialist skill that not every CER author has at a high level — and when the skill is missing, the cost shows up later as gaps in the evidence base, as Notified Body queries, sometimes as audit findings. On top of the strategy work itself, manual projects accumulate organisational overhead: papers added by hand because the search was not systematic enough, search plans written into Word documents that are not cleanly reproducible afterwards, Excel sheets that need to be fully populated even at forty abstracts.
With Cobra, several of these costs disappear or shrink substantially.
The trial-and-error in search design is largely gone. The platform synthesises a comprehensive search from the structured input on the first attempt — not because it is smarter than a senior reviewer, but because it does not need the iterative cycles that human strategy-building usually requires. The lower bound on search quality is lifted: a less specialised reviewer ends up with a more comprehensive search than they would have built alone, and a senior reviewer reaches the same comprehensiveness in a fraction of the time.
The organisational overhead largely disappears. The platform structure is the organisation. Search plans, classifications, evidence mapping, audit trails — these are not separate documents that need to be maintained; they are the working surface itself.
A CER built on 117 carefully classified studies makes a different argument than one built on 40 — particularly in front of a Notified Body that examines the evidence base.
What does not shrink — and in some cases grows — is the strategic alignment work in Layer zero. For complex projects, the conversations between consultant and device team take longer, not shorter. A Custom-Made Device under Article 61(12) where the equivalence pathway does not apply, a novel material-design combination where the regulatory framing must be carefully constructed — these require serious thinking time at the start of the project. We do not try to compress this. The platform work that follows is faster precisely because the strategic foundation is solid.
So the right description is not "the same work, faster" and not "different work of equal volume". It is closer to: less time spent on the parts that did not require regulatory expertise to begin with, more time available for the parts that genuinely do, and a comprehensiveness in the search that a constrained manual workflow rarely achieves.
Where this is going
The honest answer about AI in regulatory affairs is this: today, working with AI in cooperation is the safe and high-quality way to do this work. The cooperation model is not a workaround for AI being too weak. It is a methodological position about where regulatory responsibility sits and how it gets exercised.
What AI brings to the cooperation is genuine intelligence. A modern language model designing a search strategy understands what the device's regulatory question actually is and constructs queries that reflect that understanding — not just keyword matching but contextual reasoning. The same intelligence is at work when the platform reads abstracts: it interprets methodological function, recognises study design, identifies what kind of evidence the paper provides for the specific regulatory argument under construction. Working alongside the platform on the difficult cases is genuinely collaborative work. It is closer to working with a very capable colleague than to operating a tool.
What AI does not bring — and this is structural, not a deficiency to be fixed — is the regulatory authority that signs the CER. A Notified Body audits a person, not a model. The reviewer who is responsible for the regulatory submission carries the legal and methodological accountability for what the document says. That accountability cannot be delegated to a language model, no matter how capable.
This division of labour is not a half-measure. It is the appropriate shape of AI use in regulated industries today. The intelligence of the platform makes the cooperation possible at high quality. The validation by the human keeps the regulatory authority where it must sit. Both contributions are real, and neither is decorative.
What to take from this
We are in a new era of regulatory work — not because the regulations have changed at their core, but because the means of doing the work have shifted in ways that are not going to reverse. A reviewer who steps into this cooperation needs a particular kind of openness: not naive enthusiasm for AI, not reflexive scepticism, but a professional willingness to engage with new working forms and to let their own role evolve as the tools evolve.
The expertise of the reviewer is not threatened by this — it becomes more visible, more concentrated, more clearly tied to where it actually matters. But getting there means letting go of some familiar patterns. The screening rhythm, the sovereignty of building everything from scratch, the comfort of methods that have worked for years.
That transition is real, and it is not always easy. We do not pretend otherwise. What we offer alongside the platform is the methodological scaffolding to make the transition defensible — and the cooperative working style to make it manageable. The rest is up to the professional who chooses to engage.