The AI acceleration gap

Artificial intelligence does not cancel the structural load on conformity assessment. It changes the relative speeds of creation and assessment, and unless the assessment environment improves, widens the gap between them.

That claim is easy to overstate. AI is not magic. Certification timelines will not collapse because a language model can draft a section of a design-history file. The sharper story is system dynamics: tools that accelerate design, software, literature synthesis, and documentation increase the volume and tempo of what can enter a Notified Body queue, while assessment remains scarce-expert, evidence-intensive, and liability-bearing. At the same time, AI-enabled devices themselves raise novel evidence questions. Both effects stack on top of the base load already documented for MDR and IVDR.

The acceleration gap is not “AI versus regulation.” It is creation and documentation capacity outrunning usable expert time, unless review infrastructure keeps pace without automating judgment.

Layer one: creation accelerates

Generative tools compress work that used to gate how fast teams could iterate. Industry analyses of generative AI, outside the regulated medical-device certification literature, and therefore only directional here, place large productivity potential in software engineering, product research and design, and related knowledge work, including generative design and first-draft production of complex text (McKinsey gen-AI report, 2023). The directional implication: more design cycles, more software variants, more literature synthesis, and more technical prose can be produced per unit of calendar time.

For manufacturers, that can be a genuine advance. For the assessment system, it is also a demand shock in waiting. Queues already show applications ahead of certificates and multi-year average journeys under the Regulations (series A). Anything that increases the rate at which packages, variants, and revisions can be assembled, without a matching increase in usable reviewer capacity, tightens the constraint.

Layer two: assessment does not automatically accelerate

Conformity assessment is not a typing problem. It is the application of scarce technical and clinical expertise to evidence under rules that exist because mistakes harm patients. Team-NB’s 2025 survey still places 70% of new MDR certificates in time bands of thirteen months or more; manufacturer surveys report technical-documentation and QMS averages measured in many months (Team-NB Survey 2025; MedTech Europe 2024). Those timelines reflect completeness friction, multi-phase process, expert scarcity, and the cost of defensible judgment (series B).

Faster drafting can even increase assessment load if structure does not improve with speed. A generative assistant that produces confident, fluent text without coherent evidence linkage can expand the volume of incomplete or hard-to-navigate material arriving at completeness check, the very tax Notified Bodies already report at high rates (series B). Acceleration without quality of structure is not throughput; it is noise with better grammar.

Layer three: AI-enabled devices raise assessment complexity

Separate from generative tools used to build products is AI and machine learning as product function. Those devices bring their own evidence burdens: training and test data, performance across populations, change control for learning systems, transparency of intended use, and clinical evaluation that may not fit legacy templates. That work lands on the same finite expert pool already handling the broader MDR/IVDR transition.

Team-NB’s 2025 sector survey treated AI/ML as a tracked particular pathway and reported a shifting pattern: AI/ML submissions down by roughly 50%, while AI/ML certificates rose by roughly 40% year-on-year (Team-NB Survey 2025 press release). Absolute application and certificate counts for that pathway are not published as a clean census in the materials we cite; the directional signal is enough: AI-enabled products remain inside the pipeline while the base backlog and multi-month clocks still exist, and dual-regime questions under the EU AI Act will add further designation and evidence complexity.

So the system faces a double bind: more and faster packages can be produced with generative tools, while harder and more novel packages arrive for AI-enabled devices, both against multi-year assessment realities and incomplete intake patterns already in the data.

The system dynamics

• Manufacturer throughput can rise as design, code, and documentation assistance improves.
• Assessment complexity can rise with AI-enabled functions and denser evidence packages.
• Expert capacity grows slowly: qualification, liability, and judgment do not sprint with model releases, and 2025 showed staffing can also contract.
• Without better structure, queues lengthen, variance under load increases (series D), and usable expert time shrinks further into search and rework.

Two responses, and only one is honest

The tempting response is to declare that models will certify devices: that generative systems will replace Notified Body judgment, stamp certificates, and “solve capacity.” That response confuses fluency with accountability. Conformity assessment exists because some conclusions must be owned by qualified professionals under institutional quality systems. Automating the signature would not expand capacity; it would relocate risk into opacity.

The honest response is narrower and harder: use AI where it can make expert work more legible and less wasteful: structure intake, retrieve and link evidence, surface gaps, assist drafting of source-cited analysis, while leaving disposition, clinical and technical evaluation of the evidence, and formal conclusions with the human assessor. That is capacity infrastructure, not autonomous certification.

• No autonomous conformity decision: models do not issue certificates or replace the qualified professional’s call.
• No invisible basis for a conclusion: assisted text remains inspectable against sources; uncertainty stays visible.
• No substitution for qualification: tools do not mint reviewers; they change how reviewers spend hours they already own.

Expand capacity without automating judgment. That is the only AI story about conformity assessment that can stay true under pressure.