JUMO Regulatory Agent: evidence for regulated product imports
JUMO Regulatory Agent: evidence for regulated product imports
Importing a medical device, medicine, cosmetic or food supplement into another country creates a business problem long before the first shipment. Your team must establish how the product is classified, what evidence the destination market expects, which claims it can support, and who carries local responsibility. An approval elsewhere leaves important questions to resolve.
For management, this means uncertainty in the launch plan. For regulatory teams, it means finding the right rules, checking their dates, comparing them with the product dossier and explaining what the business must do next. Each formulation, intended use, label change and destination can create another review.
The JUMO Regulatory Agent supports this work with AI that retrieves regulatory evidence, follows a defined regulatory method and returns answers that experts can check against their sources. For a business considering its use, the decision is where that support can improve a real task while preserving accountable expert judgment.
The cost of regulatory work extends beyond the application
Regulatory affairs requires specialist time throughout a product’s life. Before launch, teams assess the pathway, coordinate evidence and prepare documents. After launch, they monitor changes, review labels, maintain registrations and respond to questions. An incorrect assumption can create repeat work or interrupt market access.
Credible sources show several dimensions of this burden:
- Work accumulates across products. In an October 2026 FDA notice on cosmetics information-collection requirements , the agency estimated 367,777 annual hours for specified reporting and recordkeeping activities across the respondent population. Initial product listings accounted for 330,337 hours. These are estimates for defined activities across respondents, not an individual company’s workload or the full cost of compliance.
- Consulting sits within a larger project. CIRS Group’s medical-device registration cost analysis separates consulting from testing, clinical trials and administrative charges. Published in 2015, it provides a historical breakdown of cost categories. Its procedural details should not be treated as current requirements.
- Rework can affect continued sales. In a June 2026 article on medical-device registration renewals , CIRS describes how missed deadlines, poorly sequenced changes and delayed awareness of updated standards can lead to certificate invalidation, renewed registration work and disrupted market access.
The practical question is how much expert effort goes into gathering and checking evidence, and how much remains available for decisions. JUMO’s supplied Value Map frames the business value around expert time, avoidable delay and preventable errors. Actual savings need to be measured in the client’s workflow.
Where AI can help regulatory teams
AI can help organise a question, search a large body of documents, extract relevant passages and structure an explanation. These tasks recur in pathway assessments, regulatory intelligence and reviews of proposed changes.
Regulatory work demands controls around those capabilities. A fluent answer may cite an outdated rule, confuse guidance with binding law or omit the condition that changes the conclusion. A source link helps only if the cited passage supports the answer and applies on the relevant date.
JUMO combines a curated evidence library with operating instructions for regulatory analysis. The intended benefit is less effort assembling the evidence and a clearer basis for expert review. Its supplied benchmarks measure answer quality; they do not establish a percentage reduction in working hours or launch timelines.
How the JUMO Regulatory Agent works
The agent connects to a compatible AI assistant, allowing a team to ask questions through a familiar interface. Behind that conversation sits a regulatory library and a method for using it. The inspected JUMO operating instructions describe five steps.
1. Identify the decision and the relevant facts. The agent classifies the request by regulatory job, such as intelligence, pathway assessment, dossier preparation, post-approval change or compliance analysis. It asks for facts that could change the answer and states remaining assumptions. Product type, intended use, import status and the applicable date matter.
2. Retrieve the evidence. The agent searches the library and reads the relevant article or section before citing it. Dedicated tools support classification lookups, exemption checks and searches of court or enforcement precedents. The method requires answers to rely on retrieved passages rather than model memory.
3. Check status and authority. The agent distinguishes rules in force from drafts, future-effective texts and superseded versions. It applies a hierarchy of sources and treats precedents as context. Historical texts remain available for questions about earlier dates. This helps an expert see which version supports the answer.
4. Explain the implications. An answer connects the rule to the product’s facts, identifies options or evidence gaps, and gives citations and a confidence assessment. Interpretation is labelled. When the library lacks a necessary product guideline, the method requires the agent to disclose the gap.
5. Leave accountable decisions with the expert. The responsible regulatory lead reviews the analysis, resolves ambiguous classifications and approves the company’s position. Signing, filing, submission and the relationship with regulators remain human responsibilities. Drafts are prepared for expert review.
The architecture starts with a shared legal foundation, then adds specialty packs for deeper product knowledge. The supplied release documentation describes dental, consumer healthcare, oncology and rare-disease packs. Consumer healthcare includes operating methods for cosmetics, toothpaste, health food, OTC-specific questions and consumer-health devices. The Value Map also describes a client-context layer for portfolio information, dossier history and internal templates; implementation scope should be agreed for the client.
For example, a team preparing a cosmetic launch could ask which evidence needs review before using a proposed efficacy claim. The agent’s role is to retrieve the relevant provisions, check their status and identify the facts and evidence needed for an assessment. The responsible expert approves the final claim. This is an illustrative use case, not a reported client outcome.
What the benchmark results show
JUMO’s supplied release infographic reports two internal studies using the same model and the same questions across three configurations: a generic AI with web search, that AI with JUMO’s official texts, and the full JUMO RA Agent with official texts and regulatory skills. Answers were described as graded blind, claim by claim, against official texts.
October 2026: 60 real client questions
| Measure | Generic AI + web search | JUMO official texts only | Full JUMO RA Agent |
|---|---|---|---|
| Safe answers: no decision-changing error | 90% | 95% | 97% |
| Claim accuracy | 86% | 96% | 97% |
| Major errors across 60 answers | 21 | 8 | 4 |
The full agent recorded four major errors compared with 21 for the generic assistant. Its reported safe-answer rate rose by seven percentage points. These measures describe performance on the tested questions, not a guarantee for future answers.
November 2026: 100 questions
| Measure | Generic AI + web search | JUMO official texts only | Full JUMO RA Agent |
|---|---|---|---|
| Claim accuracy, matched settings | 93.4% | 98.3% | 98.0% |
| Major errors across 100 answers | 21 | 3 | 4 |
Safe-answer rates were reported as 99–100% across all three configurations. The distinguishing gain was precision in details such as articles, dates and catalogue codes. Both JUMO configurations recorded substantially fewer major errors than the generic assistant.
The library-only configuration slightly exceeded the full agent on claim accuracy in November and recorded one fewer major error. The comparison highlights the importance of authoritative evidence. It does not establish that adding the full workflow improves every quality measure on every question set.
These are JUMO-reported internal results. The infographic describes the November answers as expert-validated, while also marking the research section “pending expert review of the grading.” Treat the results as provisional internal evidence. The supplied documentation does not include the complete question sets, scoring protocol or statistical uncertainty.
Put the agent to work on one defined task
The release 2026-11.6 infographic, dated 12 November 2026, reports 955 source documents: 907 regulatory texts and 48 court or enforcement cases. It lists 841 texts in force and more than 230 regulatory questions used as a release gate. These are documented figures for that release, rather than a claim of exhaustive regulatory coverage.
The important operational question is whether the evidence and method help your experts complete a selected task. A sensible evaluation would proceed as follows:
- Choose a bounded use case. Start with a cited pathway explanation, a review of a proposed change or a question about claim evidence. Confirm that the relevant product-family sources and workflow support are available before relying on the output.
- Assign review ownership and data boundaries. The regulatory lead should define the acceptance criteria. IT and legal stakeholders should confirm permitted data, access and provider arrangements before confidential material is used.
- Compare complete work. Assess a sample of your own questions using the existing process and the agent-assisted process. Measure evidence-gathering time, expert review time, unsupported claims and necessary corrections. Include the effort needed to check the answer.
- Set an expansion decision. Broaden use only if the results meet the agreed quality standard and help the team. Keep ambiguous classifications, missing evidence and low-confidence conclusions with the responsible expert.
This is recommended implementation guidance, not a claim that JUMO has achieved a particular client saving. It gives management and regulatory leaders a way to assess value without turning an answer-quality benchmark into a forecast.
To discuss a suitable regulatory task, source coverage and expert-review responsibilities for your portfolio, contact us .
Frequently Asked Questions by Executives
Q1: What business problem does the JUMO Regulatory Agent address? A: The JUMO Regulatory Agent supports the research and analysis needed to assess regulated products for market entry and lifecycle changes. It retrieves evidence, checks regulatory status and structures cited answers for expert review. Its business value should be assessed through task quality, total effort and avoidable rework.
Q2: Which product areas does the supplied documentation describe? A: The documentation describes drugs, biologics, medical devices, in vitro diagnostics and consumer-health products. Specialty packs add depth in dental, consumer healthcare, oncology and rare diseases. Teams should confirm the relevant product-family sources and workflow support for their intended task before relying on an answer.
Q3: Why does regulatory work create costs beyond filing an application? A: Regulatory projects combine specialist analysis with document preparation, testing and, where applicable, clinical work. Ongoing changes and renewals add further effort. The published CIRS cost framework separates these components, while its 2026 renewal article describes how errors can cause repeat registration work and disrupt market access.
Q4: How does the agent differ from a generic AI assistant? A: The JUMO Regulatory Agent combines access to a curated regulatory library with operating instructions for retrieval, status checks, citations and analysis. A generic assistant with web search was one comparator in JUMO’s internal studies. The reported improvement is specific to those tests and should be checked against the team’s own questions.
Q5: What does the agent check before using a regulation? A: The inspected operating method requires retrieval of the supporting passage and a check of the text’s status before it is relied on for a decision. It distinguishes current rules, future-effective provisions, drafts and superseded versions. The question’s applicable date determines which version matters.
Q6: Can the agent approve a claim or submit a registration? A: Approval and submission remain human responsibilities. The agent prepares evidence and analysis for review; the responsible expert decides the company’s position, approves claims and handles signing or filing. This separation should be maintained in the team’s workflow and acceptance criteria.
Q7: What did the October benchmark report? A: Across 60 real client questions, JUMO reported 97% claim accuracy for the full agent compared with 86% for the generic assistant with web search. Major errors numbered four versus 21. The reported safe-answer rates were 97% and 90%, respectively. These are provisional internal results.
Q8: What did the November benchmark report? A: Across 100 questions, JUMO reported 98.0% claim accuracy and four major errors for the full agent, compared with 93.4% and 21 errors for the generic assistant. The library-only setup achieved 98.3% accuracy and three major errors. Safe-answer rates were 99–100% across all three setups.
Q9: How much confidence should management place in those results? A: The benchmarks are useful internal evidence for a controlled evaluation. The supplied infographic marks grading as pending expert review, despite describing the November answers as expert-validated. Without the full scoring protocol, question sets and statistical uncertainty, the figures should not be treated as an independent performance guarantee.
Q10: Do the benchmarks prove time savings or faster launches? A: The supplied benchmarks measure answer quality, not working hours or launch timelines. Teams should measure their own total task effort, including expert review and corrections. Any claim about time saved should come from that comparison rather than be inferred from accuracy scores.
Q11: Which task should a team evaluate first? A: Choose a bounded task with a clear question and expert-defined acceptance criteria, such as a cited pathway explanation or proposed-change review. Confirm relevant source and workflow coverage first. Avoid expanding the evaluation into an entire registration project before the selected task’s quality and effort are understood.
Q12: Who should own adoption of the agent? A: The regulatory lead should own answer-quality standards and review responsibilities. The business sponsor should assess the value of the selected use case, with IT and legal stakeholders confirming access and permitted data use. This is recommended governance, not a prescribed staffing model from a benchmark.
Q13: What should be checked before using confidential documents? A: Confirm the assistant, connector, provider arrangements, access controls and permitted data use with the appropriate internal stakeholders. The supplied method describes controlled evidence retrieval but does not establish a universal data-residency or privacy guarantee. Client-context configuration should be agreed before confidential material is introduced.
Q14: What happens when a source is missing or the answer is uncertain? A: The inspected method requires the agent to disclose missing product guidelines, state assumptions and label interpretation or low confidence. The responsible regulatory expert must resolve a material gap before the business relies on the conclusion. Unavailable evidence should never be replaced with an invented reference.
Q15: How can we discuss whether the agent fits our portfolio? A: Contact us with a defined regulatory task, relevant product category and the decision your team needs to support. Discuss source coverage, workflow scope and expert-review responsibilities, then agree how a controlled evaluation would measure answer quality and total task effort.
We are Here to Empower
At System in Motion, we are on a mission to empower as many knowledge workers as possible. To start or continue your GenAI journey.
You should also read
Complex AI Agents
Article 1 minute readLet's start and accelerate your digitalization
One step at a time, we can start your AI journey today, by building the foundation of your future performance.
Book a Training