For the past several weeks I've been running a quiet experiment. Every Sunday at 6 PM Eastern, an AI agent assembles a global medical-device regulatory intelligence digest covering roughly 110 primary sources across the EU, US, UK, Canada, Mexico, China, Japan, Korea, Australia, ASEAN, the GCC, and IMDRF / WHO / MDSAP. Weekly mode covers the prior seven days. The last Sunday of each month, it produces a deeper monthly view.

I'm sharing the May 2026 monthly digest with this post because I want feedback - and because I think the more interesting story is how it gets validated, not how it gets generated.

What's in the report

Each digest scope-filters every item against a single rule: it must be a medical device, IVD, SaMD (including AI/ML and digital therapeutics), or the device-side of a drug-device combination, AND it must affect design, registration, market access, post-market surveillance, vigilance, or compliance obligations. Pure pharma, biologics, cosmetics, veterinary, and food/supplement regulations are excluded - they have plenty of other channels.

Items are tiered:

• CRITICAL - framework-level, with a concrete near-term trigger (deadline, effective date, binding regulator action)
• STRATEGIC - durable impact, guidance, cluster signals, NB capacity, jurisdictional digital-health pathways
• MONITOR - individual recalls, FSNs, Early Alerts, warning letters (collected in a Watchlist)

The monthly view layers in cross-jurisdictional macro-themes (currently tracking AI/ML and SaMD lifecycle oversight, submission digitization mandates, cluster enforcement as a durable FDA tool, UDI traceability convergence, and software-update-as-Class-I remediation) plus a forward calendar of upcoming deadlines.

How it's generated

A regulatory-intelligence agent runs a structured scan across primary regulator sites (fda.gov, ec.europa.eu, gov.uk, canada.ca, nmpa.gov.cn, pmda.go.jp, mfds.go.kr, tga.gov.au, hsa.gov.sg, and others), pulls candidate items, scope-filters them, tiers them, and writes a structured payload. A renderer turns that payload into a branded PDF and a matching HTML email body.

That part is automated. It's also where most AI-generated reports stop.

How it's validated

Here's the part I'm proud of, and the part that's still imperfect.

Every draft goes through a dual-AI verification loop with a human gate before it can be distributed:

1. Draft. The generator produces a v1 PDF and uploads it to a private Drive folder.
2. Independent review. A second AI - a different model, a different prompt - reads the PDF and checks four things: (a) scope filter integrity, (b) tier logic, (c) primary-source verification for every CRITICAL/STRATEGIC item, and (d) every date, deadline, and number cross-referenced to a primary regulator URL. Numbers and dates are the highest-risk fields.
3. Adjudication. The first agent reviews the findings, accepts what it agrees with, and rejects with documented rationale what it disagrees with. Amendments are applied; a v2 PDF is produced.
4. Second review. The reviewer takes another pass on v2: did the accepted amendments land correctly? Is the rejected rationale defensible? Anything new?
5. Human approval. Nothing leaves the loop without my explicit "approve and distribute" in chat. There is no auto-distribution, ever. There are only two review cycles - if the second pass still flags structural issues, the report escalates to me directly.

This means the published digest carries two AI signatures and one human signature. The AIs disagree often. That's the point. Where they disagree, the disagreement is documented and the rationale lives in the file's history.

It's not perfect. The May monthly review caught date errors (a Mexican standard misdated by a week, an EU AI Act applicability date that conflated two annexes), scope drift in the cardiac-cluster framing, and a template bug that leaked "This Week's Synthesis" into a monthly file. Cycle 2 will catch the rest, or escalate them.

What I'm asking for

Please tell me what's wrong with it. What did the AIs miss? Where is the scope filter too tight, or too loose? Which tiering calls would you have made differently? Which jurisdictions deserve more depth, or less? Comment on this post, or DM me - both work.

If you'd like it in your inbox every Monday at 7 AM, there's a free subscription at https://info.arazygroup.com/regtalk-subscription. You can pick weekly digests, monthly deep-dives, or both. No charge, no sales call, no upsell - I'm interested in the feedback loop more than the funnel.

The AIs are not going to write themselves out of the imperfection. The humans reading along are.

Want the May 2026 monthly digest as a starting point? Attached below, or available on request.