The landscape
Find the evidence in systems that do not talk to each other. Make a defensible call against a clock. Write the rationale in Word. Your QMS holds the record - it does not help you make the call.
CAPA & investigations
A CAPA opens. A first plausible root cause shows up within an hour. The real one - a supplier process change, a drifted calibration, a training gap that compounded over six lots - takes days, and lives in records nobody connected. The CAPA closes. Three quarters later it reopens, this time as a failed effectiveness verification.
Effectiveness is what FDA pulls during your next inspection. A closed-then-reopened CAPA reads as a systemic gap.
64% of medical-device warning letters cite CAPA deficiencies. 26 FY2025 warning letters root in CAPA. 28% of 483s flag inadequate investigation depth.
The agent drafts each CAPA stage, asks for the evidence behind every claim, and pushes back on weak root causes. Before you close, an audit-style check pressure-tests the CAPA against the FDA citations library so a thin investigation does not become a 483.
Complaints & MDR
A complaint comes in that might be reportable. Causation is unclear, severity is borderline, recurrence is unknown. Your QA Director opens seven tabs - complaints DB, MAUDE, the risk file, prior MDRs, the engineering investigation, the regulation itself, an empty Word doc - and assembles a defensible decision by hand.
30 calendar days from any-employee awareness. 5 days for events requiring remedial action. A late or wrong call is the citation FDA pulls during inspection.
Complaint-file violations were the 3rd-most-cited Part 820 issue in FY2025 (23 warning letters). FDA receives 2M+ MDRs per year.
The agent pulls from complaints, MAUDE, your risk file, and prior MDRs in one pass. It drafts the reportability rationale as a side-by-side table with a citation on every claim. An audit-style check runs on the decision before you sign.
NCR & supplier quality
A supplier ships a batch with a drift in the pressure-sensor lot. NCRs open. Engineering finds the root cause weeks later - a supplier process transition to automated dispensing on a known date. Last quarter’s CAPA on the same lot already failed effectiveness. Nobody connected the two.
Purchasing-controls citations expand into design control, CAPA, and complaint files when audits trace one root cause across silos.
Purchasing controls = 15 FY2025 warning letters. The supplier-driven NCR is the connecting tissue between complaints, CAPA, and design control.
The agent pulls supplier qualifications into a side-by-side review with sources cited. Right-click a flagged row to open a CAPA - related documents and the supplier’s gap clauses come with it.
Change control
Engineering proposes a component change. Impact crosses the design history file, the risk file, the V&V plan, the regulatory submissions library, training records, and supplier-quality docs. The assessment lives in a Word matrix that gets emailed for sign-off. Half the dependencies stay invisible until an auditor finds them.
Design control is the 2nd-most-cited Part 820 violation. A missed dependency in the impact assessment is the warning-letter pattern.
Design control = 25 FY2025 warning letters. Top-three Part 820 citation in every recent year.
The agent traces the change across your design history, risk file, V&V records, training, and supplier quality. Every downstream document the change touches surfaces with its source so nothing slips through.
Regulatory submissions
Twelve sections, dozens of predicates to compare, hundreds of citations to assemble. The traceability matrix gets rebuilt in Excel. Six weeks of consultant time. The next submission rebuilds it from scratch.
A refuse-to-accept hold restarts the clock. Each round of additional-information requests adds weeks before clearance, and the rebuild work falls on the same team.
Submissions where the same predicate analysis, traceability matrix, and citation set get rebuilt from scratch every time. The work compounds across the company; the artifacts do not.
The agent drafts 510(k) and De Novo scaffolds from one connected memory. The predicate analysis becomes a traceability table you can re-use, so the next submission starts from the last one instead of from zero.
Post-market surveillance + AI drift
Drift detection happens by manual quarterly review, if at all. New patient population, model retrains, no PCCP - every change becomes a new submission. Field signals do not trigger investigations until they cluster into a recall.
EU AI Act high-risk obligations enforce August 2027: incident reporting, drift monitoring, post-market surveillance documentation. There is no post-market workspace built for AI today.
1,350+ FDA-authorized AI/ML medical devices. Only ~16.7% of 2024 approvals shipped with an authorized PCCP.
The agent watches the KPIs you set in the PMS plan. When per-1000-unit signals trip a threshold, it opens a CAPA. Drift telemetry from Arize, Fiddler, and WhyLabs joins the evidence record automatically.
The shared answer
Every workflow needs the same thing: pull evidence from many sources, cite each one, make a call. Today that work is manual - copy, paste, stitch. The agent does the pulling. The Evidence Matrix is the shape that fits all six: rows for the factors, columns for the sources, a citation in every cell.
Open a sample matrix →The supplier change that broke this CAPA also broke last quarter’s NCR. The predicate analysis from the last 510(k) gets rebuilt by hand. Every investigation starts from zero. The agent works against the Brain - one connected memory across complaints, CAPAs, NCRs, changes, submissions, and post-market signals - so prior work carries forward instead of restarting.
See the Brain →64%
of FDA medical-device warning letters cite CAPA deficiencies (FY2022 - FY2024).
FDA inspection observations · Compliance Insight FY2024
$250K - $5M+
direct remediation per warning letter, scaled to company footprint and finding scope.
The FDA Group + Compliance Architects analyses
+73% YoY
FDA warning-letter volume, July to December 2025 vs. the same period in 2024.
ECA Academy FY2025 statistics
A CAPA fails because the supplier change lived in a different system. A 510(k) drags because predicate analysis gets built from scratch. Same failure every time - knowledge that should carry forward, doesn’t.
Pilots include a baseline measurement. No data leaves your tenant.