Your Complaint Backlog Is a Warning Letter Waiting to Happen

Walk into any mid-market device company's QA department on a Friday afternoon. Ask to see the complaint log. Then ask how many complaints have been open longer than 30 days.

The number will be larger than the QA director wants to admit.

It's not because they're bad at their jobs. It's because complaints come in faster than the team can investigate them, and the system designed to manage them was built for a much smaller company. The result is a backlog that grows slowly, invisibly, and stays manageable until a regulator shows up — and then it's catastrophic.

FDA issued 327 warning letters between July 1 and December 3, 2025 — a 73% increase over the same period in 2024.

Complaint files (21 CFR 820.198) ranked as the third most-cited deficiency in medical device warning letters during fiscal year 2025. Of the FDA's published 2025 CDRH warning letters, the majority that targeted device manufacturers cited complaint handling failures specifically: missing procedures, informal complaint handling, late or incomplete MDR submissions, and lack of systemic controls.

This is happening at the same time the FDA's new "Elsa" AI targeting system is making it easier for them to find companies with complaint anomalies. The math is shifting against any company that hasn't modernized its complaint handling operation.

Why mid-market is uniquely exposed

Companies at the extremes have different problems. Small startups have low complaint volumes; one person can hand-manage everything. Large enterprises have well-resourced complaints teams with 15+ people, dedicated medical reviewers, eMDR specialists, and mature QMS platforms.

Mid-market sits in the worst spot. Complaint volume is high enough that hand-management doesn't work, but the team is rarely big enough to keep up. A typical $100M-300M device company gets 200-1,500 complaints per year depending on indication. The complaints team is usually 1-3 people. The math is grim.

The structural problem is that complaints arrive as unstructured text. Phone notes from reps. Emails from distributors. PDFs from surgeons. Text messages forwarded from sales managers. Faxes (yes, still). Each one needs to be categorized, coded to IMDRF taxonomy, evaluated for MDR reportability, investigated, and either closed or escalated. The mechanical work of just processing an incoming complaint takes 30-60 minutes before any actual investigation begins. Multiply by 500 complaints a year and you've consumed 250-500 hours of skilled QA time just on intake.

This is why backlogs grow. The team isn't slow. The work is enormous and the system makes them do most of it manually.

What the FDA actually cares about

It's worth being precise about what triggers warning letters specifically for complaint handling. The 2025 pattern across CDRH warning letters shows five recurring themes:

Missing or inadequate complaint handling procedures. Companies that don't have a documented, written procedure for receiving and processing complaints. Or procedures that exist but are missing essential definitions like "malfunction" or "serious injury."

Informal handling. Complaints handled by individual managers without flowing through the formal QMS process. The FDA inspector sees emails from 2024 mentioning device issues that were never opened as complaints in the system. That's a finding.

Late MDR submissions. When a complaint meets MDR reportability criteria but the report doesn't get filed within the regulatory window. This is increasingly common because mid-market QA teams are stretched.

Inadequate investigation. Complaints opened, briefly noted, and closed without meaningful root cause analysis. The FDA expects investigation depth proportional to the complaint severity. A vague "device failure noted, customer satisfied with replacement" doesn't cut it for anything serious.

No trending or CAPA linkage. Complaints handled as one-off events with no analysis of patterns across complaints. When 30 customers complain about the same packaging issue, the FDA expects you to have noticed and opened a CAPA. Companies that don't are vulnerable.

If your QA director can't confidently tell you the company's status against all five of these, you're exposed.

Where AI genuinely shines (not vendor hype)

The complaint handling workflow is unusually well-suited to AI augmentation. The reason is that most of the work is structured-output translation from unstructured input — exactly what language models are good at.

Triage and severity assessment

An agent reads incoming complaints and classifies them: device problem code, severity level, likely reportability category, recommended priority. The QA reviewer sees pre-classified complaints and either confirms or adjusts. The mechanical work of categorization gets compressed from hours to minutes.

IMDRF coding

The IMDRF Adverse Event Reporting taxonomy is a structured vocabulary the FDA increasingly expects in MDR submissions. Coding complaints to IMDRF terms is painstaking manual work. AI tools can do this consistently and at speed, flagging uncertain cases for human review.

Initial investigation drafting

When a complaint requires investigation, an agent can produce a first draft of the investigation framework: similar prior complaints, potentially relevant CAPA records, recommended investigation steps, the questions to ask the customer. The investigator reviews and adapts. The blank-page problem goes away.

MDR drafting

When a complaint meets reportability criteria, the eMDR submission requires structured data in specific fields. AI tools can populate the structured fields based on the complaint investigation, generate the narrative sections, and produce a draft eMDR for human review. Submission turnaround time drops from days to hours.

Trend detection

This is the highest-value capability. Humans are bad at noticing patterns across hundreds of complaint records. AI tools can flag emerging patterns — five complaints about the same packaging issue across three weeks, an unusual cluster of complaints from a specific surgeon — before they become CAPAs. This isn't replacing the QA director's judgment; it's giving them earlier signals to act on.

What AI doesn't do

Same list as the QMSR article, with one addition specific to complaints.

AI doesn't make the final reportability decision. Whether a complaint meets MDR criteria is a regulatory judgment that has legal consequences. A model's output is input to the decision, not the decision itself.

AI doesn't replace the customer call. Sometimes the only way to understand what happened is to call the customer and ask. A model can suggest the questions. A human has to make the call.

AI doesn't replace the medical reviewer. If your complaint handling process requires medical review (and for many device categories it does), that review remains a human function.

And the addition: AI doesn't replace post-investigation root cause analysis. Identifying why a device failed in a specific case requires engineering judgment, design knowledge, and sometimes physical testing. A model can help organize the analysis. It can't perform it.

The business case

Mid-market device companies that deploy AI-augmented complaint triage typically see three outcomes within 90-120 days.

Backlog clearance. Complaints older than 30 days drop from 15-25% of the open queue to under 5%.

MDR submission timeliness. Reportable events get filed within FDA deadlines consistently, eliminating the late-MDR warning letter risk.

Trend identification. CAPAs get opened earlier in the lifecycle of a recurring issue, which means smaller scope, lower remediation cost, and lower regulatory risk.

The cost of these tools is in the $30-80K per year range for mid-market deployments. The downside scenario for not doing this is a single warning letter, which typically costs $500K-2M in remediation work, consultant fees, and management time over 12-18 months, plus the reputational and procurement impact when major hospital customers see your name on the FDA's public warning letter database.

The math is straightforward. The reason most companies aren't doing it isn't economic. It's that the QA director is too busy fighting fires to evaluate the tools that would help them stop fighting fires.

A specific recommendation for this quarter

If you're a CEO or COO and you haven't asked these specific questions, ask them this week:

How many complaints are currently open longer than 30 days? Get a number, not a vague answer.

What percentage of incoming complaints are categorized within 48 hours of receipt? Same — get a number.

When did we last open a CAPA based on complaint trending data, not on a single incident? If the answer is "I'd have to check," the trend detection system isn't working.

If the FDA showed up tomorrow and asked for our complaint trend analysis from the past 12 months, how many days would it take us to produce it? Anything over a week is a problem.

Have we evaluated AI-augmented complaint triage tools? Same question I'd ask about QMSR. Not buying — evaluating. The companies that integrate these tools in 2026 will compress their complaint cycle times by 50-70% and dramatically reduce their warning letter risk.

Complaint handling is the most-overlooked operational risk at mid-market device companies. It's expensive to ignore, cheap to fix relative to the downside, and getting more dangerous every quarter as FDA enforcement accelerates.

The QA team in your company isn't asking for help. They've adapted to the workload. They're going to keep doing the best they can with the tools they have. The question is whether you, as the executive responsible for the company's regulatory standing, are going to give them better tools before the FDA decides for you.