Interpreting Complaint Data

Complaints data presents a rich and dense source of information for root cause analysis in any business because complainants often have multiple touch points with businesses before they reach the complaints team. But making sense of this data is tricky. Most companies we work with adopt a data categorisation strategy, where handlers are asked to label complaints according to their root causes. Statistical trends and data dashboards are then built on top of this structured information. While this approach is a good start, it suffers from a number of inherent limitations: • Complaints handlers are not experts in root cause analysis and often lack the time to carry out in-depth root cause reviews. This leads to frequent misclassifications. • Data collected in this way is totally dependent on the types of root causes included in the business' framework, making it difficult to capture new and emerging trends. • Achieving a balance between a granular and an effective categorisation framework presents an ongoing challenge: while more root cause categories makes the analysis more granular, it also increases the risk of mis-classification. At CourtCorrect, we were one of the first companies to use AI for root cause analysis in complaints. Concretely, our AI technology helps handlers select the correct root cause category for any given complaint, increasing accuracy without impacting productivity. Additionally, we've been exploring the use of AI to *identify root cause categories independently from existing frameworks*, allowing for a more flexible analysis of root causes outside of the confines of existing root cause categories. We're very pleased to share a major breakthrough from our Data Science Team on the automatic identification of root cause clusters with AI. Below, you'll see our latest root cause AI model independently identify and classify common complaint types on a sample dataset. In this example, our clustering technology has identified 5 main root causes in the complaints dataset based on the actual contents of each complaint file. The granularity, i.e. the number of clusters identified by the model, can be set by the user. So you can zoom in and zoom out of the plot to reflect differing levels of granularity. Additionally, we enable analysts to ask questions in their own words, e.g.: • What caused an increase in the number of complaints in this cluster in Q1? • Which clusters are responsible for the majority of compensation payments to claimants? • Which actions can be taken to prevent these complaints from happening again in the future? The combination of accurate and flexible identification of problem clusters with a natural language interface to query information effectively promises to significantly improve the quality and speed at which businesses can derive insights from their complaints data. We're excited to continue partnering with our clients to deliver this innovation to the market.

Stop collecting customer feedback. Start decoding it. The difference? Most founders I speak to are not short of data. They are drowning in it. NPS dashboards. Support tickets. Feature requests. Slack screenshots. “Love the product but…” It feels productive. It looks customer-centric. But nothing changes. Because collecting feedback is passive. Decoding it is relational. When a customer says: “Can you add dark mode?” They might mean: “My eyes hurt.” “I use this at night.” “I don’t feel considered.” When someone says: “Your onboarding is confusing.” They might mean: “I feel stupid.” “I don’t feel confident.” “I don’t know if I made the right decision choosing you.” The words are surface. The experience is underneath. Most teams build from the sentence. Few build from the emotion. And that is why churn feels mysterious. Here’s how you start decoding instead of collecting:  1. Stop asking what they requested.   Start asking what they felt.  2. Look for repetition of emotion, not repetition of wording.  3. Translate every complaint into: “What expectation did we unintentionally break?” Because feedback is not a to-do list. It is a mirror. Collecting tells you what happened. Decoding tells you what it meant. And meaning is where retention lives. If you’re a founder reading this: What’s one piece of feedback you keep seeing… but haven’t truly interpreted yet? Drop it below. Let’s decode the first layer together. — Villia Lorna Founder, Villia Connect CX & Communication Strategist Bridging the Gap between what you mean and what people feel

Most brands handle customer complaints. But very few learn from them. DTC brands live in a data-rich world and complaint data is no exception. You know: -Who complained -Where they complained -How often they complain -The nature of the complaint -Open-to-close speed across all agents -Which products are complained about most What you’re probably not doing: -Mapping complaint trends back to root cause in manufacturing, packaging, or formulation -Quantifying the cost of complaints in terms of refunds, churn, or lost LTV -Using complaint data as a leading indicator for quality system breakdowns -Systematically closing the loop with ops, R&D, and suppliers -Turning high-friction interactions into retention opportunities through process + empathy Complaints are a goldmine of operational insight but only if you treat them like the early warning system they are. The best brands I’ve worked with don’t just respond to appease the customer - they investigate, escalate, and correct. That’s how you build trust at scale. Don't sleep on your complaint data.

Imagine you're a data consultant hired by a hospital facing constant complaints about long patient wait times in their outpatient department. The management wants to understand the root causes and find a solution, so they provide you with data, including patient arrival times, consultation durations, staff schedules, and resource availability. 🚩 Step 1: Analyzing the Data: You dive into the data and uncover various patterns: - Peak patient influx occurs between 9 AM and 11 AM. - Consultation durations vary significantly among doctors, with some taking twice as long as others. - There’s a mismatch between staff schedules and patient demand, leading to bottlenecks during peak hours. 🚩 Step 2: Finding the Central Insight: Amid the analysis, you identify a key insight: The primary driver of long wait times is the misalignment between staff availability and patient demand during peak hours. 🚩 Step 3: Building the Data Story: Your data story revolves around this central insight. You structure your presentation to guide the audience step-by-step toward this conclusion: - Introduction/Hook: Begin with patient testimonials or survey results highlighting frustration with long wait times, this will push stakeholders off their seats to pay attention - Setting the Stage: Share descriptive statistics, such as average wait times and peak hours of patient arrivals. - The Climax (Central Insight): Visualize the staff scheduling versus patient demand with a heatmap or line chart. Clearly show the misalignment during peak hours, leading to bottlenecks. - Resolution: Offer data-driven recommendations, such as adjusting staff shifts to better align with patient demand or introducing a triage system to manage peak hour surges. 📌Why the Central Insight Matters: Without the central insight, your data story would feel like a collection of random facts: a scatterplot here, a bar chart there, but no cohesive narrative to tie it all together. By focusing on the misalignment as the main point, you give the story purpose, direction, and a clear call to action. ✅️ The takeaway here is simple: before you build your data story, find the one meaningful insight that ties everything together. That central insight isn’t just a detail—it’s the reason your story exists. Without it, you risk confusing or losing your audience. 📌 In the real world, whether it's reducing patient wait times, optimizing supply chains, or boosting sales performance, the power of your data story lies in the clarity and relevance of your central insight. So this weekend, I'm excited to invite you to the Zion Tech Hub weekend webinar, where together with Louisa Igbonoba I’ll be discussing: The Art of Effective Data storytelling. Date: Saturday, 18th January Time: 5pm WAT Sign up via the link below https://lnkd.in/di_YY4Mw Keep Learning and keep growing

Section 5.2.3- Customer Complaints: PBRERs ➖➖➖➖➖➖➖➖➖➖➖➖➖➖➖➖➖➖ When authoring Section 5.2.3 of a PBRER, customer complaints often feel like the “miscellaneous” drawer of pharmacovigilance data. A broken blister, a leaking vial, a discolored tablet- they may not look like safety cases at first glance. Yet, when examined carefully, these reports can reveal patterns of risk, manufacturing issues, and even potential safety signals. In fact, many customer complaint cases overlap with medication errors, misuse, or product quality issues, making them integral to understanding the full picture of a product’s real-world performance. Why do these matter: ▪️A cracked ampoule may signal packaging design flaws that increase injury risk. ▪️Repeated complaints from a specific batch could indicate a manufacturing defect with downstream clinical impact. ▪️A labeling smudge or font issue might increase the chance of dosing errors. Concentration of analysis: 👉 Are complaints clustering around certain batches, geographies, or manufacturing sites? 👉 Do complaints coincide with medication errors, AEs, or product recalls? 👉 Are some complaints precursors to serious incidents- a “near miss” before harm occurs? Even when case counts are low, dismissing them as “minor” risks missing early warning signs. Enough cumulative data can turn a pattern into a preventive action- sometimes even before regulators demand it. In pharmacovigilance, customer complaints are like whispers. Listen closely, and they might just tell you where the next risk will come from. Share your experience how you update this section while drafting reports. #CustomerComplaints #AggregateSafetyReports #Pharmacovigilance #DrugSafety #PatientSafety #MedicalWriting #MedicalWriter #RegulatoryWriting #LinkedIn #LinkedInCommunity #KnowledgeSharing #Monday #MondayDose

𝗛𝗼𝘄 𝗰𝗼𝗺𝗽𝗹𝗮𝗶𝗻𝘁 𝘁𝗿𝗲𝗻𝗱𝘀 𝗰𝗮𝗻 𝗳𝗼𝗿𝗲𝘀𝗵𝗮𝗱𝗼𝘄 𝗿𝗲𝗴𝘂𝗹𝗮𝘁𝗼𝗿𝘆 𝗮𝗰𝘁𝗶𝗼𝗻  On March 30, the FDA posted that Intuitive Surgical has initiated 6 𝗖𝗹𝗮𝘀𝘀 𝗜𝗜 𝗿𝗲𝗰𝗮𝗹𝗹𝘀 for the EndoWrist instruments (link in comments).   According to the recall notice:  “𝘖𝘯 𝟏𝟐/𝟐𝟗/𝟐𝟎𝟐𝟓, 𝘵𝘩𝘦 𝘧𝘪𝘳𝘮 𝘴𝘦𝘯𝘵 𝘷𝘪𝘢 𝘦𝘮𝘢𝘪𝘭 𝘢𝘯 ‘𝘜𝘳𝘨𝘦𝘯𝘵: 𝘔𝘦𝘥𝘪𝘤𝘢𝘭 𝘋𝘦𝘷𝘪𝘤𝘦 𝘙𝘦𝘤𝘢𝘭𝘭’ 𝘭𝘦𝘵𝘵𝘦𝘳 𝘪𝘯𝘧𝘰𝘳𝘮𝘪𝘯𝘨 𝘤𝘶𝘴𝘵𝘰𝘮𝘦𝘳𝘴 𝘵𝘩𝘢𝘵 𝘐𝘯𝘵𝘶𝘪𝘵𝘪𝘷𝘦 𝘩𝘢𝘥 𝘣𝘦𝘤𝘰𝘮𝘦 𝘢𝘸𝘢𝘳𝘦 𝘰𝘧 𝘢𝘯 𝘪𝘯𝘤𝘳𝘦𝘢𝘴𝘦 𝘪𝘯 𝘤𝘰𝘮𝘱𝘭𝘢𝘪𝘯𝘵𝘴 𝘳𝘦𝘨𝘢𝘳𝘥𝘪𝘯𝘨 𝘧𝘳𝘢𝘺𝘦𝘥 𝘰𝘳 𝘣𝘳𝘰𝘬𝘦𝘯 𝘤𝘢𝘣𝘭𝘦𝘴 𝘰𝘯 𝘴𝘰𝘮𝘦 𝘥𝘢 𝘝𝘪𝘯𝘤𝘪 𝘚 𝘢𝘯𝘥 𝘚𝘪 𝘳𝘦𝘶𝘴𝘢𝘣𝘭𝘦 𝘪𝘯𝘴𝘵𝘳𝘶𝘮𝘦𝘯𝘵𝘴.”   This recall becomes especially interesting when looking at the the 𝗠𝗔𝗨𝗗𝗘 𝗰𝗼𝗺𝗽𝗹𝗮𝗶𝗻𝘁 𝗱𝗮𝘁𝗮 over the past years..   By plotting MAUDE complaints over the past three years for the EndoWrist device that mention “𝘤𝘢𝘣𝘭𝘦” in the problem description (see figure below), a clear pattern emerges. Out of more than 𝟕𝟑,𝟎𝟎𝟎 𝘁𝗼𝘁𝗮𝗹 𝗰𝗼𝗺𝗽𝗹𝗮𝗶𝗻𝘁𝘀 for this product during that period, 𝗼𝘃𝗲𝗿 𝗵𝗮𝗹𝗳 (𝟒𝟕,𝟓𝟗𝟓) reference cable‑related issues.   The trend shows that cable‑related complaints began increasing toward the end of 𝟐𝟎𝟐𝟒 and peaked in the 𝗳𝗶𝗿𝘀𝘁 𝗵𝗮𝗹𝗳 𝗼𝗳 𝟐𝟎𝟐𝟓.   Looking more closely at the 𝘗𝘳𝘰𝘥𝘶𝘤𝘵 𝘗𝘳𝘰𝘣𝘭𝘦𝘮 fields, many reports describe issues such as material being 𝘀𝗽𝗹𝗶𝘁, 𝗰𝘂𝘁, 𝘁𝗼𝗿𝗻, 𝗳𝗿𝗮𝗴𝗺𝗲𝗻𝘁𝗲𝗱, 𝗼𝗿 𝗳𝗿𝗮𝘆𝗲𝗱, which aligns closely with the concern described in the recall communication.   The 𝘗𝘢𝘵𝘪𝘦𝘯𝘵 𝘗𝘳𝘰𝘣𝘭𝘦𝘮 data is also informative. Most reports indicate either no clinical symptoms or insufficient information. However:  • 𝟐𝟗𝟗 𝗿𝗲𝗽𝗼𝗿𝘁𝘀 indicate a foreign body remaining in the patient • 𝟏𝟓𝟒 𝗿𝗲𝗽𝗼𝗿𝘁𝘀 reference unspecified tissue injuries • Other reports mention burns, blood loss, or perforations of vessels or bowel While MAUDE data is publicly available, extracting insights at this scale is not trivial. The FDA’s public interface limits queries to 𝟓𝟎𝟎 𝗿𝗲𝗰𝗼𝗿𝗱𝘀 𝗮𝘁 𝗮 𝘁𝗶𝗺𝗲, which complicates trend analysis when dealing with tens of thousands of reports.   By mirroring the complete MAUDE database and enabling more flexible querying, it becomes possible to uncover patterns like these earlier and to use them for post‑market surveillance, signal detection, and peer benchmarking (see my recent post in the comments)   This case is a good example of how 𝗽𝘂𝗯𝗹𝗶𝗰 𝗱𝗮𝘁𝗮, 𝘄𝗵𝗲𝗻 𝗮𝗻𝗮𝗹𝘆𝘇𝗲𝗱 𝗽𝗿𝗼𝗽𝗲𝗿𝗹𝘆, 𝗰𝗮𝗻 𝗽𝗿𝗼𝘃𝗶𝗱𝗲 𝗲𝗮𝗿𝗹𝘆 𝗶𝗻𝘀𝗶𝗴𝗵𝘁 𝗶𝗻𝘁𝗼 𝗲𝗺𝗲𝗿𝗴𝗶𝗻𝗴 𝗱𝗲𝘃𝗶𝗰𝗲 𝗶𝘀𝘀𝘂𝗲𝘀 rather than just documenting them after the fact.

🚀 Starting a 30-post series: AI Production Playbook I've spent 11 years building AI systems in healthcare and life sciences — at IQVIA, J&J, Cognizant and Infosys. Not prototypes. Not demos. Systems that had to work, at scale, in regulated environments. This series is everything I wish I'd known earlier: → What actually made it to production → The trade-offs nobody writes about → The mistakes that cost us months → How to close the gap from POC to real-world impact No theory. Just real systems. ───────────────────────── 𝗦𝘁𝗼𝗿𝘆 𝟭/30: 𝗪𝗵𝗲𝗻 𝘄𝗲 𝗯𝘂𝗶𝗹𝘁 𝗮 𝘀𝗲𝗮𝗿𝗰𝗵 𝗲𝗻𝗴𝗶𝗻𝗲 𝗳𝗼𝗿 𝟱 𝗺𝗶𝗹𝗹𝗶𝗼𝗻 𝗽𝗿𝗼𝗱𝘂𝗰𝘁 𝗰𝗼𝗺𝗽𝗹𝗮𝗶𝗻𝘁𝘀 Imagine being a brand manager at a global pharma company. Your product has a spike in quality complaints. You need to know why — fast. But the data? Spread across 200+ brands. Analyzed by different teams. Never connected. No unified view. This was the problem we walked into. 𝗧𝗵𝗲 𝘀𝗰𝗮𝗹𝗲: • ~5M historical product complaints and adverse event documents • 15–20K new documents arriving every week • 200+ brands • Zero unified structure Nothing was connected. No one had the full picture. 𝗪𝗵𝗮𝘁 𝘄𝗲 𝗯𝘂𝗶𝗹𝘁 We started with a Q&A approach. Quickly realized that was the wrong frame. What users actually needed: a configurable search engine + topic clustering. A brand manager could define a time window (say, last 6 months), filter to their sub-brand, and immediately see: → The most frequent complaint phrases → How topics co-occurred — visualized as a circular dendrogram Not a chatbot. But a dashboard with a lens on their own data. 𝗧𝗵𝗲 𝘂𝗻𝗲𝘅𝗽𝗲𝗰𝘁𝗲𝗱 𝗹𝗲𝘀𝘀𝗼𝗻: 𝗿𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹 > 𝗺𝗼𝗱𝗲𝗹𝗹𝗶𝗻𝗴 Everyone wants to talk about the model. Our biggest bottleneck was retrieval quality. We chose Elasticsearch over chasing the "perfect" vector DB (which barely existed then). Here's why: • Exact keyword matching still mattered in complaint language • Domain-specific terminology couldn't be lost in embeddings • Interpretability was non-negotiable in a regulated environment The right tool for the context beats the fashionable tool every time. Maybe I would have chosen OpenSearch now. 𝗧𝗵𝗲 𝗰𝗵𝗮𝗻𝗴𝗲 𝘁𝗵𝗮𝘁 𝗺𝗼𝘃𝗲𝗱 𝘁𝗵𝗲 𝗻𝗲𝗲𝗱𝗹𝗲 Even after nailing retrieval, users told us: "The top results aren't always useful." One small change fixed it: instead of surfacing full documents, we highlighted only the relevant text snippet within each result. Result → 1.5x increase in user engagement within a month. Not a new model. Not a re-architecture. A UX decision. ───────────────────────── But here's what we got wrong: our clustering approach was static — and it slowly broke as data scaled. Post 2: what happened when our topic clusters stopped making sense, and how we rebuilt them. If you work in AI, data science, or healthcare tech — this series is for you. #HealthcareAI #AIProduction #AIProductionPlaybook