Practical AI · 3-minute read
A complaints folder is easy to treat as a collection of isolated bad afternoons. Read together, the messages may point to a service problem worth fixing. They may also point to one unhappy person who has now emailed seven times.
AI can help sort the material. It needs boundaries around what it counts, what a category means and what the evidence can actually support.
Start with a question the records can answer
Cedar Junction Clinic is fictional. Its administrator practices with 20 invented comments about parking, appointment reminders and reaching the front desk. No patient records or clinical decisions are involved.
A useful question is, ‘Which administrative problems appear in these comments?’ A less useful request is, ‘Explain why our reception staff keep disappointing patients.’ That second prompt has already assigned blame before anyone reads a sentence.
Give each comment an ID. Keep the original wording available to authorized reviewers. Decide whether you are counting messages, unique issues or individual people. If you cannot establish that two messages concern the same incident, flag the uncertainty instead of guessing.
Ask for categories with evidence
Review these fictional or approved comments, identified by ID. Propose issue categories with short definitions. For each category, list supporting IDs, a brief supporting passage and the count of distinct comment IDs. Allow multiple categories per comment and state that totals may overlap. Include unclear items and evidence that challenges your interpretation. Do not infer motives, diagnoses, employee fault or causes. End with questions for human investigation.
Review the suggested categories before accepting the counts. ‘Nobody called me back’ and ‘the reminder arrived late’ both concern communication, but they may require different fixes. Combining them too early can conceal the practical problem.
Government guidance on thematic analysis explains that finding themes involves interpreting related ideas, not simply counting repeated words. It also emphasizes considering how the data was collected. [1]
Check the denominator and the missing voices
If eight of the 20 invented comments mention parking, report ‘8 of these 20 comments mention parking.’ That does not mean 40 percent of all patients have a parking problem. People who complain are not automatically a representative sample, and several issues can appear in one comment.
For a small set, read every comment and verify each assignment and count. Keep rare but serious concerns visible. Ask whose feedback is missing because of language, access or the collection method.
Choose one finding to investigate with the people doing the work. A pattern can suggest a question; it does not prove the cause. Use your established privacy and complaint-handling process for real material. Removing names alone does not establish that sensitive feedback is safe to upload.
Your turn: What recurring complaint would be most useful to understand before your team writes another apology?
Source [1]: GOV.UK: Analyse your data, qualitative analysis guidance. Theme development, coding and attention to data sources. The clinic, comments, counts and prompt are fictional or original guidance; no clinical advice is offered.
