AI at work · Guide
Have AI Read Every Customer Conversation and Tell You the Top Five Problems
By Andrew Ling ·
Short answer
AI can read every customer conversation you export, sort each into a fixed list of problem categories, and give you the five that come up most, with counts and real quotes. The method matters more than the tool: let AI propose categories from a sample, fix the list yourself, have AI label each conversation with one category and a quote, then count the labels in a spreadsheet. Check twenty labels by hand before acting.
Can AI really read all my customer conversations and find the patterns?
Yes. AI does this better than people, because people remember the loudest complaint and AI reads the four hundredth message as carefully as the first. Picture a one-page report: the top problem last month was "customers cannot find the rebooking link", here is how many conversations raised it, and here are five things customers actually wrote.
Most owners already have this information in the CRM inbox, the support email, the chat widget and the call transcripts. Nobody reads it all, so decisions follow whatever the last angry customer said.
Building one customer's history is a different job, covered in how to get every customer's full story in one place. This one looks across all of them at once.
What is the method that makes the answer trustworthy?
Paste 300 conversations into a chat and ask "what are the top five problems?" and you will get a confident answer you cannot check, and AI is weak at counting across a long input. Use this five-step method instead, on any tool:
- Discover. Give AI about 50 conversations and ask for 8 to 12 problem categories, each with a one-line definition and an example
- Decide. Edit the list yourself: merge overlaps, use your own words, add "No problem" and "Other". This is your codebook and it stays fixed
- Label. AI reads every conversation and returns one row each: ID, category, a short direct customer quote, and whether it looks resolved
- Count. Put the rows in a spreadsheet and count them with a pivot table or COUNTIF. The spreadsheet does the math, not the AI
- Check. Read the originals behind 20 random rows. If more than two or three labels are wrong, sharpen the definitions and label again
The quote column is what makes this useful. A category name tells you where to look. Real sentences from customers tell you what to fix, and they expose a wrong label in seconds.
What are the ways to run it, from easiest to most ambitious?
- 1. Upload an export
- What you getA one-time top five report with counts and quotes
- How it worksExport conversations, upload them to an AI assistant and walk it through the five steps
- What it takesAn hour or two. A paid AI plan helps with larger files
- 2. Label inside a spreadsheet
- What you getA category beside each message that anyone can filter
- How it worksOne conversation per row, with an AI formula in the next column applying your codebook
- What it takesAn afternoon and a Google plan with the AI function
- 3. A Monday report
- What you getThe top five for last week, sent to your inbox every Monday
- How it worksA scheduled n8n workflow pulls messages from your CRM's API, labels, counts and emails the summary
- What it takesA day or two of setup with Claude Code or n8n, plus a read-only API token
- 4. A problem table over time
- What you getEvery label kept, so you can see whether a fix made a problem shrink
- How it worksThe weekly job saves its rows to a table and a small dashboard charts each category
- What it takesA few days, a database, and someone to review the build
How do I try it this week with an export and an AI assistant?
Start with whatever export you can get: help desk tickets, chat logs or an inbox export. If your CRM has no message export button, rung three covers the API route.
Claude takes up to 20 files per chat, at up to 500MB each, and with its code execution and file creation feature turned on it can work through CSV files directly. So it can count in code rather than by reading. Ask it to: "Load this CSV, apply this codebook to the message column, save a new CSV with category, quote and resolved columns, then count each category." Then do the 20-row check on the file it gives back.
Before you upload, strip card numbers, health details and addresses. On a personal Claude plan, check the model improvement setting under Settings, then Privacy. On Claude Team, chats are not used for training by default.
Can I do the labeling inside a spreadsheet?
Yes, and for a team that lives in spreadsheets it is the friendliest version. Google Sheets has an AI function, =AI("prompt", range), on eligible Google Workspace and Google AI plans. Put each conversation in a row, write your codebook into the prompt and point it at the message cell. Sheets generates the first 350 selected AI cells per run and has usage limits, so label a large file in batches.
Excel's COPILOT worksheet function was retired on September 14, 2026: old cells keep cached results, but recalculating returns an error. The Copilot pane is the route now, which works more like rung one than a formula.
Whichever you use, keep the count in a pivot table. Pivot tables do not guess.
How would a weekly report pull messages straight from my CRM?
This is where it becomes a habit. I work in GoHighLevel every day, so here is the shape of it there. Any CRM with an API works the same way under different names.
- Read-only access
- What it does in GoHighLevelA token with the conversations/message.readonly scope for the sub-account
- What to watchTick nothing that can send or edit
- Pulling the messages
- What it does in GoHighLevelThe export messages endpoint returns the whole location's messages by channel, contact or date range, 10 to 1,000 per page
- What to watchEach page cursor expires two minutes after the last request
- What it does in GoHighLevelEmail is left out unless you ask for it with channel set to Email
- What to watchRun a second pull for email
- Rate limits
- What it does in GoHighLevel100 requests every 10 seconds and 200,000 a day by default
- What to watchA loop with no pause can hit the burst limit
In n8n the workflow is short: a Schedule Trigger on Monday morning, an HTTP Request node that pulls last week's messages, a step that groups them by conversation, then a Text Classifier node, which sorts each item into categories you define (your codebook). An Information Extractor node pulls the customer's quote into its own field, and a final step counts categories and emails the top five. Claude Code can write the workflow, or the same job as a script, from a description of the five steps.
For large volumes, Anthropic's Message Batches API runs labeling jobs at half the normal price within 24 hours. A Monday report can wait.
What does the top rung look like?
Keep every labeled row in one table: conversation ID, date, channel, category, quote and resolved flag. Now you can ask what a single run never could: did the rebooking problem drop after you moved the link? Which problem is growing?
That table can feed a page like the one in building a custom dashboard with AI: one chart per category by week, with last week's quotes under each. A ranked, living list of what customers struggle with, with evidence.
Where does this go wrong?
- Categories that drift. If AI invents new categories every run, weeks cannot be compared. Change the codebook on purpose, not by accident
- One conversation, five issues. Decide up front whether you label the main problem only or allow a second category
- Counting by vibes. Never take a count from the AI's prose. Count rows
- Grading staff instead of problems. Scoring people from this changes how they write to customers. Keep it about what customers struggle with
- Acting on a quote out of context. Read the full conversation before you change a policy because of it
When do you not need this?
If you get fewer than about 50 customer conversations a month, read them yourself. An hour of reading teaches you more than a report.
If you already know the top problem and have not fixed it, you do not need AI to confirm it. Fix it, then run the analysis to find the next one. If most contact is in person, write down the questions people ask at the counter for two weeks. The method works on a notebook too, and it is a good step before you decide what to automate first.
Frequently asked questions
How many conversations do I need before AI analysis is worth it?
Roughly 50 or more. Below that, reading them yourself is quicker and catches things a category list misses.
Why should I count the results in a spreadsheet instead of asking the AI?
AI is good at labeling one conversation at a time and unreliable at tallying hundreds. Have it return one labeled row per conversation and let a pivot table or code count.
Can this include phone calls?
Yes, once they are transcripts. Many phone systems and call recorders produce one per call. Treat each as one conversation and use the same codebook.
Is it safe to upload customer conversations to an AI assistant?
It can be, if you remove details the job does not need, such as card numbers and health information, and check how your plan treats your data. Keep a person reading the source before acting on the report.
Keep reading
AI at work · Guide
How to Get Every Customer's Full Story in One Place With AI: Messages, Calls, Tickets and Problems
How to have AI pull a customer's texts, emails, calls, meetings and support issues into one timeline with a short written story on top, from a free version you can try today to a live timeline built with Claude Code.
Read
AI at work · Comparison
ChatGPT vs. Claude for a Small Business: Which One Should Your Team Pay For?
How ChatGPT and Claude compare for a small team on price, business plans, images, connectors and building tools, which one to pay for in which situation, and how to test both on your own work before you buy.
Read
AI at work · Guide
AI Can Build You a Custom Dashboard in an Afternoon. Here Is How
Four ways to get a dashboard of the numbers you actually care about built by AI, from a chart made out of one spreadsheet export to a live page that reads your CRM, with what each one takes and where each one breaks.
Read
AI at work · Guide
What Can AI Actually Do for a Small Business Right Now? From the Basics to What Most Owners Never Try
The everyday jobs AI does well for a small business, nonprofit or church, the bigger builds most owners never try, such as custom dashboards and a full history of every customer, what you still have to check, and what it costs to start.
Read
More in AI at work.
Free newsletter
Get the weekly AI update for small teams
One short email a week: what changed in AI and in the tools small teams run on, what it means for you, and what to do about it. No spam, unsubscribe any time.
