Live Chat Transcript Analysis: What to Look For and What to Do About It

Your analytics dashboard can tell you that 312 people opened the chat last week and that you answered them in a median of 48 seconds. It can't tell you that forty of them asked whether you ship to Canada, or that the bot kept getting your return window wrong. That is what live chat transcript analysis is for: reading conversations with a short list of questions in mind and turning what you find into changes on the site, in the bot and in the team.

This guide is a working checklist. For each signal you'll see what it looks like in a real chat and what to do about it. At the end we'll cover how often to review, why reading every transcript by hand stops working, and how ChatDeck's Insights tab produces an AI report on your chats. If you want the numbers side first, start with our guide to live chat metrics; this article is about what's inside the conversations.

Why transcripts beat dashboards for the "why"

Metrics are great at telling you something changed. They are bad at telling you what to do about it.

• A funnel report shows people dropping off at checkout. A transcript shows someone asking "do you take PayPal?" thirty seconds before they left.
• A CSAT trend shows satisfaction dipping. A transcript shows that every low score this week mentions the same late courier.
• A traffic spike shows more visitors. A transcript shows they arrived from a competitor's outage and want to know if you can migrate their data.

Chat is the one channel where customers explain themselves, in their own words, at the moment they're deciding. Use metrics to find where to look and transcripts to learn what's going on.

What to look for: the signal checklist

You don't need to read everything. Work through this list for the period you're reviewing and ask, for each item: is it there, how often, and is it growing?

Repeated questions your site doesn't answer

What it looks like: "Do you ship internationally?", "Is this compatible with the 2023 model?", "Where do I find my invoice?" — the same question from different people.

What to do: five identical questions in a week means the answer is missing from the site or hard to find. Add it to the product page or FAQ, and to the bot's instruction. These are the cheapest wins you'll get.

Pricing and shipping questions right before purchase

What it looks like: "What's shipping to Texas?", "Is there a discount for annual billing?", "Will it arrive before the 15th?"

What to do: treat these as sales conversations, not support tickets. They need a fast, exact answer. Check whether delivery times and costs are visible where people decide, and track how many of these chats end in an order.

Objections

What it looks like: "Seems expensive", "X does this for less", "What if it doesn't fit?", "Can I cancel anytime?"

What to do: list them and write one strong answer for each — a guarantee, a comparison, a returns policy, a reference customer. Put the most common ones on the page itself and the rest in your canned responses.

Where the AI bot fails or hands off

What it looks like: vague answers, wrong policy details, or an instant handoff on a question the bot should know.

What to do: the cause is almost always the instruction — a missing fact or rule. Add it, then check the same topic next week. Our guide to writing a customer service chatbot prompt covers how to structure it.

Repeated handoffs on the same topic

What it looks like: ten handoffs this week, seven of them about changing an order or a refund.

What to do: decide whether that topic should reach a human at all. If it must (the agent needs access to your order system), focus on cutting the wait after the handoff. If not, teach the bot. See chatbot to human handoff for designing that moment.

Chats abandoned without a reply

What it looks like: a visitor writes, waits, and leaves. The reply arrives to an empty tab, or never.

What to do: check which hours these happen in and close those gaps with schedules, mobile notifications or an AI first line. The first response time guide has practical fixes.

Night and weekend demand

What it looks like: a meaningful share of chats arrives when nobody is working.

What to do: compare chats inside and outside working hours. If one in five comes in after hours, that's a case for an AI assistant overnight, a shifted rota, or at least an honest "we'll reply at 9am". Options are covered in after-hours customer support.

Peak hours

What it looks like: agents are swamped from 11 to 1 and idle at 3.

What to do: staff to the load, not to the calendar. Sometimes moving one lunch break fixes most of it. The support team guide covers scheduling around peaks.

Complaints and mood shifts

What it looks like: a polite opening that turns into "this is the third time I'm asking". Or a general tone that's noticeably sharper than last month.

What to do: find the step where the mood broke: a long wait, a copy-paste reply, being bounced between people. Fix the step, not just the chat.

Sudden new topics

What it looks like: a topic nobody mentioned yesterday is a third of today's chats. "The login code never arrives", "card payment fails", "your competitor has 30% off, will you match it?"

What to do: act the same day. Bugs go to engineering, partner outages get a banner and a line in the bot's instruction, competitor promos go to marketing. This is the most valuable signal on the list and the easiest to miss in a monthly review.

Differences between agents

What it looks like: one agent's chats end in orders, another's end in silence. One replies in a minute, another in eight.

What to do: learn from it rather than rank people. Read the strong agent's best chats and turn their moves into shared replies. The slower agent may simply be getting the hardest chats.

Low ratings

What it looks like: a 1 or 2 after the chat closes.

What to do: read every one in full. Often the customer isn't unhappy with the answer but with the wait, or with the fact that nothing was resolved. The CSAT score guide explains how to read ratings over time.

Pages where people open the chat

What it looks like: most chats start on the checkout page, or on one particular pricing tier.

What to do: a page that generates a lot of questions is a page to rewrite. No agent fixes confusion as well as a clear sentence in the right place.

Don't try to fix every item at once. Pick two or three changes from each review, ship them, and check next week whether those topics went quiet.

How often to review, and who should do it

A rhythm that works for most small and mid-size teams:

• Daily, five minutes: unanswered chats and anything that looks like a new topic.
• Weekly, thirty minutes: topics, bot failures, complaints, low ratings.
• Monthly, one meeting: what to change on the site, in the schedule, in the bot. Support, marketing and whoever owns the product.

The support lead is usually the right person to run the review, because they have the context. But the findings have to reach the people who can act on them — whoever edits the site, the developers, the founder.

The trap: reading 500 chats by hand

With a few dozen chats a week, reading them all is fine. Past a few hundred, manual review breaks down:

• Time. 500 chats at 90 seconds each is over twelve hours. That review happens once and then never again.
• Fatigue. By chat 150, you're skimming, and the important signal slips past.
• Memory bias. The loud, recent conversation feels bigger than twenty quiet identical ones.
• Counting by feel. "Lots of people asked about shipping" — how many? Without numbers you can't compare weeks.

The sensible split: let software do the reading and the counting, and keep the decisions for yourself. That is the idea behind Insights in ChatDeck.

ChatDeck Insights: an AI report on your chats

Insights is a tab in the ChatDeck backoffice, next to Analytics. It's available to owners and admins, on every plan. For a walkthrough with screenshots, see the Insights tutorial.

Running a report

The main flow is a report you run yourself:

1. Pick a period: Today, Yesterday, 3 days, Week or Month.
2. Pick the AI model. You see the estimated price before you run it.
3. Optionally, say what to focus on — for example, "compare delivery complaints by region and make a chart".

You can re-run the same period with a different model and compare the results. Every report shows which model produced it and what it actually cost.

Automatic reports — daily, weekly or monthly — are optional. You can turn them on once there's enough chat history to cover a full period. When a report is ready, you get an email with a button that opens it in the app.

What a report contains

• Key numbers for the period.
• A summary of what mattered.
• What visitors ask about — topics with counts and a chart.
• Real quotes from visitors, each linking to its chat.
• Where the bot failed, with ready-to-use text and an "Add to instruction" button.
• FAQ candidates — questions worth answering on your site.
• Sales opportunities — chats where someone was close to buying.
• Complaints and mood.
• Operator performance, by name.
• Signals worth your attention.
• What to do next — concrete steps.

Numbers you can trust

The usual worry about AI analytics is that the model makes up figures. Insights is built so it can't: every number and every chart comes from data our code counts from your conversations. The model only chooses what to show and groups chats into topics. Quotes are real messages, and each one links back to the original conversation.

Privacy

Email addresses and phone numbers are stripped from the text before it is sent to the model, and there's a consent checkbox before your first run.

Cost

Insights uses the same GPTunneL key as the AI assistant, and you pay the model provider for tokens directly. It stays cheap because each chat is first condensed into a short line by an inexpensive model, and later reports reuse that work. A weekly report on a few hundred chats with a mid-tier model typically costs somewhere between a few cents and tens of cents.

Your first report, step by step

1. Connect a model key. If your AI assistant is already running, you're done. If not, follow Connect the AI assistant to add your GPTunneL key.
2. Open the Insights tab in the backoffice, next to Analytics.
3. Pick a period. Week is a good first choice: enough chats to see patterns, not so many that the report gets broad.
4. Pick a model and check the estimated price. A mid-tier model is a sensible start.
5. Add a focus if you have a specific question: "why do people leave after asking about shipping?" or "what are people asking about the new plan?"
6. Tick the consent box. It's a one-time step before the first run.
7. Run the report. Read "What to do next" first, then "Where the bot failed".
8. Act on it. Use "Add to instruction" for the answers that fit, and pass FAQ candidates to whoever edits the site.
9. Repeat next week and compare. If the reports earn their place, switch on automatic runs.

The Insights tutorial walks through every step with screenshots. For the speed and rating numbers that sit alongside, see Read your analytics.

Run the same week with a cheap model and a stronger one. Comparing the two reports is the quickest way to learn whether the bigger model is worth it at your volume.

Frequently asked questions

How is an AI insights report different from chat analytics?

Analytics counts things: chats, response times, response rate, ratings. An insights report explains the content: what people asked about, where the bot went wrong, what they complained about, where a sale was within reach. In ChatDeck, the numbers in the report are still counted by code, not by the model.

How many chats do I need for a useful report?

You can run one on a single day, but patterns get clear from a few dozen conversations upward. That's why a week is a good first period.

Can the AI make things up?

Not the numbers: every figure and chart is computed by our code from your data. The model groups chats and writes the conclusions, and every quote is a real message linked to its conversation. It's still worth spot-checking the key conclusions against the original chats — it takes a couple of clicks.

How much does a report cost?

It depends on the model and the number of chats. A weekly report on a few hundred chats with a mid-tier model typically costs a few cents to tens of cents. You see an estimate before running and the actual cost on the finished report, and you pay the model provider directly.

Who can see Insights?

Owners and admins of the workspace, on every plan.

Can reports run automatically?

Yes — daily, weekly or monthly. You can turn this on once there's enough chat history for a full period, and you'll get an email when each report is ready.

Where to start

Run through the checklist once by hand on last week's chats — you'll almost certainly find a question your site should answer. When there are more chats than you want to read, hand the reading to Insights. Not using ChatDeck yet? Create a free workspace; setup takes a few minutes.