How to review customer chats for quality in 30 minutes a week

Why read chats at all

Numbers tell you how many chats you had and how fast you answered. They do not tell you that a colleague quoted last month's price, that customers keep asking a question nobody has written an answer for, or that the Arabic replies read like a translation. Only reading tells you that, and ten chats a week is enough to see patterns within a month.

The 30-minute routine

Use the same day and time each week, such as Sunday at 11:00. Whoever runs it rotates, so it is not always the manager judging the team.

The five points

Score each chat yes or no on:

1. Fast enough: the first useful reply came within your target.
2. Correct: every fact and price matched the published version.
3. Complete: the customer's actual question was answered, or they were told what happens next and when.
4. One voice: one person owned the chat; no double or contradictory replies.
5. Closed well: the chat ended with a clear outcome, and any promise was noted.

Five yes or no answers per chat, ten chats: fifty marks. You do not need a percentage; you need to see which point fails most.

A scoring sheet to copy

Turning findings into fixes

Each failing point has a typical fix:

• Not fast: look at when the slow chats arrived and adjust the rota or the reply hours in your welcome message (response-time targets).
• Not correct: find the source of the wrong fact. Usually a price or policy changed and the published answer did not; publish the correction.
• Not complete: the question probably has no written answer. Write one and publish it (the 20 answers to start with).
• Not one voice: remind the team of "take it before you type", and check that presence is visible to everyone (five rules against double replies).
• Not closed well: agree a closing line and the habit of a note for every promise.

Limit yourselves to two fixes a week. Ten fixes announced at once become none.

Where the tools help

In the StayReplied unified inbox every website conversation is in one searchable list with the full history and notes, so picking and reading ten takes minutes. Tags are on every plan, as listed on the pricing page, so you can tag reviewed chats and avoid picking the same ones twice. When a fix is a new or corrected answer, documents and price records are drafted, reviewed and published, with every document version kept and rollback available, as the docs describe. If you use AI drafts or replies (no plan includes AI today; they need your own Anthropic key), review those chats with the same five points, and use StayReplied's list of questions the assistant could not answer, with suggested facts a person can approve, as a second source for "not complete". A wrong AI answer almost always traces back to a missing or outdated published answer, which is the same fix.

Keep it from becoming blame

• Review the work, not the person. Hide names if that helps the team relax.
• Start with something done well. One good chat read aloud teaches more than three bad ones.
• Fix the system first. Most errors come from missing or outdated answers, not careless staff.
• Let the team pick chats sometimes. People learn from the ones they found hard.

When this is not for you

If you get fewer than ten chats a week, read all of them instead of sampling, and skip the scoring sheet. And if your support is mainly long e-mail threads, a review routine still helps, but a tool built for e-mail tickets may fit better; when a general help desk is enough helps you decide. E-mail in StayReplied is coming soon.