Why read a transcript instead of a list of tips
Advice about chat is almost always written as rules: be concise, set expectations, do not make promises you cannot keep. The rules are right and they are nearly useless on their own, because the hard part is never knowing the rule. It is noticing, in the middle of a conversation, that this is the moment the rule applies.
So this guide does the opposite. Below is one ordinary conversation — a billing question of the kind most teams see several times a day — written out in full, with a note after each turn about what that turn is doing and what it would have cost to get it slightly wrong. Nothing here is dramatic. That is deliberate: the conversations worth studying are the routine ones, because they are the ones you will have ten thousand of.
Read it once straight through for the shape, then again for the notes. If you are training somebody, this is a better first exercise than any checklist.
The conversation
Visitor: “I was charged twice this month.”
What is happening: the visitor has opened with a conclusion, not a question, and the conclusion contains an accusation. They are not asking whether they were charged twice; they have decided. Anything that sounds like doubt in the next message will be read as a defence of the company. Note also what is missing: no account, no dates, no amounts. You need all three and you cannot ask for all three at once without sounding like a form.
Agent: “That should not happen and I can check it right now. Can you confirm the email on the account?”
What is happening: three things, in a deliberate order. It validates the concern before asking for anything, which costs six words and buys the rest of the conversation. It states an action in the present tense — I can check it right now — which is the single most reassuring thing you can say to somebody who expected a fight. And it asks for exactly one piece of information, the one that unlocks everything else.
What the wrong version costs: “Can you provide your account email, the dates of the charges and the last four digits of your card?” is the same request and a much worse message. It reads as a form, it front-loads work onto an annoyed person, and it has not acknowledged the problem at all. The visitor who was mildly irritated is now dealing with a bureaucracy.
The middle, where most chats are won or lost
Visitor: “[email protected]”
Agent: “Thanks. Looking now — give me about a minute.”
What is happening: this is the most skipped message in live chat and one of the most valuable. The agent is about to go quiet for sixty seconds. Unannounced silence in a chat window feels roughly three times as long as it is, and a visitor who thinks they have been abandoned will either leave or send a second, angrier message. Announcing the gap and naming its length converts dead air into a wait the visitor has agreed to.
The refinement: give a number you can beat. “About a minute” and returning in forty seconds reads as competence. “Just a second” and returning in ninety reads as a lie, even though ninety seconds is a perfectly reasonable time to look something up.
Agent: “Found it. There are two charges on the 3rd, both for the same plan. That is our error, not yours — I have refunded the second one.”
What is happening: the finding is stated before the explanation, which is the right order almost every time. The visitor's question was was I charged twice, and they get the answer in the first four words. It also assigns fault plainly and without ceremony. There is no paragraph about why it happened, because nobody has asked yet, and volunteering a technical explanation at this moment reads as excuse-making.
What the wrong version costs: “I can see that during our billing run on the 3rd there was an issue with retries that in some cases caused a duplicate authorisation to be captured…” — every word of which may be true, and none of which answers the question. The visitor has to read forty words to find out whether they are getting their money back.
The part everyone forgets
Visitor: “ok when will I see it”
What is happening: notice the register has changed. No capitals, no punctuation, shorter. The visitor has relaxed — the problem is solved and they are now doing admin. This is a reliable signal and it is worth learning to read: when someone's typing gets more casual, the emotional part of the conversation is over and you can be brisker too.
Agent: “Usually three to five working days, depending on your bank. If it is not there by Friday the 12th, reply to this and I will chase it.”
What is happening: a range rather than a promise, an explicit reason for the range, and — the important part — a named date and a named next step. “Three to five working days” alone requires the visitor to do calendar arithmetic while mildly stressed, and it quietly makes the refund their problem to track. Converting it to a date and offering to own the follow-up is the difference between a resolved conversation and one that comes back next week.
What the wrong version costs: “It should be back shortly.” This is the single most expensive sentence in support. It sounds helpful, commits to nothing, and guarantees a second contact from a customer who now has no idea whether four days is normal or a problem.
Agent: “Anything else while I have your account open?”
What is happening: this is better than “anything else?” for a small reason that matters. It offers a concrete benefit for asking now — the account is already open — rather than sounding like a closing formula the visitor is expected to decline. It is also where a meaningful share of second issues surface, the ones that would otherwise have become a separate conversation tomorrow.
Visitor: “no thats great thanks”
Agent: “Sorry again for the double charge. Refund is on its way — have a good week.”
What is happening: a one-line summary of the outcome before the sign-off. The visitor will not re-read the transcript, so the last message is what they remember, and it should contain the resolution rather than only pleasantries. The apology is brief and comes once. Apologising three times in a conversation makes a small error feel like a large one.
The whole exchange is nine messages. Nothing clever happened. The agent acknowledged before asking, announced a silence, answered before explaining, converted a duration into a date, and summarised at the end. Those five habits are most of the craft.
Two of the messages above are obvious candidates for a saved reply — the wait announcement and the refund-timing line are near-identical every time. Storing them as canned responses is worth doing, and it is also where saved replies start to hurt if you are careless.
The rule that keeps them useful: save the structure, not the whole message. A saved reply that ends mid-sentence, expecting the agent to complete it with the specific date or amount, stays accurate. A saved reply that is a complete, polished paragraph gets pasted unmodified into a conversation it does not quite fit, and visitors detect that instantly — usually because it answers a question slightly different from the one they asked.
What tooling cannot do is the reading. Noticing that the visitor's punctuation relaxed, or that they opened with a conclusion rather than a question, is judgement, and it is the thing worth spending training time on. If you hand a new agent a library of saved replies and no practice at reading a conversation, you get fast, fluent, slightly wrong answers.
How to use this with your own team
- Annotate your own transcripts, not invented ones. Take three real conversations from last week, strip the identifying details, and write a note after each agent turn. The exercise is the value; the finished document is a by-product.
- Pick ordinary chats deliberately. The instinct is to review the disasters. Disasters teach you about disasters. The routine billing question teaches you about the other nine thousand conversations.
- Annotate what went right. A review culture that only marks errors produces agents who write defensively, which is its own failure mode — hedged, over-qualified messages that take longer to read and commit to nothing.
- Do it with two people. Where two reviewers disagree about whether a turn was good is precisely where your team has no shared standard yet, and that disagreement is more useful than either opinion.
- Keep the good ones. An annotated transcript is the best onboarding material there is, and it costs nothing beyond the review you were already doing.
What to measure
- Second contacts about an already-resolved issue. The clearest signal that endings are vague. If people come back to ask whether something is still coming, the closing message is not doing its job.
- Silences longer than about thirty seconds with no holding message. Readable straight from transcripts, and one of the few things you can fix with a single habit.
- How often a saved reply is sent unedited. Very high is a warning, not an efficiency win.
- Length of the first agent message. Long openers usually mean the agent is asking for several things at once, which is the form-not-a-conversation failure.
- Disagreement rate between two reviewers. Falling over time means a shared standard is forming, which is the actual goal of transcript review.