Label setup checklist: minimal taxonomy, usage criteria and errors that dirty your reports

Tags are the only cheap way to turn a tray full of conversations into data that can be read. Without tags, every report you pull will say the same thing: messages came in, messages went out. With tags applied incorrectly, the report says something worse: it says something false and you will believe it. This page is a checklist to tick off, in order, before signing off on your team's tag setup. It is not a list of ideas: each item says what to check and how to know it is right.

Order matters. First you define the taxonomy (which tags exist and what they are for), then the criteria (when each one is applied and who applies it), then the operation (how they are assigned without slowing down support) and finally the verification (how to know the reports came out clean). If you start at the end, you will end up with tags nobody uses or with reports nobody looks at.

Stage 1: define the minimum taxonomy

The minimum taxonomy is the smallest set of tags that answers the questions your business actually asks itself. It is not the most complete set: it is the smallest one that works. If you have 40 tags and nobody fills them in, you have zero tags with more steps. The practical rule: each tag must be able to answer a specific question that someone on the team asked themselves at least once in the last month.

• Choose one main axis and do not mix it with others. The main axis is usually the conversation status (new, in progress, waiting for customer reply, closed, transferred).
• Add a topic or contact reason axis, but keep it limited: between 5 and 10 reasons that cover 80% of what comes in.
• Add an outcome or result axis if your business needs it for measurement (sale, inquiry resolved, not interested, complaint).
• Do not add a channel axis if you already have only one channel. If you handle WhatsApp, Instagram and Messenger, then it does make sense.
• Do not create tags for one-off cases. If a tag was used once in a month, it was not a tag: it was a note.

Stage 2: set usage criteria

A tag without written criteria is used differently depending on who applies it. That is the origin of almost all dirty reports. The criteria must answer three things: when it is applied, when it is removed and who can apply it. If any of the three is left to interpretation, the report will mix criteria and you will not be able to compare months.

• For each label, write a sentence describing when it is applied, with an example and a counterexample. 'Angry customer' doesn't work; 'customer who asks to speak with a supervisor or mentions a delay longer than promised' does.
• Define whether the label is cumulative or exclusive. 'New' and 'in progress' cannot coexist in the same conversation; 'complaint' and 'in progress' can.
• Define who can apply each label. If anyone can apply any label, the outcome label will be applied by whoever handles the conversation, not by whoever closes it.
• Define the moment of assignment: at opening, at closing, or during. Status labels are applied at opening and updated; outcome labels are applied at closing.
• Define what happens when a conversation is reopened. If you don't define it, you'll have closed conversations with an open status label and reports that contradict each other.

Stage 3: operate without slowing down service

Label configuration falls apart when labeling costs more than serving. If applying a label takes three clicks and a decision, it won't get done during peak hours. The goal of this stage is for labeling to be part of the normal response flow, not a separate task left for later. And what's left for later, in an inbox, doesn't get done.

• Keep frequently used labels at hand and hide the occasional ones. If all are at the same level, the team picks the first one they see.
• Use templates or quick replies that already come with the associated label when the case is repetitive.
• Define a fixed moment for labeling. The most reliable is usually when closing the conversation, not when opening it.
• Check whether the inbox allows multiple people to work on the same conversation without stepping on each other. If not, define one owner per conversation.
• Don't use labels as a personal agenda. If a label is 'I'll look at it Monday,' it's not a team label: it's a task.

Stage 4: verify that the reports are clean

A clean report is one where two people looking at the same number understand the same thing. Verification isn't looking at the report: it's looking at a sample of conversations and comparing what they say with what the report says. If the sample doesn't match, the problem is in the labels, not the report. And if the problem is in the labels, it won't be fixed by looking at the report.

• Take a sample of closed conversations from the last week and label them again blindly. Compare with what they already had.
• Count how many closed conversations were left without an outcome label. If it's more than 10%, the closing criteria aren't clear.
• Count how many conversations have two labels from the same axis. Each one is a criterion that wasn't followed.
• Check whether reason labels are concentrated in a single one. If 80% is 'general inquiry,' the axis isn't measuring anything.
• Compare this month's report with last month's. If a number moved a lot without anything happening, check labels before drawing conclusions.

The mistakes that dirty your reports

There is a handful of mistakes that repeat and ruin tag reports more often than anything else. None of them are technical: they are about criteria. Recognizing them is half the work, because once you know which one is yours, the fix is limited.

• Mixing axes in one tag. 'Urgent closed sale' mixes result with priority. When you want to count sales, you also count urgencies that were not sales.
• Using tags as notes. If the tag says 'customer asked for invoice A', it is not a tag: it is a piece of data that belongs somewhere else.
• Changing criteria mid-month. The month's report mixes two definitions and cannot be compared with anything.
• Tagging at the end of the day from memory. What is tagged from memory is tagged wrong or not tagged at all.
• Letting each person create their own tag. In two weeks you have five tags for the same thing and none of them comparable.
• Not retiring old tags. A tag that is no longer used keeps appearing in historical reports and confuses the reading of the present.
• Tagging the conversation by what the customer said and not by what happened. 'Asked for a discount' is not the same as 'was given a discount'. They are two different axes.

What to do when something fails

When a tag report does not reconcile, the temptation is to add tags. It is almost always the opposite: you need to remove or clarify. The procedure is always the same and does not depend on the tool. First the problem is isolated, then you decide whether it is about criteria or operations, and only at the end do you touch the configuration.

1. 1Choose a tag that is giving a strange number and take 20 conversations tagged with it.
2. 2Read the 20 and mark how many really meet the written criterion. If fewer than 15 meet it, the problem is about criteria or training, not the report.
3. 3If the criterion was not written, write it now with an example and a counterexample, and communicate it to the team before looking at the number again.
4. 4If the criterion was written and was not met, review whether tagging is too costly in the flow. If it is costly, simplify: fewer tags, more at hand.
5. 5If the criterion was met and the number is still strange, review whether two tags overlap. Overlap is fixed by merging or retiring, not by adding.
6. 6Set a cutoff date. Criteria changes apply from the following Monday, not from the moment you decide them.
7. 7Measure again after two full weeks. Before that, any comparison mixes the old criterion with the new one.
8. 8If after two weeks the number still does not reconcile and the criterion is clear and met, escalate: it may be a problem with the tool or with how the data is being read, and that has to be seen by whoever administers the account.

About timing: how long it takes to get a label setup in order depends on the size of the team and how many conversations come in per day. There is no deadline that can be promised. What can be said is which signals indicate it is on the right track: the team labels without asking, two people label the same conversation the same way, and the weekly report is read in five minutes without questions. When those three signals do not appear after a couple of weeks of work, it is worth checking whether the problem is with the criteria or whether the tool does not support the flow.

What a healthy taxonomy looks like, concretely

So the checklist does not stay abstract, this is what a minimal applied taxonomy looks like. It is not the only possible one, but it meets the rules: each label answers one question, the axes do not mix, and the team can explain what each one is used for. If your taxonomy looks like this, you are on the right track. If it has three times as many labels, you probably have several that do the same thing.

An inbox with AI suggestions and well-configured labels becomes much more useful, because suggestions can be sorted and filtered by label and the team does not start from scratch on each conversation. Tools like Wando, which suggest replies and learn from the business's real replies, perform better when the taxonomy is already defined: AI does not fix misapplied labels, but it makes good use of the ones that are clear. If you want to see how it applies in practice, you can try the inbox with the 14-day free trial, no card, and configure the labels on real conversations.

Final checklist to print

This is the summary of everything above, in the order in which it is best to review it. It can be checked off in a single pass if the setup already exists, or used as a guide if you are building the labels from scratch. Each item can be verified by looking at the system and a sample of conversations, without depending on anyone else.

• Every active tag answers a specific question that someone asked themselves last month.
• No tag mixes two axes (status, reason, outcome, priority).
• Every tag has a written usage sentence, example, and counterexample.
• It is defined who can apply each tag.
• It is defined when each tag is applied and when it is removed.
• It is defined what happens when a conversation is reopened.
• Frequently used tags are at hand and occasionally used ones are not.
• Closed conversations without an outcome tag are less than 10%.
• No conversation has two tags from the same axis.
• No reason concentrates more than 80% of conversations.
• There are no active tags that have not been used in the last month.
• The criteria were not changed halfway through the period you are measuring.
• You can take 20 tagged conversations and at least 15 meet the written criterion.
• Two people on the team tag the same test conversation the same way.
• The weekly report is read in five minutes without questions.