Cx Tag Taxonomy Hygiene

Tag and category hygiene

Tag sprawl is the normal end state of any helpdesk more than a year old. Free-text tagging, several people with admin rights, campaign tags nobody removed, and three spellings of the same concept.

The reason to fix it is not tidiness. It is that every report built on tags inherits their inconsistency, usually without saying so, and the resulting numbers are quietly wrong in ways nobody can see from the chart.

Measure the damage first

Before proposing anything, quantify:

• Tag count versus meaningfully-used tag count — how many are applied more than a handful of times. The gap is usually enormous and is the argument for doing the work.
• Tags per conversation. A rising average means agents are hedging because they cannot find the right one.
• Untagged share. If a third of conversations carry no tag, tag-based reporting describes a self-selected two-thirds. This single number often invalidates a report someone is already using.
• The near-duplicate clusters: case variants, singular/plural, hyphen/underscore, abbreviations, translations, misspellings, and payment_issue next to payment-issues next to Payments.
• Concurrent-lifetime overlap. Two tags that were both actively applied in the same period are genuine duplicates. Two where one stopped when the other started are a rename, and merging them without noticing creates a false step change in every historical series.

Decide what tags are for

Most sprawl comes from one taxonomy trying to do several jobs at once. Separate them, because they have different owners, lifespans and rules:

• What the contact was about — the contact driver. Should be a controlled, mutually exclusive list, and it is what most reporting actually wants.
• What happened to it — escalated, refunded, bug-linked. Workflow state.
• Attributes — VIP, vulnerable customer, regulated, language.
• Temporary markers — a campaign, an incident, an experiment. These need an expiry date at creation, and their absence is the largest single source of sprawl.

Trying to encode all four in one flat tag list is why the list is unmanageable.

A contact-driver taxonomy in particular has its own design problems — deriving categories that name a removable cause rather than a topic, and keeping the "Other" rate honest — which are a bigger question than tag hygiene. Hygiene is about making the mechanism reliable; taxonomy design is about making the categories worth reporting.

Rules that keep it clean

• Controlled list, not free text, for anything used in reporting. Free-text tags are a reporting dead end from the first day.
• One naming convention, mechanically enforced: case, separator, singular or plural. Pick and apply.
• Mutually exclusive within a dimension. A conversation should carry exactly one contact driver. Multiple drivers means the reporting cannot sum, and every percentage built on it is over 100 or arbitrarily attributed.
• Every tag has an owner and a definition. A tag whose meaning is folklore will be applied inconsistently within a month, and the definition needs to be visible where agents tag, not in a document.
• Expiry dates on temporary tags, enforced by review.
• A creation gate. If anyone can create a tag from the ticket view, the list grows without bound. If nobody can, agents will use the nearest wrong tag. The workable middle is a request path with a fast turnaround plus a visible "suggest a tag" channel — and treat frequent requests as evidence the taxonomy has a gap.

Consolidating without breaking history

This is the part that goes wrong.

Merging tags rewrites history. After merging payment-issue into payments, last quarter's report of payments changes. Anyone comparing to a previously published figure will find a discrepancy and lose confidence in the reporting.

So:

• Map, do not just merge. Keep a documented mapping from old tags to new, with the date it took effect.
• Prefer a reporting-layer mapping to a destructive retag where you can. It is reversible; a bulk retag usually is not.
• If you do retag, treat it as a reviewed mutation — a plan reviewed before it runs, a bounded batch size, an audit record per change, and the ability to resume. Bulk tag operations across a helpdesk are rarely reversible and frequently trigger automations and triggers as a side effect. Check what fires before running anything.
• Announce the change, with the date and the mapping, and mark it on dashboards.
• Do not retag historical conversations to match a new taxonomy unless you have a strong reason. Historical tags record what someone thought at the time, which is itself information — and a retro-fitted taxonomy makes a trend look like it always existed.

Improving tagging behaviour

Tag quality is mostly a design problem, not a compliance problem:

• Fewer options tagged correctly beat more options tagged badly.
• Put the definition where the agent tags, not in a wiki.
• Do not make tagging a QA criterion without checking the tag list is usable first. Penalising agents for a broken taxonomy produces tagging that is compliant and meaningless.
• Automate what can be automated — channel, language, and attributes derivable from the record should never be typed by a human.
• Audit agreement periodically: have two people tag the same conversations independently and measure how often they match. Low agreement means the categories overlap, and that is a taxonomy fix, not a training fix.

Present results to the user

1. The damage — tag count vs used count, tags per conversation, untagged share, duplicate clusters. Lead with the untagged share, because it bounds every existing report.
2. The four dimensions, and which of them the current list is conflating.
3. A proposed consolidated list, with definitions and owners.
4. The mapping from current tags to proposed, flagging renames versus true duplicates by their concurrent-lifetime overlap.
5. The impact on existing reports — which series change and by how much.
6. The migration plan, stating what is reversible and what is not, and what automations the change would trigger.
7. Governance — creation gate, expiry policy, review cadence, and who owns it.