Two tickets can describe the same symptom without describing the same problem. A customer who submits a web form and then emails the helpdesk may have created a duplicate. Two customers who report failed deliveries may need entirely separate investigations. A useful AI triage workflow helps an agent distinguish these situations before any records are combined.
The starting objective should be a better review suggestion, not a smaller ticket count. Closing records can make a queue look tidy while hiding unresolved work. Design duplicate detection around the customer obligation that must survive the decision: who needs help, with which event, and what still needs to happen.
Define three relationships before building a matcher
Give reviewers separate options for duplicate, related and independent. Define a duplicate as a second record of the same support request, where retaining both creates unnecessary parallel handling. Define related tickets as distinct requests sharing a possible cause or investigation. Independent tickets require separate handling even when their wording overlaps.
In a hypothetical Durban equipment supplier, a customer submits a portal request about an undelivered replacement pump, then forwards the same order confirmation by email. These records may be duplicates. Another customer reporting an undelivered pump from the same distribution centre could belong to a shared delivery investigation, but their order, correspondence and remedy remain separate.
Document boundary decisions with the support lead. Does a fresh failure after a completed repair reopen the earlier case or start a new one? Can one request contain several affected assets? The matcher needs these operating rules before a similarity score can become a useful recommendation.
Build a comparison record without erasing differences
Extract a small comparison record from each ticket: customer account, order or asset reference, issue type, reported event time, contact channel and unresolved request. Keep the original message accessible. Mark an absent field as unknown rather than inferring a convenient value from another candidate.
Remove repeated email signatures and quoted boilerplate from the text used for comparison, but preserve negation and corrections. In a hypothetical pair, "the replacement has arrived but is damaged" is not equivalent to "the replacement has not arrived". Both concern the same order, yet the requested response has changed.
Treat identifiers according to their actual system rules. Normalising whitespace may be harmless; dropping leading zeroes might change a reference. Ask the source-system owner which transformations preserve identity. Record the original value alongside the normalised value so an agent can inspect a mismatch.
For attachments, store whether useful evidence exists and what authorised extraction found. A failed attachment read is not proof that two tickets contain identical evidence. Flag that gap for review.
Use similarity to retrieve candidates, not authorise merges
Semantic matching is one possible candidate-finding method. The Sentence Transformers documentation describes comparing text embeddings with similarity measures. That comparison concerns textual meaning; the business decision about whether two support obligations are identical needs additional checks.
Combine candidate retrieval with exact-reference searches. A shared order number can uncover differently worded contacts, while text similarity can suggest candidates where a reference was omitted. Keep the reasons separate: "same order reference" is a different signal from "similar description".
Apply access boundaries before exposing candidate details. A support agent should not learn another customer's private correspondence through a duplicate suggestion. For cross-account incidents, show only the incident information that the agent is authorised to view, not unrestricted ticket snippets.
Do not label an arbitrary similarity score as a probability that merging is correct. Establish review thresholds using labelled examples from the intended queue. Even a strong match should remain a suggestion when identity, event timing or outstanding obligations conflict.
Make the review screen explain the disagreement
A reviewer should see both records side by side, with matching fields, conflicting fields and missing evidence distinguished. Put the difference that could change the decision near the proposed action. Requiring reviewers to expand several panels to find a different serial number invites superficial approval.
Useful review details include:
- The exact messages supporting a proposed match.
- Account, order and asset references, including unknown values.
- Event dates and the most recent unresolved request.
- Current assignees, response deadlines and promised follow-ups.
- Attachments or private notes that will not transfer automatically.
Offer "link as related" and "keep separate" as first-class actions. A reviewer should not have to merge tickets merely to tell colleagues that they concern the same outage. Allow a short rejection reason, such as different asset or repeat failure, which can later inform evaluation.
Keep the queue manageable by showing the strongest few candidates and allowing further search. A long list of vaguely similar tickets transfers the search burden back to the agent.
Treat merging as a controlled record change
Before enabling merges, test the helpdesk's behaviour in a non-production environment. Establish which record survives, what happens to attachments, whether customer recipients change and how response timers are handled. Do not assume that a merged ticket preserves every operational detail.
Require the reviewer to choose the surviving record and confirm its owner. Carry forward every unresolved commitment explicitly. In the hypothetical supplier example, a promise to phone the customer before dispatch must not disappear because it appeared only in the email ticket.
Check for changes again immediately before applying the merge. Another agent may have replied, changed the assignee or added evidence since the suggestion was created. If material fields changed, return the pair for review instead of executing a stale decision.
Store the source record identifiers, reviewer, decision reason and merge result. If the platform cannot reverse merges cleanly, prefer linking until the business accepts that limitation. Keep any recovery instructions practical: identify who reconstructs separated records and how outstanding work is restored.
Test false matches and missed matches separately
Create a labelled test set containing genuine repeated contacts and deliberately difficult non-duplicates. Include the same customer with two orders, two customers using identical wording, reopened incidents, missing identifiers and mixed-language messages. Where examples are synthetic, label them as hypothetical test cases.
Measure how many suggested duplicates reviewers confirm, and separately check how many known duplicates the system fails to surface. Review both errors by issue type. A queue dominated by password enquiries may behave differently from a queue handling equipment repairs; one blended score can obscure that difference.
Add a cluster test. If ticket A resembles B and B resembles C, do not assume A and C should merge. Require the same identity and obligation checks for each proposed member of a group.
Pilot with suggestions and visible ownership
Start in one queue with suggestions only. Keep normal assignment and response duties active while an item awaits duplicate review. Track review time, rejected suggestions, reopened work and any lost commitments, alongside the number of confirmed duplicates.
Give the support lead a straightforward way to pause suggestions when a recurring false-match pattern appears. Correct the extraction or matching rule, then replay the relevant cases before resuming. A focused support automation workflow should help agents find the right relationship between records while leaving every customer's unresolved request visible and owned.