AI for duplicate case detection: is it the same incident?

Two contacts about one incident

In this hypothetical case, a customer writes, “I couldn't pay for my order.” They later call, and the agent records, “Customer following up on a payment error with yesterday's purchase.” Both tickets contain the same order number, and the call concerns the same failed payment attempt. The agent answering the phone should be able to find the earlier email and see what has already been established.

Another ticket contains exactly the same sentence, “I couldn't pay for my order,” but describes a different purchase and another payment attempt. That is a separate incident. Even a shared order number does not settle every case: after resolving a payment issue, the customer might report a delivery problem with the same purchase.

A useful suggestion therefore shows both messages side by side, along with the order number and when the problem occurred. The agent can read the earlier findings and confirm whether this is a follow-up about the same incident. If a detail is missing, the proposed match still needs clarification.

What comparing text can tell you

Semantic similarity search helps find descriptions written in different words. The Sentence Transformers documentation describes finding pairs of texts with the same or similar meaning. That can help connect a customer's email with an agent's brief note. Similar meaning alone does not establish that both describe the same incident.

The result should make it easy to compare the text with the case details. A long list of messages about payment problems offers little help to an agent looking for an earlier contact about a particular purchase. It also matters whether the tool can find a pair where one message describes the error and the other merely refers to a previous conversation. Without that conversation or the recorded case details, the model may lack the information needed to suggest a match.

Start with the system you already use

Start by reviewing order numbers, existing links between contacts, and search in the current support system. If agents cannot see an email during a call because the channels store records separately, connecting those records may be the first job. Further model training will not fill in a missing history.

Once the records are available, compare the current search with an existing text similarity model. Such a model may already find descriptions that use different wording. Customization is worth assessing when particular mistakes keep recurring: perhaps the team's abbreviated notes are missed, or similar descriptions of separate incidents repeatedly appear as suggested matches.

Fine-tuning is one form of customization. It involves further training an existing model and, in this application, may help it compare the language used in tickets. The solution still needs case details that distinguish separate incidents. Experienced agents help explain why certain contacts belong to the same case while others should remain separate.

Assess suggestions before deployment

The comparison should include cases that were not used to customize the model. The team reviews similar-sounding tickets about separate incidents as well as genuine duplicates written in different words. Look at which suggestions agents reject and which connections the solution still misses. Otherwise, a short list can look convincing while overlooking most of the contacts that need to be handled together.

A service manager also needs to know how much reading the suggestions require. If an agent works through similar messages without reaching earlier findings any faster than in the current system, further investment needs justification. Comparison with the existing model will show whether customization changes the work enough to be worth maintaining.

Finding a related case and merging records are separate actions. Zendesk states that merged tickets cannot be unmerged. The solution proposed here leaves the decision with the agent and does not automatically merge or close tickets.

In quality analysis, many separate complaints may deliberately belong to one group. Our article on grouping quality complaints explains that use. Duplicate detection looks for contacts about the same incident.

What you can commission from Syntalith

We propose an AI application that suggests related cases as part of the support team's work. We start by identifying where work is being repeated and which records establish that contacts concern the same incident. We compare search in the current system with an existing model and assess customization where errors recur.

We connect the selected approach to a case view where agents can review the suggested pair, read earlier findings, and record their decision. Suggestions draw on cases the individual agent is authorized to access.

Your team explains difficult cases and corrects suggestions, helping establish whether the tool makes earlier work easier to find. For an initial conversation, describe an occasion when two agents handled the same incident and how they currently search for previous contacts. We can agree on how examples will be used during that conversation. See Syntalith pricing for information about the offer.