AI support analytics

Resolve the company in a cascade: the ticket's own association, then the contact's, then email, then normalized phone (country and area code plus the last 8 digits). An email domain only counts if it is not a free email provider. Anonymous widget visitors, who can be half of the tickets, stay out of the per-customer analysis but count in volume. Publish the link rate.

Remove the team's internal tickets, test companies and messages from the team's own numbers. Flag the tool's AI bot replies and automations. Agent productivity counts only messages from an identified human operator; system events and internal notes are not conversation.

In chat, the customer writes before the ticket is created: in a real base, the first message was at 07:43 and the ticket at 07:59. Measure waiting and resolution from the customer's first message in the transcript. A technical close of the conversation is not resolution. Timings are summarized by median and percentiles, never by sum or average of averages.

Rebuild the full thread (customer, agent and bot) before classifying. In email, cut the quoted history from each reply ("On ... wrote:", "Em ... escreveu:", "De: ... Enviada em:", lines starting with ">"). Call transcripts have very short utterances: join them into 60-second windows with a timestamp. Never use the call tool's ready-made summary as a source: it is a third-party interpretation that loses context.

The AI picks one primary reason from a closed, versioned list, and separately lists every reason that appears with evidence. For a cancellation or downgrade request, classify its nature: leaving or rightsizing. Without enough evidence, the safe default for retention is to treat it as leaving. When the list changes, store the version on each row and map the old reason to the new one.

Each reason, sentiment and risk signal comes with the literal quote from the new text. Normalize the keys of the AI's response (accents and lowercase) before consolidating. Track the class distribution every run: a model that returns everything as neutral is almost always a call error, not a happy customer.

Classify only what is new: a ticket enters the run if it was never classified or if it received a message newer than the last one read. The watermark is per ticket; a single watermark for the table gives each ticket one chance and loses messages that arrive later. Use a fixed window and a cap on calls per run, and limit how many types go to the AI in frequency analyses.

Rank topics by affected customers and their revenue, not by raw message count. Recurrence is the same customer opening the same reason (or the same tag combination) more than once in 30 days: it signals a problem not solved at the root, not a customer who uses support a lot. Counting distinct weeks with a ticket separates recurrence from an isolated spike.

In real bases, the category was empty in about 60% of tickets and almost half had no owner. Use queue or pipeline as the main split, treat a null owner as unassigned and never discard it: a queue without an owner is itself a support health signal. Priority, origin and stage are dimensions, not quality measures.

Before sending conversations to an external model or a voice-of-the-customer tool, replace email with the user id, phone with a placeholder, and mask CPF and CNPJ (individual and company tax IDs). In a source shared by several clients (for example, one account that gathers reviews for several brands), the per-client filter is mandatory: a forgotten filter does not throw an error, it leaks silently.