Complaint classification and deadline routing

A classification is useful only when a handler can see why it exists. The record should quote the inbound complaint, attach a type and urgency only when those quotes support them, and point to the deadline rule that starts the response clock. If the model cannot name a type from the text, type, urgency, and deadline stay blank. Compliance still writes the response. Nothing auto-closes. No one invents a statutory number of days.

That bar is the quality outcome for this workflow. A tag in a workbasket without those cites is a queue move, not a classification.

Inbound complaints arrive as email, portal forms, regulator forwards, and call notes. Queue names and source systems are weak signals. The same letter can be a claims dispute, a billing error, a coverage misunderstanding, or a conduct issue. Treating the channel as the type is how urgent regulator correspondence sits behind a billing queue.

Core systems already hold the case: policy admin platforms such as Guidewire and Duck Creek, and service desks such as Salesforce and ServiceNow. Classification sits on top of those records. It does not replace the complaint file, the handler assignment, or the response letter.

When the complaint is really a coverage question, handlers still need the policy, not a faster label. Pair this workflow with the policy RAG coverage assistant so the cited route and the cited policy language stay in one thread.

Start by loading the full inbound text into the case, including attachments that are already in the file. Do not classify from a subject line or a one-line summary. If the complaint arrived as a call, load the transcript or the handler note as written, not a cleaned paraphrase.

Run classification against that loaded text. For each proposed type and urgency, require a cite: a short quote from the complaint that a reviewer can find in the source. For the deadline, require a cite to the deadline rule your compliance library already maintains, such as a regulator acknowledgment window, an internal SLA by complaint class, or a mapped rule legal has approved. The model maps the cited text to that library. It does not write new law.

Then route. Assignment goes to the handler group that owns that type, with the clock visible on the case. Routing is a recommendation plus evidence. A supervisor can reject it. A rejected route returns the case to unclassified, not to a default "other" bucket that looks complete.

A typical letter shows the shape. A policyholder writes that a claim payment was delayed, that they have already chased twice, and that they will take the matter to the ombudsman if they do not hear back. The model quotes the delayed payment and the ombudsman threat for type and urgency, then cites the firm's mapped rule for claim-service complaints that sets the acknowledgment and substantive-response clocks. The handler still drafts the reply. If the same letter only said the writer was unhappy with the policy and named no event, type stays empty, no deadline starts from a guessed class, and a human reads it before anything is labeled.

Voice-originated complaints need the same load-then-cite path. Notes from an ai voice agent for policy servicing are source text, not a finished classification. Classify the note the way you classify an email.

Mid-term change complaints often mix servicing and product. Before you treat a "you changed my cover" message as a conduct issue, check whether an endorsement actually posted. The mid-term endorsement processing agent is the companion check so you do not route a processing delay as a regulatory complaint, or the reverse.

Empty is a valid state. Ambiguous letters are common: a billing dispute that mentions "unfair," a coverage question that mentions "complaint," a cancellation that might be a cooling-off issue or a simple request. If the quotes do not pick a class, do not pick one for the model.

Blank fields keep the case in a human triage queue. That is slower than a wrong label and safer. A wrong label starts the wrong clock, notifies the wrong team, and can look like the firm already decided what the complaint is.

Do not fill blanks with a catch-all type to keep dashboards tidy. "Other" with a clock is still a classification. If you need a holding queue, name it unclassified and keep deadline fields null until a person sets them from the rule library.

Three failure modes show up as soon as the workflow looks efficient.

The first is a route with no deadline-rule cite. The case has a type, a handler, and a due date, but the due date is a model guess or a copied number from a similar case. That date is not a control. Strip the due date until a reviewer attaches the rule your library actually contains. If no rule applies yet because type is unset, there is no due date.

The second is treating the route as closed. Classification and assignment are not a response. Auto-closing on "routed," on a template acknowledgment, or on a confidence threshold hides open regulatory time. The complaint stays open until a human records the response, and any required follow-up, in the case. Acknowledgments can be drafted. They are not sent as a close event.

The third is inventing a statutory deadline. Models will offer round figures that appear in other jurisdictions and in generic training data. Unless the cited rule in your library states that period for that complaint class and that product, the field stays empty or holds only the internal SLA your team has documented. Guessed statutes create false comfort and false breaches.

Bind-time checks do not set service-complaint clocks. A product that passed a regulatory compliance check at bind can still generate a valid post-sale complaint. Do not copy bind findings into the deadline field.

Handlers and compliance write what goes back to the customer or the regulator. The system stores the inbound text, the cites, the chosen rule, the assignee, and the clock. That split is the operating model: automation proposes and tracks, people decide and send.

Reviewers should be able to open a case and answer four questions without hunting: what the complainant said, what class was applied and why, which deadline rule is running, and who still owes the reply. If any answer is missing, the case is not ready for a dashboard green status.

Tune the classifier against those four questions, not against volume classified. A higher auto-route rate with more empty or wrong clocks is a regression. Watch whether cited routes survive human review and whether blank cases are actually read, not whether the workbasket looks empty.

Vendors in this stack remain systems of record. Guidewire, Duck Creek, Salesforce, and ServiceNow keep the complaint, the customer, and the task. The classifier writes structured fields and evidence onto those records. If a field cannot be cited, it is not written.