How to automate request triage without losing context

Automatic triage organises requests so they reach the right person with the necessary context. It can combine rules based on known fields with AI to interpret free text. The useful result is a verifiable assignment with a route for ambiguous cases, rather than merely a label produced quickly.

By Lyntev. Updated on .
About 4 minutes to read.

A request that stays connectedThe same reference follows the request and prevents duplicates.

Translation published on . Original Portuguese guide; source revision .

By the end, you can:

  • Design categories without unnecessary overlap
  • Measure correct routing by category
  • Handle ambiguity and failures without losing requests

Describe what arrives and who should receive it

Inventory the channels used: form, email, inbox or application. Initially choose one and record available fields, attachments, languages and the current owner. The same sentence may need different destinations depending on the product or contract; identify the context that actually changes assignment.

Separate classification from priority. “Invoicing” identifies a topic; “reply by tomorrow” expresses a need; an operational priority depends on company rules. Do not automatically turn words such as “urgent” into permission to bypass controls. Preserve the original message even when showing the team a summary.

Write a definition and boundary for each category

Give each category a description, included examples, excluded examples and a destination. If two team members classify the same requests differently, resolve the rule first. Automating an ambiguous taxonomy may merely speed up cases being sent to the wrong place.

Anthropic’s routing documentation proposes studying the support process and testing with classified requests. The matrix below is an editorial proposal for a fictional service business. Adapt boundaries to your work and retain a review category; it has an owner and deadline, rather than being a place where requests disappear.

Write a definition and boundary for each category
CategoryIncludesRouting
SalesRequest for conditions for a new serviceSales team
SupportProblem with a service already contractedSupport team
AdministrationDocument request or contact-detail changeAdministrative team
ReviewTwo topics or insufficient informationIntake owner

Use rules where the decision is already defined

A form explicitly identifying the service and reason can follow a direct rule. AI is an option when information arrives as variable text needing interpretation. Do not add a model call just to copy a field already determining the destination.

Request output limited to the expected categories, with missing data and brief textual evidence. Structured outputs help meet the format, as Anthropic documents; they do not prove the chosen category is correct. The application checks structure, permitted destination and necessary conditions before assigning the request.

Examine a request needing review

In fictional example PED-309, the person writes: “I cannot open the customer area and would like to know the price of two more licences.” There are two legitimate topics. The proposal is to route the access problem to support and flag the sales interest, keeping the work linked to avoid two contradictory replies.

If the system permits only one destination, the policy may place the request in review for someone to split it. That is preferable to pretending the dominant category resolved the entire message. Record the policy in the example set so evaluation does not penalise human routing that was precisely the expected behaviour.

Illustrative result for review, without automatic executionJSON
{
  "pedido_id": "PED-309",
  "categoria": "revisao",
  "assuntos": ["acesso", "novas_licencas"],
  "motivo": "The message contains a support issue and sales interest.",
  "acao_proposta": "separar_assuntos_com_referencia_comum"
}

Measure errors that change the destination

Prepare a sample annotated by people who know the operation and reserve part for final evaluation. Include rare categories, short messages, long texts and cases that should go to review. Do not tune the system using every example and then present results on those same examples as independent evidence.

In a fictional test, 50 requests are routed to support and 45 actually belong there: precision for that destination is 45 ÷ 50 = 90%. If the sample had 60 support requests, recall for that category is 45 ÷ 60 = 75%. The first calculation measures the chosen destination; the second shows how many support cases were missed. These are not universal targets.

Measure errors that change the destination
ObservationQuestion helping correction
Sales sent to supportDoes the definition confuse intent and problem?
Support routed to reviewIs context missing or is the case genuinely ambiguous?
Many requests with no destinationIs a category missing or is there a technical failure?
Correct category, wrong teamIs the responsibility map current?

Preserve statuses and allow assignment correction

Keep statuses such as received, classified, assigned and under review. An interruption during classification must not delete the request or present it as handled. The operation reference enables resumption and recognition of repeats; another attempt does not accidentally create a second support case.

Begin in suggestion mode: the team sees the proposed category, confirms or corrects it, and the system records the decision. Observe triage time, rerouted requests and the pending queue. A reduction in classification time accompanied by more correction work does not demonstrate improvement of the complete process.

Document the rule before automating it

The operational owner should be able to explain the category and correct its destination without editing model code. Define how new categories are approved and which cases are retested. When the offer or team changes, review classification and assignment mapping together.

Worksheet for your project

Request channel and reference:
Available fields and necessary context:
Categories, definitions and exclusions:
Rule for messages with several topics:
Destinations and responsible people:
Conditions for human review:
Structured output and validations:
Development sample and reserved sample:
Precision and recall by category, with counts:
Triage time, corrections and pending cases:
Repeated events, unavailability and recovery:
Policy version and next review:

Sources and criteria

Content prepared with AI assistance and research using technical sources, with checks of the relevant examples and calculations. The references support the guide’s decisions; examples identified as fictitious let you try out the method.

Wellinton Oliveira

Written by

CEO of Lyntev

Technology with context, original analysis and sources you can consult.