Classification
Give an agent fast, typed judgments about a conversation, with a built-in test panel
Classification is a built-in platform tool that lets an agent ask a small classifier model typed questions about the conversation and get back structured answers with confidence. It works in chat, voice (cascaded and realtime) and process runs.
A classifier is not a chat model. It never writes replies. It answers questions you define, such as "is the caller asking to stop being contacted?" or "which team should handle this?", and reports how sure it is. Your agent then decides what to do with the answer.
When to use it
| Scenario | Question type |
|---|---|
| Is this urgent, or is the caller upset? | Yes/No, Rate |
| Which department should take this? | Pick one |
| Is this lead qualified? | Yes/No, Pick one |
| Did the caller ask to opt out or stop calls? | Yes/No |
| Should the agent escalate to a person? | Yes/No, Rate |
Use a normal LLM reply for anything that needs generated text. Use Classification when you want a calibrated answer the agent can rely on.
Question types
| Type | Answer | Example |
|---|---|---|
| Yes/No | Probability that the statement is true | "The caller expresses urgency" |
| Pick one | The best option, a probability for each option, and confidence | billing, technical, sales |
| Rate | A score across ordered levels, with confidence | calm, frustrated, very angry |
All questions in one tool run together against the same conversation, so adding questions barely changes the response time. Keep each question about one thing. If you need a combined judgment, ask separate questions and let the agent combine them.
Create the tool
Open Build → Tools and create a Classification tool (under Built-in).
Choose a classifier model. Only classifier models are listed here, never chat models. Each model shows what it supports. Use Taliiq-funded usage, or connect Your key under Manage → Integrations → Classifier Providers.
Write When should the agent use this? This text is what the agent reads to decide when to call the tool.
Add questions, start from a template, or edit the JSON view. Question types the chosen model does not support are disabled.
Set the confidence floor. Answers below it are reported to the agent as uncertain.
Open the Test panel, paste a conversation or load a recent session, and press Run. Nothing is saved by testing.
Attach the tool on the agent Actions tab and add one line to the agent instructions, for example: "When the caller sounds upset, use the classification tool before deciding whether to offer a transfer."
What the agent receives
The agent gets one compact line, for example:
urgent=yes (0.97); department=technical (confidence 0.78, billing 0.15); frustration=1.0 of 2. Uncertain: none.Answers below the confidence floor are listed as uncertain. The tool description tells the agent to ask the caller instead of guessing, and never to read scores aloud.
Test panel
The Test panel runs your unsaved draft through the same service the agent uses, so what you see is what the agent gets. It shows probability bars for every option, the confidence of each answer, an uncertain badge, the input and output tokens used and the response time.
Runs on Your key are never limited. Runs on Taliiq-funded usage are limited per workspace per minute, and count toward the daily Taliiq-funded limit. Testing is not charged, because it is not part of a session. Any answer that changes when you reword a question is a sign the question needs to be more specific.
In sessions and logs
- In the conversation: each check appears as a Classification card in the Agent Console and in the session's conversation, with the answers, confidence, the credential used and what the agent received. Older sessions and other tools look exactly as before.
- Classification tab: in Observe → Logs, a session shows a Classification tab when the agent made checks. It lists every check in order and counts the uncertain ones. It works for a single agent and for Squad sessions, in voice and in chat.
- Cost: the Cost Breakdown adds a Classifier member only when the tool ran in that session. It shows the input and output tokens reported by the provider. With Taliiq-funded usage the cost is added to the session cost using the provider's published rate plus your plan markup, in the same single session charge. With Your key the tokens show at $0.00.
Good questions
- Write the exact condition. "The caller asks to be removed from the call list" works better than "opt out".
- Put boundary cases in the option or level descriptions.
- Keep arithmetic, dates and counting in your own systems. Classifiers judge meaning, not numbers.
- Send only the conversation the question needs. The agent sends the last 12 turns by default.
- Treat answers as advisory. A caller can try to steer any model, so keep confirmations on sensitive actions such as transfers or messages.
Limits and availability
- Up to 20 questions per tool.
- Maximum checks per session (default 20, up to 100) stops a looping model. At the limit the agent is told the check is unavailable and continues. It applies to every session type and both credential options.
- Text only. English is the tested language today; other languages are added to each model card once measured.
- A check takes about a second end to end. In a live call the agent waits for it, so use it where a judgment matters, not on every turn.
- In process runs there is no chat, so the tool checks the run input and session variables instead of a conversation.
- If the classifier is slow or unavailable, the agent is told the check was unavailable and carries on.
- Taliiq-funded usage has a daily limit per workspace. Connect Your key to remove it. Your key has no Taliiq limit.
- The tool is not available on workspaces with Zero Data Retention or HIPAA enabled, or on sovereign hosts, because conversation text is sent to the classifier provider.
See also Integrations and Squads.