AI Assist — context and reply assist

AI Assist is the human-in-control AI layer. It never sends anything on its own. Its job is to make a human agent faster by reusing the same Ticket Context that triage, Suggest, auto-resolve, handoff, and KB gap detection use.

What Assist shows you

Open any ticket. The right rail shows Ticket Context with:

• AI summary — customer intent, current state, last ask, key facts, and what not to repeat
• Grounded context — relevant KB evidence, supported claims, unsupported risks, and open questions
• Suggested next action — answer from KB, ask a clarifying question, escalate, wait for customer, or resolve
• Customer memory — durable facts, preferences, and unresolved cross-ticket issues

Reply drafts stay in the composer. Use Suggest when you want AI to draft the actual response from the current Ticket Context and latest conversation state.

Credit cost

Ticket Context uses the summary credit action when it refreshes after inbound customer messages. Refreshes are debounced per ticket so rapid message bursts do not burn duplicate credits.

Opening and reopening the same ticket reads the persisted context. Manual refresh is available from the panel when an agent wants to force a new context pass.

Configuring Assist

Settings → AI → Context:

• Shared context. Context is generated once and reused by the panel, Suggest, auto-resolve, handoff, analytics, and KB gap review.
• Grounding contract. Evidence includes source chunks, supported claims, unsupported risks, and a safe grounding state.
• Suggest behavior. Drafts are generated from the composer and use fresh Ticket Context before falling back to live retrieval.
• KB gaps. Weak or missing matches can be reviewed in Knowledge coverage.

Reading the draft quality

Ticket Context shows a grounding state before any draft is generated:

• Grounded: evidence supports an answer. Suggest and auto-resolve can use the supported claims.
• Weak match: the KB may be relevant but not enough to answer safely. Ask a clarifying question or verify the missing fact.
• No match: there is no safe KB answer. Find or create KB coverage, escalate, or ask for more information.

Assist + canned responses

Suggest won't use canned responses by default. If you want it to reference your canned library (e.g., "use the welcome canned response for new signups"), add canned-response names to your brand-voice instruction:

Prefer the welcome canned response for new signups. Prefer the refund-approved canned response when a refund has been approved by finance.

Simple and effective. No special setup.

What Assist is not

• Not auto-send. Suggest never sends a reply. If you want AI to close tickets on its own, see auto-resolve.
• Not a co-pilot for everything. Ticket Context is scoped to support replies, KB grounding, handoff, and gap detection. It doesn't edit customer records.
• Not a draft preview. The sidebar shows context and next action. Draft text belongs in the composer.

Best practices

• Trust the grounding state. Spot-check weak or no-match cases before answering with facts, especially numbers, dates, and prices.
• Edit the tone to match the customer. If the customer is upset, soften the opener. If they're technical, drop the "hope this helps!"
• Give AI more context via internal notes. Write a quick internal note ("this customer had the same issue in ticket #1234, resolved by X"). Context refresh reads internal notes and incorporates them into future drafts.

Summary

• Ticket Context refreshes after inbound customer messages and is reused by AI features
• Grounding includes evidence, supported claims, unsupported risks, and open questions
• Drafts stay in the composer through Suggest
• Auto-resolve and handoff read the same context before falling back to raw history