# Support & help desk: the help desk that reads the ticket before you do BlueHill's help desk is built around the **BlueHill Assistant**, an agent that reads every new ticket and hands the team a plan, and **Autopilot**, a rules layer that lets the assistant send routine replies itself inside guardrails the workspace controls. ## How the assistant works a ticket One analysis runs when a ticket arrives and again when the customer replies. 1. **Read.** The assistant reads the conversation, the customer record, the account's recent tickets, the SLA clock and the public help-center articles. 2. **Plan.** It ends with a single plan: a summary, triage proposals (priority, assignee, tags, category, ticket fields, related or duplicate tickets, a follow-up) each with a confidence level and a "Why", and a reply draft that cites the articles it relied on. The draft is signed as the brand, never as a person. 3. **Decide.** Autopilot rules are checked top-down and the first match decides what happens to the draft: *Shadow* (record what would have been sent), *Hold, then send* (schedule it after a delay anyone can cancel), *Send right away*, or *Never autopilot*. With no matching rule the draft stays a proposal for a person. 4. **Record.** Automatic replies carry an Autopilot badge on the ticket timeline; every event they cause reads "by BlueHill Assistant automatically" with a Why link to the rule and the reasoning. A person who presses *Send now* is recorded as the approver. ## What the agent sees on a ticket - The **Assistant card** in the customer rail: summary, TRIAGE rows with Apply / Dismiss / Apply all, and the REPLY DRAFT box with source chips, a split **Send / Send & set status** button, Edit in composer and Dismiss. - The **Needs review** inbox view lists every ticket with open proposals; the row shows a proposal count. - The **Autopilot queue** view lists replies on hold with an amber countdown on each row. The scheduled reply card shows the rule, the draft, what happens after sending, and **Cancel send** / **Send now**. - In the composer: **Draft reply**, **Rephrase** (Polish, Fix spelling & grammar, More professional) and **Send reply** with an optional status change. ## Rules and preview Rules live in Support settings → AI assistant. Each rule has a name, a mode, an optional canary share of eligible tickets, a hold time, a daily cap, and conditions in four families: the trigger (new tickets, customer replies, first message only), the ticket (brand, channel, category, priority, status, tags, custom fields, assignment), the requester and customer (sender domains, sender kind, segments, customer ids), and timing, message and model (business hours, SLA risk, must-mention and never-mention phrases, maximum length, minimum confidence, minimum cited articles, skip upset customers). After-send actions set the status, assignee, tags, an internal note, a follow-up reminder, notifications and the satisfaction survey. **Preview on recent tickets** replays a rule over stored analyses and shows which tickets would have matched, which the safety floors would have blocked, and which would not match — without a model call. Every change to the rule set is versioned and **Revert last change** restores the previous version. ## Guardrails no rule can lower - **Safety floors:** minimum cited articles, minimum confidence, one automatic reply per ticket, a maximum reply length, blocked phrases, and hard skips for teammates, shared and no-reply senders, suppressed addresses, threads with other people on Cc, messages with attachments, and customers who sound upset. - **Hold window:** a workspace and per-brand minimum hold; longer on brands that send an auto-acknowledgement. - **Re-check before sending:** a scheduled send is dropped if the ticket was merged, marked spam, resolved, assigned when the rule requires unassigned, or touched by a person since the analysis. - **Human takeover:** an agent's reply or internal note cancels any send on hold for that ticket. - **Budgets and breaker:** daily caps per workspace, brand, rule and customer; a circuit breaker switches autopilot off after a run of failed sends and rings the admin bell. - **Audit:** an activity card with 30 days of sent, would-have-sent, on-hold, cancelled and failed counts, per rule, plus how often customers wrote back. Autopilot is enabled per workspace by BlueHill, switched on per brand by the workspace, and starts with no rules. ## The help desk underneath Email-to-ticket across multiple brands with their own domains, senders and signatures; SLAs with business hours; canned responses, internal notes, watchers, merges and bulk actions; a public help center that doubles as the assistant's source of truth; CSAT surveys; time tracking; and support analytics by agent and brand — all next to the customer record, their onboarding board, deals and portal. ## FAQ **Does it send anything on its own?** Only when an autopilot rule says so. With no matching rule, the draft stays a proposal. **Can we try a rule without sending?** Yes — Preview replays it over recent analyses, and Shadow mode runs it live without sending. **How do we know what was automatic?** Autopilot badge on the timeline, "by BlueHill Assistant automatically" on every event, a Why link to the rule, and the 30-day activity card. **What if it starts failing?** Daily budgets cap sends; the circuit breaker turns autopilot off after repeated failures until a person re-enables it. --- Start a free trial at https://app.usebluehill.com/sign-up or book a demo at https://cal.com/himanshu-minocha/bluehill-intro.