Agent Copilot: AI support right at the agent's desk

Agent Copilot does not talk to customers. It works alongside the human agent in the Agent Workspace: it supplies context, suggests replies, offers the next step. The agent stays at the wheel; every reply needs his approval. That sets Copilot apart from the AI Agents, which resolve tickets with no human involved. Two different tools, often confused.

What Agent Copilot is

Not a single feature, but a bundle of AI functions in the Agent Workspace, booked as the Copilot add-on. The add-on costs around 50 USD per agent per month on annual billing, and requires at least the Suite or Support Professional tier; the Team plans are left out. It is the successor to the earlier «Advanced AI» add-on. The main parts:

• Suggested First Replies: one suggestion for the first response, generated from existing Macros and Help Center articles. Not new knowledge, but what you already have, faster to hand.
• Auto Assist, the core: an LLM reads the ticket and suggests how to resolve it, as reply text or as an Action. More on that shortly.
• Ticket Summaries, Similar Tickets, Suggested Macros, Enhance Writing, Intelligent Triage, the supporting cast: a summary of the thread, similar resolved cases, matching Macros with a reliability badge, a rewrite of the draft, automatic sorting by topic, sentiment and language.

The common thread: Copilot invents little of its own. It draws on what is already there, on Macros, the Help Center and resolved tickets, and makes it visible at the right moment.

How it helps the agent

Auto Assist runs on procedures: instructions that admins write in plain language in the Knowledge admin, «for this request, proceed like this». A procedure also includes «Instructions for agents», a note to the human on the ticket that does not go to the customer. A trigger marks the matching tickets with the tag agent_copilot_enabled; Auto Assist only activates there. Set the tag only once the matching procedure is in place, otherwise Copilot comes up empty. The suggestion appears in the composer: the agent sees what Copilot would do, and decides.

Two things matter here. First, approval: replies and writing Actions only go out once the agent confirms them with a click. Since February 2026 there is one exception, the pre-approved Actions: purely reading Actions, such as looking up order or account data, run without a click because they change nothing. Anything that writes or sends stays with the human. Second, the Actions themselves: a procedure builds in existing Macros and calls Standard Actions (shipped with the product) or your own Custom Actions, which reach into other systems via API. Custom Actions come with hard limits: at most 100 per account, a 10-second timeout per call, 2 MB response size, at most four suggestions at once. That is how Copilot suggests concrete steps in the system that the human triggers or discards.

Auto Assist only kicks in for tickets from email and messaging, not in live chat or voice, and only where a human is sitting on the ticket. For channels without agents, the autonomous AI Agents are in charge.

That is where the value sits: repetitive tickets get handled faster and more consistently, without removing the human decision.

Putting it in context

Agent Copilot is only as good as the material it draws from: a well-kept Help Center, clean Macros, clearly written procedures. The AI changes nothing about a thin knowledge base; it only makes it visible.

And the line to the AI Agents stays the compass: Copilot = human in the loop, every reply approved. AI Agent = autonomous, with no agent in between. Both draw on the same knowledge and process world, but the one who decides is a different one. Knowing that line tells you which tool to set up for which case.