what is copilot mode in customer support?

When a customer message arrives, the AI agent reads the full conversation history — including orders, stock or any connected data — and drafts a proposed reply. That suggestion appears in the human agent's inbox, not in the customer's chat. The person can:

• Approve it as-is and send with one click.
• Edit it — adjust tone, add a detail, fix something — before sending.
• Discard it and write their own reply from scratch.

The customer never knows there was a draft in the middle: they only see the final, reviewed reply. This sharply cuts the time a human agent spends composing each message — especially repetitive ones — without taking away control of what actually gets sent.

Because the suggestion is generated from the same knowledge base, catalog and order data the agent uses on autopilot, a copilot draft is rarely a blank slate. Even when a person rewrites it, the useful facts — the order status, the return window, the item that's actually in stock — are already surfaced, so the human is editing wording rather than researching the answer. That's the practical difference between copilot and a generic canned-response macro: the draft is specific to this customer and this conversation, not a template the agent has to adapt by hand.

Autopilot is the mode where the AI agent replies to the customer directly, without human review, and only escalates when it should (by rules, keywords or uncertainty). It's the mode that scales furthest: it covers 24/7 without depending on someone being online.

Copilot is the better fit when:

• The business is still tuning the agent's tone and answers, and wants to supervise before letting it run fully.
• There are sensitive conversation categories (complaints, legal cases, VIP customers) where human judgment matters more than speed.
• The team wants to cut drafting time without fully delegating the decision of what's said on the brand's behalf.

Autopilot is the better fit once the volume of repetitive conversations — stock checks, hours, order tracking — outstrips what the team can handle by hand, and those categories are already well trained and tested.

Most teams don't jump from no AI to full autopilot overnight, and they don't have to. A common path is: start in copilot to watch how the agent responds with real data and real customers, correct the tone or the knowledge base where needed, and once the suggestions are consistently good, move to autopilot — sometimes only for certain query types first, leaving the more sensitive ones in copilot longer.

That change isn't all-or-nothing or irreversible: you can go back to copilot at any time, or pause the bot on the spot — per conversation, channel or globally — if the team needs to take control.

In heysteff, copilot mode isn't a separate feature but a state of the same agent: Steff keeps reading every conversation and generating suggestions, and anyone on the team can take over in one click — per conversation, channel, workspace or globally. The system automatically detects when a human has already replied ("echo" detection) and pauses Steff so it doesn't step on that reply. Escalation flows by rules and keywords work the same in both modes, so switching between copilot and autopilot doesn't mean reconfiguring anything from scratch.

• What Is an AI Agent for Ecommerce?
• What Is an AI Employee?
• Copilot mode in heysteff
• Conversational AI glossary
• AI customer support