Arman Afzar

AI Copilots & Agents

A support copilot pilot: a realistic plan

Four stages that take a support copilot from demo to an everyday team tool.

The Arman Afzar team2 min read

A support copilot only matters if the team actually uses it. Most pilots stall at the demo stage because they weren’t designed for daily work. This plan comes from building Q&A assistants for trading and enterprise platforms.

1. Pick a queue

Choose one category of requests whose answers already exist in documents — for example identity-verification questions or order tracking.

2. Drafts, not auto-replies

In the first weeks the copilot only drafts replies; an agent approves or edits them. Every edit is valuable data for improving the knowledge spine.

3. Measure

Compare the share of drafts approved unchanged, and response time, against the baseline. If the share is low, the problem is usually the sources, not the model.

4. Expand

Once results are acceptable, enable auto-replies for simple questions, keep the human hand-off path, and add the next queue.

Related articles

  1. 01AI copilots fail without a knowledge spineMost enterprise copilots don’t fail because of the model — they fail because knowledge lives in ten different places. What a knowledge spine is, and how to build one.
  2. 02Automation without a baseline: why AI projects can’t show resultsIf you don’t know how long a process takes today, you’ll never be able to say how much automation helped.
  3. 03An LLM on your own servers: where to startWhy organisations run language models on-premise, and what a good pilot looks like.

Tell us your project in one sentence — we’ll find the right facet together.

Start the conversation