Fellix Co-pilot: how AI-assisted agents close conversations faster

Most conversation about AI support is about replacing the reply. Co-pilot is the opposite idea: it assumes a person is answering, and removes everything around the answer that is not the answer.

That matters because of what reaches a person once an agent is handling the repetitive half. What is left is, by definition, the conversations that needed judgement — the exceptions, the complaints, the ones where a sale is in play. Those are the conversations where the time goes, and where speed is worth the most.

Where the time actually goes on a hard conversation

Watch someone pick up an escalated thread and the typing is the small part. The work is:

• Reading back. A long thread, sometimes across channels, to find out what happened.
• Working out what they want. Often not what they said first.
• Finding the relevant fact. The policy, the price, the previous order.
• Composing something careful, because these are the conversations where wording matters.

The first three produce nothing the customer sees. They are pure overhead, and they are the same overhead on every escalation.

What Co-pilot does about each

The summary

A conversation arrives with what happened already written down: what the customer asked, what has been said, where it got stuck. The person starts at the point of deciding rather than the point of reading.

This is the largest single saving, and it grows with thread length — which correlates with how difficult the conversation is.

Detected intent

What the customer actually wants, classified, rather than inferred from the opening line. A message that starts “I ordered on Tuesday” might be about delivery, returns, or a payment problem, and the difference decides who should handle it.

Intent is also what makes routing work: it is the thing being routed on.

Suggested replies

A draft, from your own information, that a person edits and sends. The point is not that the draft is always right — it is that editing is faster than composing, and a draft grounded in your policy stops the most common error on a rushed reply, which is stating something almost true.

The person stays responsible for what goes out. That is the whole design: the agent handles conversations, Co-pilot handles the agent’s paperwork.

Why “assist” is the safer shape

An autonomous agent answering a complaint is a risk. The same model summarising the complaint for a person is not — the worst case is a summary someone has to correct, and they are reading the thread anyway.

That asymmetry is why assistance is worth deploying earlier and more broadly than automation. The failure mode is recoverable, and it is recovered by the person who was already there.

It is also why Co-pilot is the right first step for teams nervous about AI touching customers at all. Nothing it produces reaches a customer without a person pressing send.

What to measure

Handle time is the obvious one, but on its own it is misleading — a team can get faster by being worse. Watch it alongside:

• Reopen rate. Conversations that come back are conversations that were closed rather than resolved.
• Escalation depth. How often a conversation changes hands. Good summaries reduce this, because the first person can act.
• Share of suggestions edited before sending. If it is near zero, nobody is reading them, and that is a risk rather than a success.

That last one is the honest health check on any assistive feature.

Where it fits with the rest

Co-pilot is the second half of a pair. The AI agent handles the conversations that do not need a person; Co-pilot makes the ones that do cheaper to handle well.

Deploying only the first gives you a fast front door and an unchanged bottleneck behind it. Deploying only the second makes a team more efficient at a queue that is still full of questions nobody needed to read.

---

Co-pilot works on the conversations that reached your team, inside the same inbox. See how Co-pilot works.