An appointment-setting agent, built in a day

In early October, I built an agent that books meetings from prospect replies. It took one day. The next day, it was working live at Oliverlist. Today, it books 60% of the meetings that come from a reply. Here is how one day was enough, what that number says, what it does not say, and the safeguards around the agent.

What the agent does

At Oliverlist, Sorter classifies every prospect reply every 5 minutes: urgent, positive, negative, unsubscribe, and so on. Before the agent, replies that needed a follow-up were handled by hand by the team, between calls.

Now the agent runs every 15 minutes. It picks up those replies and answers with the text drafted by Echo, which offers available slots. As soon as the prospect accepts a specific time, it books the meeting in the calendar. Phone calls stay with people.

The agent also handles the less simple cases: a request for a call, a different date, a question asked before committing. The goal does not change: get the meeting as soon as possible, without pushing it back.

Why every 15 minutes? Because a prospect who has just replied is available now. The longer the answer takes, the lower the odds of a meeting.

Why one day was enough

The build did not start from scratch. Every building block already existed: reply sorting, drafting by Echo, calendar access. Behind them are years of automation, since 2020, and AI in the product, since 2023. The agent assembles those pieces. It does not invent anything.

What took time that day was not the code. It was setting the rules: when to reply, when to book, how to handle a prospect who suggests another date or asks for a call back. Those rules are written in plain sentences, not in code. That is what makes them easy to review and quick to correct. If you sell in several countries, this is also where working days and public holidays go.

The lesson applies to any CEO. If your data is connected and your tasks are already broken down, an agent takes days to build. If not, that is where to start, and it will take longer than the agent itself.

What the numbers say, and what they do not

The agent now books 60% of the meetings that come from a reply. That number tells you who books the meeting. It does not tell you how many extra meetings we get.

The net effect is still being measured, and I will not quote a gain I have not measured. It is hard to isolate. Some of the meetings booked by the agent would have been booked by the team anyway. The time released goes into phone calls. The volume of replies varies with campaigns and seasons. It will take several more weeks to settle the question.

The measure I follow is simple: the total number of meetings that come from a reply, relative to the volume of replies received, before and after the agent. Alongside it, I look at what the team gets on the phone with the time it has regained. If both go up, the net effect is positive. If not, the agent has only moved the work around.

The upstream step, on the other hand, is measured, and it matters: an appointment-setting agent is only as good as the sorting of replies. Measured in October, the sorting is right 99% of the time on urgent replies and 77% of the time on positive ones. Since its instructions were reworked, as I mentioned in early October, it detects 19% more hot leads, and false alerts are down 77%. An audit of 7,612 replies found about 4 genuine errors a day.

The safeguards

An agent writing to prospects on behalf of our clients needs a frame. None of the safeguards below is sophisticated. Together, they let the agent work on its own without anyone losing sight of it.

• A meeting is booked only on a specific time accepted by the prospect.
• When I add a sensitive type of reply, the agent starts in observation mode. It decides, sends nothing, and I review its decisions before letting it act live.
• Every decision is logged, and an audit reviews the agent's work every evening. Errors spotted in the evening are fixed the next day.
• Verifier and Judge, two other agents in the product, weed out fake meetings and settle claims.
• When a reply is wrong, I fix Echo's instructions, not the message. Fixing a message by hand solves one case. Fixing the instructions solves every case that follows.

Key takeaways

• An agent is quick to build when the building blocks exist: connected data, reply sorting, drafting, calendars.
• Separate the agent's share from its net effect. The first can be counted within days; the second takes weeks to measure.
• Start in observation mode, then go live one category at a time.
• Measure the upstream step: an appointment-setting agent is only as good as the sorting of replies.
• Fix the instructions, not the messages. That is how the agent improves.