How We Built and Use AI Agents

Why We Built It

One of our SDRs left sooner than I expected, and answering every reply was a job I didn't want. Our goal is to 10x the meetings we book for clients. An outsourced developer's version needed heavy edits, so on a Friday I built it myself. Version 1 was ready on the Monday.

The Agents We Run

• Our reply desk. Reads every reply, researches the person, writes a personal answer and books the meeting.
• Our reviewer. Checks every draft before it goes, sends it back for fixes and gives the final approval.

What Our Reply Desk Does in a Normal Week

We run hundreds of multi-channel campaigns, and every reply lands on the reply desk.

• Researches the person. Their profile, challenges and KPIs.
• Offers two real times. It reads our clients' live diaries.
• Books the meeting. It tells us and the client a gong (a booked meeting) is in, with the whole conversation.
• Follows up. It chases people who said yes and then got busy.
• Handles out-of-office replies. It writes to the colleague they point to.
• Keeps in touch. It checks in every 8 to 10 weeks.
• Replies fast. A reply that waits goes cold, so the first hour is the golden hour.

How We Built Them

I'm not technically minded. I started with an AI chat tool and ended up with an AI coding tool.

• Plain prompting. I describe what I want and which systems need to connect.
• 2 days of teaching. I approved every draft. A "teach me" button applied a lesson to one campaign or all of them.
• Our own material. It learned from thousands of messages our SDRs sent, and from my Gmail, so it writes like me.

The Guard Rails, and What Stays Human

• Automation in stages. A person approved every reply at first. At around 85% with no edit, we automated that campaign.
• A reviewer agent. Once we trusted it, it became the gate.
• A 5-minute stop window. Every automatic reply waits 5 minutes so a person can stop it.
• Human takeover in Slack. Objections it can't answer, such as pricing, come to us.
• Kept contained. It only uses campaign content our clients have approved, in an isolated environment.
• Data requests go to me. If someone asks where we got their details or asks us to delete them, I handle it and the outreach stops.
• Cold calling stays with people. AI calls aren't good enough yet in my view.

Results

• Meetings. In the talk I said that, like for like over one week, our human SDR was booking 4.8 meetings a week and our reply desk 31. That's my spoken comparison with no dates attached, so read it as a snapshot.
• My own diary. On 28 Oct 2026 I had no gaps for the next 2 weeks, a first. The agent now books our sales reps instead of me.
• The team. More time on demos and closing.

Speed, share on autopilot and cost per meeting, with their windows, are on our reply desk on autopilot.

Mistakes and Lessons

• Outsourcing the build was slow. Someone from the business has to be in the loop.
• The reviewer first sent work to people. The fix was to send its suggestions back to the reply agent, with the reviewer giving final approval.
• The agent is a small part of the process. If the message isn't delivered and read, there's nothing to reply to.

The Foundations Underneath

What has to be in place before an agent has anything to answer.

• Several sending domains, so your real domain stays protected.
• Two inboxes per domain, at very low volume: around 13 emails a day.
• Domain monitoring, so a domain whose health drops is swapped for a new one.
• Campaign copy aimed at your ideal customer, with verified leads so nothing bounces.
• A safe daily limit for each LinkedIn profile (for one example, 10 messages and 10 connection requests a day).

Advice for Anyone Starting

1. Start with the part your team likes least, usually prospecting.
2. Put the foundations in place first.
3. Approve everything at first, and teach it with your past messages.
4. Automate a campaign only when a person rarely edits its work.
5. Add a reviewer agent.
6. Decide what always goes to a person, such as pricing and objections.
7. If your agency grew on word of mouth, outbound is revenue you haven't turned on yet.