What we got wrong in our first six builds

We've now built agents for over forty businesses. The last thirty went well. The first six taught us everything.

Here's what we got wrong, in the order we got it wrong.

Mistake 1: We built what clients asked for instead of what they needed

Our first client ran a landscaping company. He asked for a chatbot that could answer questions about his services on his website. We built it.

What he actually needed was something that captured leads from people who visited after 6 PM and never filled out the contact form. The chatbot answered questions but didn't push for contact info. Visitors got their answers and left.

Three weeks in, we rebuilt it as a lead capture agent — ask a few qualifying questions, offer a free estimate, collect name and number. Bookings from the website doubled within 30 days.

What we changed: We now start every project with a discovery call where we specifically ask: "What does success look like in 90 days? What would have to be true for you to call this a win?" The answer shapes the build. The feature list follows from that, not the other way around.

Mistake 2: We under-trained on tone

An agent we built for a physiotherapy clinic was technically correct and completely wrong. It answered questions accurately. It booked appointments. It confirmed times. It also sounded like a legal document.

The clinic had warm, reassuring front-desk staff. Patients with chronic pain don't want to interact with a system that sounds like a terms-of-service page. We heard about it in the first week.

What we changed: Every agent now goes through a tone calibration process before launch. We collect samples of how the business actually communicates — emails, texts, social posts, even voicemail greetings — and use them to define a voice profile. We test the agent against it before the client ever sees it.

Mistake 3: We didn't build fallback paths

Our third build was a voice agent for a restaurant handling reservations. It worked well for standard requests. Then one Friday night a caller asked about a private dining room for a corporate event — something not in the agent's training data.

The agent looped. The caller hung up after 45 seconds of confusion. They left a one-star review.

What we changed: Every agent now has a defined fallback for anything outside its scope: acknowledge, apologise, and hand off. Either transfer to a human, send a text to the owner, or offer a callback. The agent never loops. It exits gracefully.

Mistake 4: We ignored the calendar edge cases

A booking agent we built for a dental practice didn't account for blocked-off time during staff training days. Patients booked appointments that couldn't exist. Two people showed up to a closed office.

We'd integrated with the calendar but hadn't mapped every exception. The client hadn't thought to tell us. We hadn't thought to ask.

What we changed: We now run a two-week shadow period on every booking agent before go-live. The agent runs in the background, logging what it would have booked, while a human still handles real bookings. We review the logs together and catch the edge cases before they become incidents.

Mistake 5: We launched without measuring the baseline

For our fifth build — an invoice follow-up agent for a consultant — we had no data on what the collection process looked like before. When the client asked if it was working, we had nothing to compare it to.

We thought the results looked good. The client thought about the invoice she'd been chasing for ninety days and wasn't sure.

What we changed: We now require a two-week data collection period before any build begins. We log baseline metrics — call volume, response times, collection days, no-show rates, whatever's relevant — so we have something real to measure against. Month-one reporting is built into every engagement.

Mistake 6: We thought go-live was the finish line

Our first five clients all got agents that worked well at launch and drifted over the following months. New services weren't added. Edge cases that came up in the real world weren't folded back in. One agent was still quoting prices that had changed.

We were treating each build as a project with an end date. AI agents aren't projects. They're systems that need to evolve as the business evolves.

What we changed: Every engagement now includes a monthly tuning call. We review what the agent handled, what it missed, and what changed in the business. The agent gets updated. This is now standard, not optional.

We're proud of what we build now. But we wouldn't build it this way without having built it badly first.

If you're evaluating AI vendors, ask them what their worst build looked like. Anyone who says they've never had a bad one is either lying or hasn't built enough.