Where to Start With AI:Automate the Job You Would Rather Not Do

Start with something you would rather not do that still has to be done decently well. For me that was interviewing, and building a voice agent to run the first screen taught me more about AI adoption than a pilot project would have.

The candidate who called back three times

One of the women applying to be our nanny called our AI interviewer back three times. The first call, the agent glitched. The second, it hung up on her by mistake. She called back a third time and finished the whole interview anyway.

On paper she was not our first choice. We hired her, she watched our son for four days at a trade show, and she was the best hire we made all summer. On day two she hit a freeway accident and texted her ETA before she was late, then kept updating until she arrived.

That is the whole job, right there.

Pick the task you avoid

A lot of people tell me they want to use AI and have no idea where to start. Here is where I landed.

For me that is interviewing. I am too nice. I let people talk far too long because cutting in feels rude. I cannot think of a sharp follow-up on the spot. And I hate the scheduling more than the interview.

So I built a voice agent to do it. Vapi for the phone layer, and a skill where I hand it a candidate's profile and it writes her question pack, loads the agent, and dials at the time she picked.

The first problem was not technical

Some candidates asked why they had to talk to an AI at all, and said they would rather speak to us. Fair question. My answer was honest: we cannot interview everyone, and this way you get interviewed this week instead of maybe never.

But the fix that worked was one line, and it was not in the agent. I put it in the job post. We start with a short AI phone screen. Everyone who applied after that already knew, and the objection stopped coming.

Steal that one. What people object to is the surprise, more than the machine.

Fix it with a loop, not a rewrite

The bugs were real. When someone paused to think, the agent would cut in with "okay" and talk over them. Other times it went silent. Once it mistook a call-screening message for voicemail, said "Goodbye" and hung up on a qualified candidate eighteen seconds in.

So after each call I had Claude read the transcript, tell me exactly what the agent got wrong, and fix it so it could not happen twice.

That loop is the whole trick. Around ten calls in, it was reliable.

We still interviewed eleven people in one week. Most ran about fifteen minutes.

Leave the judgment where it belongs

One did not. She asked twenty-six questions. What does a normal day look like. Does tidying up mean his things or the family's laundry. By the end she had spoken more words than the agent had.

She had less experience than the others and few reviews. We listened to that recording several times and hired her, because someone who asks what is expected of them before taking the job is telling you how they will work after.

The agent had no idea any of that mattered. It asked its questions, kept the recording, and left the judgment to us.

Where I would start

You do not need to build a nanny screener. But you probably have your own version of the thing you avoid and still have to do well.

That is where I would start.