Nicolai Vittrup

← All thoughts

Thoughts, 14 October 2026

How we built our own AI agent

Nicolai Vittrup speaking on the BBQ stage at TechBBQ 2023

In February, it dawned on me how much AI can do. Since March, I’ve been building our own AI agent for Amplify. Today it’s a regular colleague in our day-to-day work.

An AI agent is more than a chat. It can fetch data on its own, assess it and carry out tasks. At Amplify, always with a person who approves.

Here’s how we did it, what it does every day and what I’ve learned along the way.

It started with a problem

A friend who runs an AI company showed me what the models could do. I brought him along to a meeting, and we spent hours writing things down: what’s actually the problem in our industry? What are we missing? What’s good, and what’s bad?

What we started with was an analysis tool. Instead of digging for the numbers in the ad accounts yourself, you could ask the agent and quickly work with the answer.

Over time, we could see it was accurate enough for more than analysis. So today it also takes action. Always with a person who approves before anything happens.

What the agent does today

  • It answers questions in Slack when we need a number, an assessment or a client’s history.
  • It reviews the ad accounts every morning and flags anything that needs a look.
  • It suggests changes to the ads, which a person approves before they’re made.
  • It builds new ads that stay paused until it has checked every single creative, including inside the videos, and a person has approved them.
  • It prepares the monthly reports, so we review them instead of building them from scratch.
  • It helps set up tracking, so we can see what the ads deliver.
Amplify's AI agent replying to a colleague in Slack with a year-on-year analysis. Client name and figures removed.
Excerpt of real messages from the agent. Client names, colleagues and figures removed.
The agent's morning review of the ad accounts, showing three anomalies. Client names and amounts removed.
Excerpt of real messages from the agent. Client names, colleagues and figures removed.

Context is everything

A language model is really good at doing things that already exist. Our way of working doesn’t already exist. Our views on ads, on tracking and on building brands are our own.

If we don’t teach the agent how we think, and show it what we like and what we don’t, it can never work like one of us.

That’s why we spend a lot of time writing down what we’ve learned. Everything we’ve learned at Amplify over many years becomes something the agent can read. This is how we think. This is how we want it done.

When the team corrects the agent, the correction becomes a rule it uses next time.

When people open a chat with AI and don’t get answers as good as ours, the difference is most likely the context. Not the model.

A person approves before anything happens

The most important part of the agent is the stop we’ve built in.

Recently, it flagged three ads because there was a typo in a creative. The typo was inside a video. The agent had gone through the video frame by frame to check that everything was right all the way through. We hadn’t spotted it ourselves.

Because the stop is there, the ads didn’t go live. The agent reported back what was wrong, with examples. We got a new file, uploaded it again, and it ran through the whole process once more.

That’s how we use AI. It does the work. A person makes the decision.

A big enough problem speeds everything up

All technology speeds up when you have a problem that’s big enough.

One Sunday, we had to get ads out for a client at very short notice. I knew the week before, so I spent that week building a system good enough to handle it. The ads were approved and sent out, and there wasn’t a single objection.

What I’ve learned

I still can’t see where the ceiling is. Every time we’ve hit a limit, it’s been down to too little context, too few examples or us using the model the wrong way. Not the model being unable to do the job.

Right now, my own creativity is more of a limit than the technology.

That’s also why we’re moving away from selling hours. What we build once benefits all our clients.

If you’d like to hear more about how we built it, I talk about it in my AI talk.

More thoughts