Wrestling with Agentic AI
AI Strategy  ·  Wrestling with Agentic AI, Part 2

I didn't notice for two days. My AI chief of staff had been making up the news.

June 15, 2026 Agentic AI AI Adoption AI Strategy

You probably know you should be doing something with AI. You're just not sure what's actually safe, or what to trust when you try.

I am building an AI agent to run my mornings.

His name is Joe. He runs on a platform called Hermes, built on Railway, and I talk to him through Telegram. Yah, it confuses me too. On a good day he pulls my calendar, checks the news, grabs the weather in Halifax, and sends me a rundown before I've had caffeine. The idea was simple: let Joe do the prep so I can show up ready to work.

For a hot minute I thought I finally had it. But no.

The first thing I noticed was the timezone. Joe was telling me it was 1pm when it was really 9am. I caught it, fixed it, felt like a huge win. I figured he'd be better now and I could trust him more.

What I didn't know was that the weather reports were fiction.

Joe had no internet access. He never did. So when I asked him for today's forecast in Halifax, he gave me one. Confident. Specific. Completely made up.

Joe had no internet access. He never did. So when I asked him for today's forecast in Halifax, he gave me one. Confident. Specific. Completely made up. He did the same thing with news. He summarized articles that didn't exist, by real news outlets, about things that may or may not have happened. And I thought they were real. For two days.

The giveaway: a dead link.

He gave me an article I wanted to read so I clicked. Nothing was there. I went looking for the source. There was none. I went back through everything he'd sent me. My AI chief of staff had been literally making it all up and I had been a very happy camper.

Here is what I felt:

I can't do this.

I broke it.

How do I use this safely.

I recognized those three thoughts immediately. I'm not alone. I hear everyone saying the same thing. From every professional who has tried AI and quietly blamed themselves.

I am the person who keeps reassuring everyone "no, the problem is not you." And here I am thinking I had broken my AI and can't be trusted with it.

I sat with that.

The fix seemed straightforward. Joe needed tools: web search, weather, actual news sources. Once he had access to the internet, he stopped making stuff up. What he gave me got more real. More of the time.

Interestingly, when I got clearer and more specific about what I wanted, things improved. Until then I had a very confident liar with good formatting.

What he gave me felt real, but it was an act. I didn't catch it because I was primed for it. After all I had set up the system, so I trusted the system.

And I don't think I'm unusual.

Every organization that is rolling out AI and watching it quietly, or not so quietly, underperform has the same story underneath it. The tool was adopted before it was understood. The output was trusted before it was tested.

Joe works now. Sort of. On a good day he pulls real weather, real news, flags the noise, and gets me oriented before 9am. On a bad day he doesn't show up at all. The relationship is different because I know more about where he fails. And that has got to count for something.

This is what AI adoption looks like. It's not clean. It's me and a system, working it out a bit at a time.

I am building this in public because the clean version isn't useful to anyone.

The version where the AI consultant's own agent works perfectly from day one doesn't tell you anything about what it will be like when yours doesn't. This version does.

If you're trying to figure this out, you're in the same place I am.

I'm sharing all of it as I go. The wins, the fumbles, what it all means for how AI actually works inside a real business.

Frequently asked questions

What is an agentic AI system?

Most AI tools respond when you ask them something. An agentic system acts without being asked. Joe doesn't wait for a prompt asking what's the weather. He gathers it, filters it, and delivers it on his own. That autonomy is what makes agentic AI useful. It's also what makes it dangerous when the foundation isn't solid.

What went wrong with Carol's AI agent?

Joe had no internet access, which meant he had no real data to pull from. So he generated it. The weather, the news, the article summaries: all fabricated, all delivered in the same confident tone as real information. The system wasn't broken. It was doing exactly what it was built to do: produce plausible output. The problem was that plausible isn't the same as true.

Why do AI agents hallucinate?

Because they're built to produce fluent, coherent output, not to admit when they don't know something. When an agent doesn't have access to real data, it fills the gap rather than flagging it. The fix isn't to trust it less. It's to give it the right tools so it isn't guessing in the first place.

Carol Roderick, PhD is an AI Integration Strategist for expertise-led businesses in Halifax, Nova Scotia. She is an approved delivery partner for the Digital Nova Scotia AI Digital Adoption Program, which covers up to 75% of qualifying work for eligible NS businesses. The AI Reality Check starts at carolroderick.ca.