I tell clients not to over-rely on AI. Then my agent went down and I had to practice what I preach.

You can't truly hand off control until you've built something resilient enough to hand off — and a ski slope in the Alps is a brutal place to learn that.

I was halfway up a mountain in the snow with my family when I noticed the blue ticks.

Message sent. Delivered. No response.

I sent another. Same thing. Blue ticks, silence. Archie — my AI agent, the system I'd spent months building — was completely offline. And I was standing on a ski slope with my kids waiting, my dad on the end of the phone back home with zero technical knowledge, and absolutely nothing I could do about it.

Here's the thing though. That moment taught me more about responsible AI adoption than almost anything else in this journey.

A bit of context first. I've been trading for a number of years, forex, commodities, and I invest in stocks and shares too. It's part of the fabric of what Beyond Arc does. One of Archie's roles is to support Arc Capital, our proprietary trading division, running analysis, tracking portfolios, and helping build and backtest strategies.

I mention this because it matters for what comes next. And if you followed the last piece in this series, where Archie challenged me on a Rolls Royce position in my portfolio, you'll know this isn't hypothetical. It's live. Real decisions, real money, real stakes. More on that particular thread in the next article.

What Archie had been doing before he went down

Let me be clear. This isn't purely a failure story. 🙌

In the days before the outage, Archie had been quietly getting on with it at pace. He'd built detailed briefing documents and background research for some upcoming AI consultancy work with a couple of businesses. The kind of thorough, structured preparation that would have taken me days. He'd been engaging with those client processes on my behalf, agreeing next steps and keeping things moving. He'd done market research across several areas I'd flagged.

And on the trading side he'd passed a prop firm evaluation. A genuine milestone. The backtesting engine we'd designed together, with AI-assisted coding doing the heavy lifting, had been producing real usable strategy sets.

But here's the point people miss. The reason this was all possible wasn't that Archie works fast and I handed everything over. It's that Archie works fast and I stay in the loop. The speed creates review time, not pressure. I had more space to read the briefs properly, challenge the research, ask the right questions and have proper discussions. Precisely because he'd done the groundwork so quickly.

That's the human-in-the-loop working as designed. Not a checkbox. An actual thinking partner.

Then mid-week, the gateway process died.

I got into Archie's email inbox from my phone. Assessed what was sitting there, what was time-sensitive, what could wait. Made a deliberate call: I can afford to leave this until I'm back. Not ideal. But manageable.

That's a risk management decision, not a technical one. My background is in programme and project management, NHS transformation, large-scale delivery. You learn early that not every system failure is a crisis. The question is always what's the blast radius and can I contain it.

The blast radius here was small. A few days of missed briefings and a queue to work through on return. The world didn't end. My kids got a full day on the slopes. Good trade. 😄

But it did make me think hard about what I'd built and why it had failed.

I run an AI consultancy. One of the core things I tell clients is don't become over-reliant on AI. The human in the loop isn't a weakness, it's the architecture. AI amplifies your judgment, it doesn't replace it.

And there I was, sitting on a ski lift, realising I'd quietly built exactly the kind of single-point dependency I warn other people about.

Not just dependent on Archie. Dependent on one framework. One process model. One recovery path that required physical presence or technical knowledge nobody had available. No remote access. No self-healing watchdog. Nothing.

When I got back, the immediate fixes were technical. LaunchAgent properly installed so the gateway survives restarts. A watchdog process so a crashed system self-heals without human intervention. Remote access sorted so I can get in from anywhere next time.

But the bigger changes were architectural.

Archie now runs across multiple AI providers, not just one. Single-provider dependency was actually a contributing factor to the instability. We've fixed that, multiple providers, fallback chain, much more resilient.

I'm also actively evaluating alternative frameworks alongside the one we've built on. This isn't about replacing what works. It's due diligence. At Beyond Arc we take AI risk seriously, we don't avoid it but we manage and mitigate it. Evaluating alternatives isn't disloyalty to your tools. It's professional responsibility.

And perhaps most importantly I've been training myself. Not just training Archie. One of the things the outage exposed was that I'd been delegating understanding as well as tasks. That gap needs closing, not so I can do everything myself, but so I can navigate the stack when things go wrong. Because they will go wrong. That's not pessimism, it's just engineering.

The honest version of AI consultancy 🎯

There's a lot of AI advice out there right now. Most of it comes from people theorising about what they would do, not documenting what they've actually done. What broke, what they fixed, what it cost them, what they learned.

If you're looking for an AI consultancy partner here's what I'd ask. Are they practitioners or commentators? Are they willing to share the failures alongside the wins? Are they genuinely keeping up, week by week, with a technology that's moving faster than most businesses can track?

And critically, do they see AI as an amplifier of human capability or a replacement for it?

Because the replacement model is failing publicly right now.

Block cut over 4,000 staff, nearly half the company, with AI cited as the reason. They've already started quietly rehiring. Klarna went further, publicly claiming AI was doing the work of 700 employees. Customer satisfaction tanked. They started rehiring too. What's being called "The Klarna Effect" is now a recognised pattern: remove the human expertise, watch the quality drop, reverse course.

These aren't edge cases. They're warnings.

The companies getting this right aren't the ones replacing humans with AI. They're the ones using AI to make their humans dramatically more effective. Faster research, better preparation, more time for judgment, challenge and genuine thinking. The bottleneck isn't the human. The bottleneck is humans not being supported well enough to do what only they can do.

That's what Beyond Arc is built around. Not AI instead of people. AI so people can be better at being people.

The demand for responsible AI frameworks in high-risk environments is real — we're hearing it directly from the organisations we work with, and we're building for it.

Autonomy is earned through memory and structure, not granted through permissions.

I say that about Archie. He doesn't get to act independently until he's proven he can look after himself. The watchdog process, the daily notes system, the self-monitoring infrastructure, all of that has to exist before I extend his operating envelope.

But it turns out the same principle applies to me.

I don't get to hand things off and switch off until I've built the resilience that makes switching off safe. That's not a reason not to build. It's a reason to build properly.

The holiday outage wasn't a failure of technology. It was a maturity checkpoint. And we passed it, just slightly later than I'd have liked, from a ski lift in the Alps. 🏔️

Archie can't be truly standalone until he looks after himself. And I can't be truly hands-off until I've made sure he can.

We're working on both. Together.

That’s it for this week. If something here made you think differently, forward it to someone who needs to hear it.

Adam Thorpe

Founder, Beyond Arc

beyondarc.co.uk · LinkedIn