Here is what I built instead. Not the unicorns and rainbows world. The real one.

This week is Mental Health Awareness Week 2026, led by the Mental Health Foundation. The theme is Action.

The Mental Health Foundation has been tracking the numbers for years, well before AI entered the workplace:

  • 2 million people in England were already waiting for mental health support
  • 70 million working days were already being lost each year in the UK to mental health problems
  • That was already costing employers approximately £2.4 billion annually
  • For every £1 invested in workplace mental health, the return is £5

Then AI came along and everyone was supposed to be instantly more productive.

What the rollout actually looks like from the inside

At a recent HR event focused on AI and workplace wellbeing, a speaker shared a story. A man whose wife was dying of cancer could not bring himself to lean on his friends, so he went to AI and asked how he could support her. Before answering, AI asked him how he was doing. It was, he said, the first time anyone had asked. That’s tragic and alarming in equal measure. AI is not HUMAN. It is an IT, at least for now.

The story landed in a room full of HR professionals because it captures something true and uncomfortable at the same time. AI is filling gaps that humans have left, and for some people, in some moments, that matters. The same event noted that AI use for mental health skews significantly male, the inverse of talking therapies, which skew female. For people with nobody to turn to, AI is stepping in.

That is the potential. The risks sit right alongside it.

There are currently around two dozen legal actions against AI providers for mental health-related harms. AI has been documented as directly encouraging dangerous behaviour around eating disorders and suicidal ideation. It hallucinates, it agrees with you, it reinforces what you bring to it including your fear, your uncertainty, and your worst assumptions about yourself.

People need training on how to use it. Not given a prompt library. Proper training.

The failures that built the system (there were about 500 or so at last count)

I know this because I have lived it and logged every failure, here are two.

Failure one. I fed three hundred pages of council documents into AI and received back confident, well-structured summaries quoting the wrong regulations. Hours lost. The tool was not broken, it was doing exactly what I asked. What I had not done was tell it what good looked like or where its limits were. That failure taught me to brief the machine the way I brief any project resource. Scope, Standards, Defined outcome.

Failure two. A working session ran too long. The context window filled up. The tool started drifting, quietly, without warning, and by the time I noticed, the thread was frozen and the work was gone, and I missed a deadline, luckily not a client facing deadline, an internal one, but nonetheless I missed a deadline, and that really ticks me off. That failure taught me to monitor thread length, write session handover notes, and never let the machine run without a governance structure around it.

I call these moments Claudisms. The moments AI demonstrates, live, exactly why you need a system around it. That failure went straight into the methodology. Oh and just in case you’re thinking build skills love, build skills. Done that mate, a lot of that is fairy dust too.

Here is what nobody selling you a skills package is being honest about. The difference between using AI for one project or drafting the occasional email, and using it across multiple complex, heavy projects simultaneously with high accuracy standards - that gap is real. It has to work how you want it to work, not the other way round. The human has to stay in control throughout, and holding that together takes discipline, and that has to come from the human operator, in this case me.

But each failure has made the system tighter. The document proliferation, the drift, the repetition, the guardrails and standards, all of it.

What I know now that I did not know then

AI has burnt me out in a way nothing else has ever been able to achieve. It is the best and worst thing I have ever experienced. Like that gaslighting boyfriend you couldn’t see for what he was charming, capable, confident, and quietly making you feel like the problem was you.

The problem was never me. It was the absence of understanding the system limitations and building a real foundation around it, based on use not hype.

A context window is not infinite, a model does not have proper guidelines, and anyone who tells you they do is embellishing, and I can prove it. A thread that runs too long will drift. An AI given no values produces output shaped by someone else’s. These are not flaws, they are specifications. Once you know the specifications, you can build around them.

Did it go smoothly? No. Were there a lot of failures to learn from? Yes. But I now have a template, a workflow, and a scalable process I will keep improving. That is what working with AI actually looks like in the real world, not the unicorns and rainbows hype world. A system you build, test, break, and rebuild until it holds. Like we used to do in the old days, find workarounds, push it to its limits, and then push it just that little bit more. Then and only then can you understand it.

What this has to do with mental health

Everything.

AI is not the panacea it is sold as. For every story of a man finding comfort in a machine that asked how he was doing, there are legal actions, documented harms, and a workforce quietly drowning in the gap between what leadership expects and what anyone has been equipped to deliver.

Understanding what AI actually is: what it can do, what it cannot, where it loses track, what it does when nobody has told it who you are is not optional. It is the difference between a tool that compounds your capability and one that quietly compounds your overwhelm.

The Mental Health Foundation’s theme this year is Action. The action worth taking is this: learn the machine before you trust it. Know its limits before you hand it your work, your thinking, or your wellbeing.

I have figured out how to contain it. Have you? Follow for more.

Sam

configure YOUR system. contAIn the chaos. control YOUR outcome.

This article was originally published on Substack.