I keep seeing the role prompt debate pop up over and over again. Some people treat it like a cheat code while others roll their eyes and move on.

Something delicious is happening in consulting.

Prompt injections are trending again. They are back in the spotlight because agents now browse the web and take actions on your behalf.

You know that moment when a client says, “We’ll think about it,” and you’re left wondering if that means “Maybe next week” or “Never call us again”?

You would not hire Iron Man to do strategy, marketing, and business development.

Yet that is exactly how most people treat their AI.

AI was meant to clear the clutter. But too often, it multiplies it. A new tool here, another “game-changer” there, and suddenly you’re spending more time managing the stack than enjoying the results.

Your AI isn’t stupid, it’s overfed.

The more documents, slides, and half-baked notes you throw at it, the more it starts to sound like an overconfident intern who skimmed the company wiki once and now insists on answering every question.

You feed your AI like a freeloading flatmate. You throw it prompts, upload knowledge, connect it to apps, and fix its mistakes just to keep it alive. At first it feels novel, like training a clever pet.

AI failure has become the new corporate pastime.

I’ve watched my AI answer questions like it was centre stage on karaoke night. It was loud, confident, and completely off-key.

Sam Altman said last week that only 7% of paid users have ever touched the o-models.*

I built my AI cocktail bar. Gemini 2.5 Pro handles most of my research while claude 4 Opus analyses the data and GPT 4.1 gives me direct feedback. Everything worked exactly just as how I engineered it right down to the last pip.

Does your AI agent sound like it moonlights at a budget call centre, even after you’ve fed it your best ideas (and a few hundred prompts)?

It’s official! “AI expert” is the new avocado toast.

But if you’ve ever opened an invoice and wondered, “Wait, how did my ‘free’ AI experiment turn into a line item bigger than my caffeine budget?”

AI models keep getting stronger. Claude 4, Gemini Pro 2.5, KIMI K2, and GPT-4.5 all push the limits of coding, analysis, and reasoning.

For the past year and a half, I’ve been steadily growing my caffeine tolerance.

You feel it every time you scroll… someone is running off with your lemonade stand.

If launching an AI assistant or agent is now as easy as sending an email, yet, why do most of us still feel… stuck?

If you haven’t been a hermit for the past couple of years, you’ve heard it all: