Where should the human be in human-in-the-loop AI? Before the prompt, not after the output.

Sally is always at the water cooler. She might as well roll her chair all the way back there and flop her laptop onto her wrinkled lap. She owns the real estate. I have watched these laps of hers enough times to know that she never quite finished eating before she starts walking there to fill up her mug. The crumbs dropping from the corners of her mouth leave a Hansel-and-Gretel-level fantasy trail to prove it. This is where she does her best procrastination (and spying), whilst her daily tasks lie with gaping, gasping mouths waiting for her to throw them back into the sea for someone else to catch tomorrow.

We don't get to choose our coworkers, in as much as we don't get to choose our family. We don't control their work ethic, their productivity, their team spirit or their people-pleasing tendencies. Employees fall into this category too, and more often than not it takes an entire other bucket of skills to navigate. Particularly when revenue is dangling from the untied shoelaces of their productivity sneakers.

We don't get to choose, until we do, thanks (or no thanks?) to AI.

Human-in-the-loop or human-in-the-poop?

So when do we choose to keep a human in the loop because it is the best risk-based approach, versus doing it because we're averse to changing the way we work? Who decides this? You might think, by the way I have framed my intro, that I am leaning towards AI as our saviour. I assure you that I very much do not believe you can (nor should you attempt to) slap AI on anything until you have clearly examined your processes, understood why you do things the way you do them, and assessed exactly which fruits are ready for the AI fruit salad. But this is not my point. My point is that human-in-the-loop is cardinally important. We are just slotting it into the wrong step in the process.

AI governance frameworks place the human checkpoint after generation. Review the output, catch the errors, approve or reject. That checkpoint protects accuracy, perhaps, but it does nothing for originality.

Research on creativity points the same way: a 2025 causal stimulation study found creative output depends on disrupting network integration before idea formation, not on correcting the output after it. A separate multi-country study of 2,433 people tied creative performance to the brain's ability to switch modes during the thinking process itself, not during review.

The checkpoint that protects originality sits before generation starts.

It lives in the unstructured space between noticing something and doing anything with it. No framework I have come across to date names that step. No deliverable comes out of it. It's the reason the deliverable is worth anything.

Human sign-off after the AI fact is quality control. It was never meant to protect original thinking, and treating it like it does is the biggest mistake we are making today. It feeds the illusion that because we cast our eyes over the output and no obvious mistake slapped us across the face, we have successfully integrated the human edge. No, Sally. You must still be dehydrated.

Protect the gap between having a thought and touching a keyboard, and the review stage barely matters. Skip that gap, and no amount of careful editing gets your originality back.

The coworker who doesn't chew loudly, steers clear of gossip and never, ever says no

So how do we create human-centred AI balance, particularly in a workplace with so many variables? I get asked this question A LOT. Every person in your organisation is unique, with their own values, integrity, personality, adaptability and skills. A symbiotic outcome that lands the plane with zero friction every single time can seem near impossible.

If there is one feature in Claude that has changed my life, it is, most definitely and unequivocally, Skills. A Skill is a set of instructions, examples and files you give Claude once, so that it can do a specific task your way, every time.

The best way I can explain it: a Skill is the obedient employee, the happy coworker who just gets on with it. You show them once, maybe they ask one round of questions, but as long as your teaching (your SOP and your data) is clear, you will never have a problem with them again. You always get what you asked for, exactly the way you asked for it.

Garbage in, garbage out

It has to be said first and loud: garbage in, garbage out. Just like an A-type employee, Claude (or any AI) loves clarity and accurate data, in a format and pattern it can work with, and which doesn't make it want to hallucinate. The only way to give clarity is to have clarity, and you are going to hate what I tell you next:

You need to get serious about documenting your procedures. You need to get serious about structuring your data in a clear, cohesive and AI-friendly way. Until you are ready to grease your elbows and dive into that, I cannot give you any news that will make you happy.

After many, many, many teeth-pulling AI testing rounds whilst building my own IP Bedrock Brain, and figuring out the hard way what NOT to do, I developed two proprietary AI tools that extract know-how from employees' and founders' minds, from scattered systems and from undocumented processes. That is, and always will be, the first and most important step before you let AI touch anything in your business.

Where to start

The Process Automation X-Ray is where those tools live. It captures each process step by step and scores it for automation, and the Founder X-Ray maps what only you know. If you want to talk about succession, a sale or getting your business ready for AI, email me for 1:1 time.