AI can do the legwork, but you need to do the thinking

AI is still more often than not presented as an end-to-end solution. While OpenAI’s promotion for ChatGPT Astra focused more on humans, it still showed them asking the model to generate 3D models and slides, delighted with the results and never touching a wireframe or text box themselves.

Four years after beginning the clumsy process of integrating AI with my daily life, I feel I’ve settled into the right rhythm for its proper use – and my experience is quite different. For LLMs to save time, rather than creating more work, you still need to do the thinking yourself. AI can handle the busywork, but you need to maintain control of the creative direction and quality.

A collaborative project

The thought occurred to me as I was putting together a simple slide deck for an upcoming team meeting. My global colleagues had recently attended a series of strategy meetings in London, and I needed to share the key decisions with the wider team so they were aware of the upcoming changes.

I distilled my written notes into hierarchical bullet points in Notepad. Then I pasted the list into Copilot and asked it to generate a PowerPoint deck in a basic template. Within a few minutes, I was downloading the PPTX file – complete magic compared to the process of manually populating slides.

Things have changed so much in the last few years that some people might even consider my quasi-AI approach to be wasteful.

I spent around ten minutes reviewing the deck, which had six or seven slides. I added detail where I thought it would be useful to the team. I corrected a few errors and inaccuracies (for example, Copilot interpreted our MTTE metric as “mean time to engage”, not “mean time to escalate”). Before long, I was happy the deck represented a good enough outline for me to speak around.

I reflected afterwards on how the whole process would have been unthinkable a few years ago. At that time I’d have had to build every slide by hand. But now things have changed so much that some people might even consider my quasi-AI approach to be wasteful. Why not just have AI do the whole thing? We’re on the cusp of artificial super intelligence, after all.¹

The missing ingredient

As I rolled the question over in my mind, it naturally broke down into a few more, which helped me to straighten out my thinking:

Why didn’t I just ask AI to create the deck from scratch? Lack of context, primarily. Had I asked AI to build a PowerPoint deck for my next team meeting with no further instructions, I’d likely have received a useless, generic presentation with an agenda full of high-level items like “updates”.

Why didn’t I ask AI to review my written notes for me? My notes were bullet points, jotted down in the order the items were discussed at the meetings. An LLM wouldn’t have the organisational knowledge to know which were important enough to spend time on, which were significant enough deviations from the team’s current ways of working that they needed to be discussed, which disparate bullets were actually connected, and so on.

Why didn’t I take the AI-generated slides straight to the meeting? Even after doing my best to guide Copilot with my prompt, there were still details it didn’t understand (like the MTTE confusion). LLMs also have a habit of adding flavour text to their creations, so there was some confusing messaging in the final deck that I needed to rectify through some tweaks.

AI saved me time shuffling elements around in a slide deck, but its output wasn’t a finished product. It provided an approximation of my desired output, and I needed to refine it by applying a layer of context over the top.

The rhythm of AI

There are two main ways I now use AI: to handle busywork, and (to borrow Anthropic’s phrasing) as a thinking partner. Both of these use cases require caution. AI-generated results require the same review that anything created by a human would – probably more, given that the human has more context, and would likely pause and ask questions where they lack it. I’ve noticed that the biggest complaints I see about AI usage tend to involve either:

1. Somebody who wouldn’t usually perform a task asking AI to do it for them. They’re unfamiliar with the subject matter and don’t know how to properly check the output. Therefore, they don’t spot errors that someone more familiar would. The classic case is the non-developer who vibe codes software and takes it to market, missing security vulnerabilities and code structures that cause issues at scale.²

2. Somebody who is a subject matter expert leans on AI too heavily, trusting its output without question. This over-reliance leads to errors that the expert would never make if they completed the work themselves – like the confusing messaging I’d have brought to the team meeting if I hadn’t reviewed the AI-generated presentation.

My opinion is that AI shouldn’t be doing the things you’re really getting paid to do. If you spend your time on repetitive tasks then you need to be thinking, and if you think then you need to be good enough that AI can’t replicate it. The insight and care that you bring are what set you apart from other humans, and they’re what will continue to set you apart from AI models.

Even in the workplace the AI companies envision, where an LLM sits in a central location and sees all of your documents and meetings, it still won’t have the full context. Your role is to apply that context, along with taste and discernment, while making the most of AI to waste less of your time on low-leverage tasks like nudging text boxes and rectangles around a screen.

But the care – the vision, the substance, the messaging, the quality – is a human touch applied on top. AI can generate a wireframe of your presentation deck, but that last layer of polish, the extra context and colour you bring through your delivery, are what makes it engaging and effective.

AI labs make their case to humanity as agents run amok →

AI companies are softening their doomer predictions – just as they're losing control of their models

Notes and references

1. I originally included the term in jest, but on the eve of this post’s publication Microsoft CEO Satya Nadella published an X post calling for greater oversight of “black box” AI systems, in which he pivoted to the “super intelligence” nomenclature.

2. Craig Mod has an excellent essay titled Software Bonkers exploring the appropriate uses of vibe coding. In short: It’s useful for building custom, local software for an audience of one. In that scenario you don’t (usually) have to worry quite so much about scale and security.