Mind the Gap: Why AI 'Adoption' Stalls

The Methodology Gap

On the surface level the models and platforms feel very different. They have different, at times very confusing terminology. Copilot calls a pre-saved system prompt an agent, Claude calls it a project, Gemini calls it a gem. It can be very confusing to jump from one to the other, because we have to relearn vocabulary. And because AI is supposed to be the intelligence that does the job for us, we're mostly left to look at the chat box and figure out how to 'talk to it' ourselves. This is very different from the more traditional software tools we learned in the last few decades:

But with AI, it becomes overwhelming: it's not just one new tool, but a whole new way of working wrapped up in hundreds of tools:

Getting to know a model is like getting to know a person. You get to know how they speak, what's important to them, how you interact with them in a way that gets your point across... and it takes a little while to suss out a new human. In many ways it takes a little bit of time to suss out a new model. But a little bit like humans with similar backgrounds, AI models all have very similar backgrounds and similar shortcomings. The biggest shortcoming of current AI might be it being extremely literal and lacking a broader view or context. While AI has a fuzzy knowledge of everything, in context it really only knows whatever you give it. It forgets anything you've been working on the moment you open a new chat.

Different platforms are looking at different solutions to overcome this amnesia. ChatGPT saves little sentences with facts about you and your life, Claude writes an essay about you and builds it's own library of mini instruction manuals as you co-work or code with it, Gemini reads your emails and calendar,... What remains is that an AI can never have all the contextual information at the same time, it needs to rely on it being given the right context at any given time. Either by its memory system or by you. When memory is automatic, it can lead to interesting problems like that holiday you planned last month suddenly influencing the pitch you're writing tomorrow. And when it's down to us, to give AI the right document at the right time we often give either too much or too little.

AI doesn't know what's in your mind. It don't know your experience. It don't know what you actually want unless you say it in very clear language. We can't rely on shared lived experience the same way we do when we meet a new human... Every time we start a new task, we sort of start from scratch. And we have to make sure we build that rapport in.

And nobody taught us how to build that rapport deliberately, the tool churn outside is mirrored by the amnesia inside: every week a new tool, every session a blank context. Once we start to build some routines, some ways of working that are built on that understanding, learning starts to stack up and become transferable. Something that you've learned in one tool, in one model today, transfers almost one-to-one to the next tool or next model that's coming out tomorrow.

The core loops or processes, the core ways of working don't really change just because you're in a different tool. I've been writing about AI as a relationship rather than a tool since 2024 and my MSc research showed that whose who actively reflect on and shape their relationship with AI get a lot more out of it, while staying more independent.

Why "just add AI" stalls

1. The wrong metric is being maxed. We came up with ridiculous mandates like "tokenmaxxing" measuring AI and the humans using it by consumption — tokens, deployed agents, dashboards full — not by outcome. We have not designed processes to make the most of humans. We've designed them to optimise basic measurable metrics. (see Metrics for a GenAI World)

2. Even when people would talk, there's no shared language. Our conversations are stuck between superlatives and existential fear. Executives in "cyber psychosis" running on sycophancy loops (Handy), mid-level managers squeezed between mandate and reality, workers hoarding prompts and hiding their AI use.

3. The orchestration layer is making decisions invisibly, well before we write anything into the textbox. The system prompt no-one sees, the router no-one chose, the data flow no-one can inspect. These have become the de-facto policy layer of AI use, and the conversation about whose values they embed isn't happening. (see The AI Supply Chain Nobody's Figured Out)

4. The artefact short-circuits evaluation. AI outputs look pretty and sound professional, but are often hollow, repetitive, or even plain wrong. And we're not equipped to make sense of AI outputs at the speed they get produced. Anthropic's own AI Fluency Index found that users become "more directive but less evaluative" once the AI produces something.

Artifacts in Anthropic's language are largely pretty, formatted documents, mini-apps, and visualisations. They look finished. They include a lot of information in a very dense surface, and they're titled and structured the way the user asked for. That's the Tool doing the work for us.

Mark shared what that can look like. He asked an AI to generate an org chart and identify missing roles for a new team. AI invented two job titles and he sent the org chart up the chain before checking.

The problem then gets put on the individual user, you know, sense-check AI. But don't spend any time on it. Mark framed the lesson like this:

The failure wasn't the model. Nothing in Mark's toolkit asked him to consider what good looked like before the answer arrived. That's not a prompting problem. That's a methodology problem.

AI is Fast. Slow Down.

The point isn't really to get faster. The point really is to slow down. We used to value our contributions by how much time we spent on the work. Now, AI promises is that if a task takes you less time, you can do five tasks in the same time you used to do one. But the due diligence of properly instructing, informing and checking AI still needs to happen at the speed of human thought.

Think back to the last time you were trying to make AI do the thing you wanted by repeatedly saying: "no, I didn't mean this, I meant that" This ultimately confuses the model. When we fill the chat, or as they say the context window, with bad drafts of something we're working on it can never give us the output we want, because it keeps re-reading all the bad examples before giving the next response.

The antidote is to pause before you prompt.

We need to intentionally slow down. Try fewer tools, collect fewer prompts. But most importantly we need to slow down in how we interact with AI: taking a breather, thinking about what we want, building mutual understanding with the model, and then only asking it to produce the thing. Or maybe even just an outline of the thing, that we then fill in. The sloppiness in much of AI's outputs is an unexpressed want, not AI's failure.

By slowing down, we protect what makes us employable. We still get to add value because we still get to keep that context. We still get to do the critical thinking. And we get to do it with a lot of assistance, with someone that can work with us, that can help us refine our thinking.

Slowing down... actually helps us learn more, faster. Almost counterintuitively. The whole discourse is optimising AI for efficiency — more, faster, cheaper. The question nobody is really asking is how to make it effective: how the relationship with AI has to be set up so that we still get to do the doing, keep and even grow our competence, and keep adding value and feeling valued along the way.

The Practice

In my dissertation research, I found that people form archetypal relationships with AI and often switch between them, depending on the task at hand, time pressure, etc.

Sometimes we just need to get things done, and the Tool is great. Sometimes we want to learn about something new, expand our own capabilities, then the Teacher might be useful. And sometimes we just need to really think about something, refine our argument, push our thinking a little farther then the Sparring Partner really makes a difference.

Anthropic's research led by Dakan and Feller coined the 4Ds: qualities an AI literate person should practice. Delegation, Description, Diligence and Discernment.

The first D almost asks us which archetype we're working in. They don't map one to one, but the first signal is: what are we delegating? Do I delegate the whole work? Then I'm probably in a utilitarian Tool relationship. Am I delegating the holding, the structure, the background knowledge I might not have, so I can concentrate on the things I really need to do? That is very much a Teacher or a coach quality. Or do I want to hold everything myself, but delegate the critique? Do I want to work with a peer, but might not have a coworker to bounce ideas off right now? Then it's more of a Sparring Partner relationship.

Description is often thought of as just the prompt. We need to write nice, big, long prompts. Very detailed, two-page things with several headings and acronyms... do this, don't do that. That leads to that hoarding prompts that might have been written for a different model than the one I'm using, or worse for a different use case. Whereas if you think about description as a collaborative process with the model you're working in, it suddenly becomes transferable. You don't need to learn seventeen new prompting frameworks. You need to think about description as a collaborative process with the AI. Alignment on what's been previously unsaid.

In the next conversation you're having try adding "Keep asking clarifying questions until we both agree that we are on the same page." to any of your requests and see what happens.

Discernment is often conflated with diligence. "You just need to check the outputs, don't let the model hallucinate." Models are always hallucinating. That's the only thing they can do. We call it hallucinations when it gives us something we don't want. And how often do we not spot the things we don't want?

Discernment is about thinking with the model in parallel, actually engaging with the content, noticing when the conversation drifts into a place we don't want it to be. Rather than just saying yes, let's keep going, we might say: no, actually, I want to do this. Or: I agree with these two paragraphs, but I don't agree with the third or the fourth.

Think back to Mark's invented job titles. Taking a moment to sketch out a rough chart of his team himself before asking AI to do it all, likely would have allowed him to catch the pretty-but-useless output AI gave him.

I'm arguing it's not a 'distrust your AI intern' framing, but a 'know your AI colleague's strengths' framing. By making the work we do with AI conversational, we recruit the LLM to help us practice all four Ds. Anthropic's own data backs the conversational reading: in conversations with iteration and refinement, key behaviours are substantially more prevalent (clarifies goal +23.6%, identifies missing context +17.1%, questions reasoning +14.7%).

Now, how do we do this? Think of delegating work to AI as a three step process. First we collaborate with the AI to create a description of the work at hand. It's not about hoarding prompts that other people wrote. And it's not about using AI to create outputs really quickly. It's about the descriptive artifact in the middle — the pre-work agreement. Call it a plan, an SOW, a spec, a letter of intent if you want. This is the key document you need to read, critique and refine the most, it's about giving AI the context only you know. Secondly we let the AI execute the plan we just made. Lastly, we compare the output to the plan, refine if needed and learn what could be better next time. The learning might be for you and your AI. Along the way we might create other documents: a decision log, a maybe-later list, and descriptions of what worked well (Anthropic calls these skills, and the industry is slowly adopting this standard). The things that are still there next week, in whatever tool you open.

Diligence in many ways also includes just knowing when to take a break, when to step out, when to go for a walk, when to talk to a human. Our brains are a lot slower than AI, and they can't just keep going. There's a biology in our brains that stops functioning after twenty minutes, an hour of deep thinking. And AI makes us think really deep, there's a lot of information just screaming at us. So having the diligence of stepping out, taking a break, coming back with fresh eyes, might be almost the most effective single thing to do. Then, once we've detached from the race, we can actually talk about our experiences with our colleagues and peers. Talking about how and when we use AI really helps build a better practice for both the more and less fluent users.

The agentic stake

Getting the AI to collaborate and learn with you is the thing standing between you and the next phase. Agentic AI is the buzzword of the year. And no, we're not talking about Copilot Agents, we're talking about completely automated workflows where AI decides on behalf of you. The promise of agentic AI is that we can stop doing the work, our agents do it for us now. A promise that's both oversold and dystopian at the same time, and maybe why so many of us feel stuck between "needing to learn more about agentic AI" and "resisting it". It opens an existential question. If agents can do everything, what is left for me to do? And yet, outside of software development, agents are still lagging:

The examples we're given are examples where the labs could get an end-to-end workflow to work consistently. At least in a demo environment, at least in an American context, working with very specific big brands. Ordering on food for delivery is already easy. I don't need my ChatGPT to order for me. It's a cute idea, it shortcuts a bit of form-filling, but it's not something people see as a big pain. What people do see as pain is finding the time to go to the dentist that doesn't clash with work and everything else. And it's not as simple as finding an empty slot in the calendar. It's knowing that on Tuesdays you're always expected to be in the whole day, because the CEO is in and has chats with everybody, and you never know when he's going to pop up at your desk. That's not something the agent knows. And so it would book the dentist appointment at 3 pm on that Tuesday, exactly when the CEO comes over. And if your agent can't even get that right, how should it be able to deal with complex work tasks that involve many unknowns and stakeholders?

What agents are fantastic at is doing a well-scoped piece of work completely independently. But because the language around AI is all about outsourcing it to the machine, most people think end-to-end: take this task off my list. Once we go from "just do the whole thing for me" to "which parts are scope-able and outsource-able", it becomes a different promise. The methodology gap is what stands between you and agents being worth anything.

Scoping a subtask well is the method. The alignment conversation, the plan, the memory an agent can run on. A plan-less, memory-less way of working gives an agent nothing to hold. It comes back to building in the friction for the human in doing their thinking.

There are still the things where we need the context, the gut feel, the relationship building that an agent cannot do. Refocusing on the parts where we get to do the thinking, the context holding, the curating. Notice which ones you actually enjoy doing. Can you double down on those?... A lot of the things agents are good at are the things humans are bad at. Doing hours of web research, matching long lists of data line by line,... If we can outsource some of those things, we might have more mental capacity left over to do the hard thinking.

Stop the hoarding, get back to the doing

What if we could do the being better? If AI saves us so much time, we should have a lot more time for being. Being the people, the partners, the family members, the coworkers that we wanted to be, that we maybe didn't have time to be because we were so busy before.

You probably don't need another AI tool. Instead you might go for a walk or have a chat with a friend. And then tomorrow, try using this simple suffix when you start a new task with AI:

Off the back of my MSc research, my experience teaching hundreds of organisations about AI, and ongoing interviews with professionals navigating AI, I'm developing a course on how to collaborate with AI in a way that makes us both more effective at work, but also retains our agency and capabilities. Not a course that teaches you how to use a specific AI app or model, but the meta-skills we need to navigate the world as AI is pushed into more and more corners of our work and lives.

If that sounds interesting, sign up for the course waitlist.