AI Is Extraordinary. Prompting Is Harder Than It Looks.

AI companies have converged on a single interaction model: prompting. You type what you want; the AI responds. It works remarkably well for many tasks, and the capabilities have advanced faster than almost anyone predicted. I use these tools daily and I’m not remotely sceptical of their value.

But prompting is only three years old, and it is worth asking (honestly) whether it is the right model for the full range of people who need to use AI. Not for early adopters and power users. For everyone.

Who We’re Actually Talking About

One problem I observe regularly is that technical users are the most vocal early adopters, and they tend to set the terms of discussion. For every bright youngster or enthusiast throwing themselves into the new AI world, there are probably ten people on the shop floor or in the accounts team who are indifferent, uninterested, or frankly hoping it all goes away.

Benedict Evans made this observation particularly well in a recent piece on AI and transformation. He noted that most people aren’t tool builders: they don’t spend their working day thinking about how to optimise their processes. Most developers, and certainly most of Silicon Valley, are tool builders, which makes it genuinely difficult for them to understand the majority of their users.

So when we ask whether prompting is the future of AI, we need to be clear about who we’re asking the question for.

The Blank Page Problem

Prompting confronts users with something unfamiliar: a blank page. You have to know what you want and know how to ask for it. This is the opposite of conventional software, which provides interfaces that give users context for their task and a framework for how to approach it. Many UIs do this poorly (mobile apps in particular seem to delight in requiring users to guess which gesture does what) but even a bad interface is a framework of sorts.

The blank prompt requires a different set of skills entirely. I’ve seen less technical users genuinely struggle with it. They simply don’t know what to type at all.

AI companies would argue that their tools solve this problem: the AI is an assistant who helps you work out what you need. And to a degree, this is true: AI does ask clarifying questions when given something vague, and increasingly searches the web on your behalf. But AI tools still tend to be far better at solving problems that exist in the external world than problems tightly bound to existing company knowledge. The AI can research flights and plan your holiday, but it doesn’t know about the food allergies or the grumpy uncle that make up real life. At work, this gap is even wider: most tasks involve referencing company data. You can’t address the problem of chasing late invoices without access to the invoicing system.

Typing Commands Feels Like a Step Backwards

Prompting requires people to type commands. This comes naturally to younger users, for whom typing is second nature, but it is considerably harder for those who aren’t used to it.

There’s also a more subtle psychological barrier. For people who have spent thirty years clicking on buttons, interacting by prompt can feel like a backwards move, distinctly technical, geeky even. It’s easy for those of us who are comfortable with technology to forget that there is still a meaningful proportion of users for whom computer use is occasional, and for whom this is a genuinely significant adjustment.

AI companies push prompting because they’re looking towards a future where AI is smart enough and knowledgeable enough for the conversation to feel indistinguishable from talking to a human, at which point it really should work for everyone. The challenge is getting there.

The Psychology of Iterative Chat

Prompting introduces psychological complications that don’t get enough attention.

Complex tasks rarely come right first time. The AI produces a blog post or a report; you review it, request changes, review the revised version, request more changes. This iterative cycle is a fundamentally different mode of working from just doing the task yourself, and it presents two distinct challenges.

First, it requires intermittent attention. You ask the AI to do something, it goes away for seconds or minutes, then comes back needing a response. What exactly do you do in those ten-second gaps? It is still faster than doing it yourself, but people’s experience of time is subjective, and intermittent attention is not a comfortable way to spend it.

Second, you have to explain the changes rather than simply making them. This is a completely different skill from doing the underlying work. Some people can produce an excellent presentation but find it genuinely hard to articulate to a third party what needs to change. That ability (to direct, review, and iterate) is separate from the skill of doing the task, and it doesn’t come naturally to everyone.

What exactly do you do in those ten-second gaps? It’s still quicker than doing it yourself, but people’s experience of time is subjective, and intermittent attention isn’t a comfortable way to spend it.

Cost, Rationing, and Too Many Moving Parts

Modern AI is expensive to run. A single conversational exchange costs very little, but asking the AI to research a topic, write a detailed report, and refine it across several exchanges is a meaningfully different proposition.

AI companies are already under significant economic pressure. Data centres cost tens of billions of pounds and take years to build, while demand keeps growing. The response has been to squeeze users: rationing access at peak times, ring-fencing the most capable models behind premium tiers, and quietly degrading response times at the standard level. Claude measures usage over a five-hour window; push too hard and it pauses until the window resets. Imagine if your intern told you they’d had enough questions for the morning and would pick things up again after lunch.

On top of rationing, users face the challenge of not knowing how far to push. Give the AI too complex a task and it fails; ask too little and you leave value on the table. And because AI capabilities are improving rapidly, the answer to “can it do this?” changes month by month. In our office, we found that AI wasn’t capable of producing network diagrams, until, a month or so ago, it suddenly was.

Then there are multiple providers to navigate: Anthropic, OpenAI, Google, Microsoft, and more. The most capable model changes every few months. Many power users now maintain accounts on more than one platform to ensure they’re always on the best available option, which doubles the learning required and adds another layer of complexity for anyone trying to keep up.

AI is only three years old and has only recently got out of nappies. It would be surprising if the interface model were already optimal.

What AI Currently Asks of Its Users

Taken together, these challenges paint a picture quite different from the vision AI companies sell. Rather than a helpful assistant waiting at your beck and call, AI currently places significant demands on users: they need to learn how to prompt effectively, providing context and structuring requests well; they need to work out how hard to push and when a task is too complex for a single exchange; they need to understand rationing and, for power users, when to switch between models and systems.

This is a new kind of literacy. Not insurmountable, but substantial. We shouldn’t assume that users will naturally take to it. Many won’t, and that’s reasonable.

What Comes Next?

AI companies are working actively on all of this, and there are a few directions worth watching.

Voice seems like a natural alternative to typing, and AI is now remarkably good at understanding spoken English, dramatically better than the previous generation. But voice-based AI chat is technically demanding: a convincing conversation requires a sub-second response time, which is very hard to achieve. It’s also economically expensive, since server resources must be available at the exact moment the user speaks. And culturally, it remains awkward: people find it embarrassing to chat with an AI in a shared office. Voice will grow enormously as costs fall, but it isn’t yet a practical solution for most workplace contexts.

Integration into existing systems is the obvious route, and it’s already happening in customer support and operational applications. But it is slower than people expect. Truly taking advantage of AI often requires redesigning large systems from the ground up (ferociously expensive) and AI tokens carry an ongoing, non-trivial cost that is genuinely difficult to account for in a budget.

New interaction modes are still emerging. There are promising directions (AI that observes your work and builds genuine understanding of your context over time, becoming a real assistant rather than a sophisticated query engine) that could eventually change the picture substantially. We are, genuinely, only at the beginning.

The future of AI may be one where prompting is what the tools use, not what the users do.

At Blueberry, our view is that prompting works well for people who have taken the time to get good at it, but that most users need something different: a model where the prompting complexity is abstracted away, and AI is embedded into purposeful, managed tools rather than offered as a blank interface. That is a significant part of what we are building with VibeMagic: an environment where non-technical users can direct AI to solve real business problems without needing to master the full prompting model themselves.

The question of whether prompting is the future may turn out to have a more nuanced answer than either its enthusiasts or its critics suggest. It will probably remain the dominant model for those who are motivated to learn it. For everyone else, the interface will need to evolve, and it will.