Getting Better at AI: Why Progress Feels So Slow

• Getting better at AI is harder than expected, as most people and companies are still repeating early mistakes from years ago.
• The gap between AI capability and actual human skill at using it is widening — getting better at AI requires deliberate practice, not just access.
• Organisations that invest in AI literacy see measurably stronger returns than those that simply deploy tools and hope staff figure it out.
• Structural and cultural barriers — not the technology itself — remain the biggest obstacle to meaningful AI improvement at scale.

• Getting better at AI is harder than expected, as most people and companies are still repeating early mistakes from years ago.
• The gap between AI capability and actual human skill at using it is widening — getting better at AI requires deliberate practice, not just access.
• Organisations that invest in AI literacy see measurably stronger returns than those that simply deploy tools and hope staff figure it out.
• Structural and cultural barriers — not the technology itself — remain the biggest obstacle to meaningful AI improvement at scale.

Getting Better at AI Is Supposed to Get Easier — So What’s Going Wrong?

We’re now several years deep into the mainstream AI era. ChatGPT launched and quickly became a cultural landmark. Google, Microsoft, Anthropic, and dozens of startups have spent billions building tools that are, by any objective measure, dramatically more capable than anything available even eighteen months ago. And yet the honest question — the one the Financial Times recently put plainly — is this: why aren’t we actually getting better at AI? Why does it feel like so many people and organisations are still fumbling with these tools at roughly the same level they were when they first signed up?

It’s a fair challenge. And the answer isn’t flattering for anyone involved — not for the companies selling AI products, not for the enterprises deploying them, and not for individual users who genuinely believe they’re making progress.

The Capability Gap Nobody Wants to Admit

Here’s the core tension. AI model capabilities have been advancing at a pace that would have seemed implausible five years ago. OpenAI’s latest models and their successors can handle a remarkable range of tasks. Anthropic’s Claude models have made similar strides. Google’s Gemini family is embedded across an enormous product surface. The raw capability is there, and it keeps improving.

But raw capability and actual user proficiency are two completely different things. And right now, the gap between them is widening, not narrowing. The tools are getting smarter faster than people are getting better at using them. That’s not a technology problem — it’s a human one. And it’s the kind of problem that doesn’t get fixed by the next model release.

Think about how the same dynamic played out with spreadsheets in the 1980s and 1990s. Microsoft Excel became one of the most powerful productivity tools in history — and a generation of office workers used it to do little more than make neat tables. The sophisticated functions, the data modelling, the scenario analysis — most of that went untouched. Not because people couldn’t learn it, but because learning it required deliberate effort that went beyond just having access to the software. Getting better at AI in 2025 is a direct replay of that pattern, just compressed and accelerated.

Why Organisations Keep Getting Stuck

Ask anyone who works in enterprise technology and they’ll tell you the same story. A company buys a suite of AI tools — Microsoft Copilot, or a custom GPT deployment, or a raft of AI-assisted coding tools for the engineering team. There’s an enthusiastic internal announcement. A few champions emerge. And then, three months later, usage metrics plateau, ROI conversations get awkward, and the tools become another line item that quietly collects dust alongside the enterprise software subscriptions nobody cancelled.

What goes wrong? Usually a combination of things. First, there’s almost never enough structured training. Organisations assume that because these tools have friendly chat interfaces, employees will simply figure them out. They won’t — or at least, not to the depth that produces meaningful productivity gains. Getting better at AI requires the same thing that getting better at any complex skill requires: guided practice, feedback loops, and enough psychological safety to experiment and fail without consequence.

Second, there’s a workflow problem. AI tools that sit outside a person’s normal working environment — requiring them to tab away, rephrase their thought as a prompt, wait for a result, evaluate it, and then integrate it back into whatever they were doing — create enough friction that people default to their old methods, especially under time pressure. The tools that see genuine, sustained adoption are the ones woven directly into existing workflows. That’s why GitHub Copilot has had more genuine traction than many standalone AI writing assistants — it’s inside the editor, in the moment when it’s most useful.

Third — and this one doesn’t get talked about enough — leadership behaviour matters enormously. When senior people at a company visibly don’t use AI tools, or use them performatively without real engagement, that signal travels fast. Culture in most organisations flows from the top. If the people running teams aren’t genuinely investing time in getting better at AI, why would their reports?

Individual Users Aren’t Off the Hook Either

It’s tempting to put all the blame on organisations and their structural failures. But individuals bear real responsibility here too. The most common pattern among casual AI users is what you might call ‘prompt and hope’ — typing a rough request, getting a mediocre result, and concluding that the tool isn’t that useful. What’s missing is the iterative, curious engagement that separates people who are genuinely getting better at AI from those who aren’t.

Effective AI use is a learnable skill. It involves knowing how to frame problems clearly, how to give models useful context, how to spot when an output is subtly wrong versus when it’s genuinely good, and how to build on outputs rather than accepting first drafts. None of this is particularly difficult to learn. But it does require time and intention — two things most people feel they don’t have.

There’s also the confidence problem. A lot of people who’ve been professionals for twenty or thirty years feel quietly uncomfortable admitting that they’re not sure how to use a technology that twenty-two-year-olds seem to navigate effortlessly. That discomfort discourages the kind of open, exploratory engagement that builds real skill. Getting better at AI, paradoxically, means being willing to be bad at it first.

What Actually Works

The organisations that are genuinely seeing returns from AI investment tend to share a few characteristics. They treat AI literacy as a proper training investment, not an afterthought. They identify internal champions — people who are enthusiastic and skilled — and give them time and resource to help others. They measure specific, workflow-level outcomes rather than vague productivity metrics. And they build in space for experimentation, accepting that some AI projects will fail and that failure is part of the learning process.

At the individual level, the pattern is similar. The people making the most genuine progress with getting better at AI are those who’ve committed to using these tools for a specific category of work — research, drafting, data analysis, coding — and doing it consistently enough to build real intuition. Not dabbling. Committing.

That’s a boring answer. It doesn’t have the urgency of a new model announcement or the excitement of a funding round. But it’s the honest one. Getting better at AI isn’t going to happen through osmosis, and it’s not going to arrive with the next version of the software. It requires the same thing that getting better at anything requires: deliberate, sustained practice with good feedback.

The Stakes Are Higher Than They Look

This might all sound like a soft, human-interest angle on what’s really a technology story. But the implications are significant. Companies that genuinely close the AI proficiency gap — that build workforces that can use these tools at something close to their actual potential — are going to have a real, durable advantage over those that don’t. That’s not a speculative claim. It’s already visible in early productivity research, which reportedly finds that the distribution of AI benefit is heavily skewed toward people who engage deeply rather than casually.

Getting better at AI as an organisation isn’t just a training exercise — it’s a strategic imperative. The tools are powerful. They’re only going to get more so. But power that goes unused — or used badly — doesn’t translate into competitive advantage. The next frontier in AI isn’t another breakthrough model. It’s the much harder, much less glamorous work of actually teaching people how to use the ones we already have.

Source: Financial Times

Frequently Asked Questions