I'm knackered. AI browsers, Copilot features everywhere, AI note takers, AI in my terminal, AI in my search. It's exhausting. And the worst part? Most of it isn't helping.

The current trend of sprinkling AI into every product isn't going to help anybody. The issue isn't that AI is bad - it's that tacking "Copilots" onto everything with complex new UX patterns makes tools harder to understand and use, not easier. And without critical knowledge of the tools we're using, we end up with lots of people flailing about, failing to understand or use things properly.

Competence still matters

Microsoft recently launched "vibe working" with Agent Mode in Excel¹. The feature generates complete spreadsheets and financial models from natural language prompts, handling everything from loan calculators to P&L statements. According to Microsoft's own benchmarks, it achieves 57.2% accuracy on the SpreadsheetBench test suite¹.

Consider what that means. Financial models that are wrong nearly half the time. If you lack the Excel knowledge to audit what's being generated, you won't catch the errors.

This isn't hypothetical. The JPMorgan Chase "London Whale" incident in 2012 involved a risk model built using Excel spreadsheets requiring manual copying and pasting of data². An operational error in the calculation - dividing by sum instead of average - meant the model showed trades as half as risky as they actually were. The result: a $6.2bn loss and $920m in fines³.

Or consider the Reinhart-Rogoff economics paper that influenced austerity policies across multiple countries after the 2008 financial crisis⁴. A basic Excel formula error that excluded five countries from calculations transformed their finding from a 2.2% average GDP growth rate to a -0.1% decline⁵. This error provided justification for austerity measures that pushed unemployment above 10% in the eurozone⁶.

The pattern is clear: when AI helps people create things they don't understand, the results can be catastrophic. This applies equally to programming. Developers accepting AI-generated code without understanding what it does, how it handles edge cases or its performance characteristics will build systems that work in demos but fail in production.

When AI actually works

AI can be extremely useful. The successful implementations share common characteristics.

The clearest examples come from medical imaging, where AI triages patients by identifying scans likely to contain critical findings. One study showed a system detecting intracranial haemorrhage on brain CT scans flagged 94 of 347 routine cases as emergencies, with 60 true positives and reducing reporting time from 8.5 hours to 19 minutes⁷.

These implementations work because the problem is specific - not "make Excel better" but "reduce time radiologists spend on initial CT scan review". Humans stay in the loop - the radiologist makes the diagnosis, the AI helps prioritise. Success is measurable. It fits existing workflow. And crucially, medical imaging specialists built these systems with deep domain knowledge.

This is proper engineering applied to AI. Find a specific bottleneck where pattern recognition genuinely helps, build a system that fits existing workflows, keep humans in the loop and make the impact measurable.

The case for AI everywhere

To be fair, there are arguments for widespread AI deployment. Recent surveys show workers using generative AI save 5.4% of their work hours⁷. Economic projections estimate AI will increase GDP by 1.5% by 2035⁸. Studies have found AI tools improve performance by 66% on average, with the biggest gains for less-skilled users⁹.

The theoretical benefits include network effects (more users improve the models), democratisation of capability (helping less-skilled workers) and discovery through experimentation (you can't know what works without trying).

But targeting a random percentage increase in "productivity" is fundamentally misguided. You can't just declare your organisation "AI native" or "AI first" and expect meaningful results. AI can't possibly know the ins and outs of your business, which is built entirely upon human behaviour, communication, relationships, networks and insights.

More comprehensive research examining a broad range of employees and tasks found that despite heavy investment, AI chatbots had modest impact, saving just 2.8% of work hours¹⁰. More critically, two-thirds of that saved time was redeployed to non-value-adding activities¹⁰. The productivity gains rarely led to higher pay.

You need to deeply understand how your employees actually work before embarking on AI implementations. Not broad strokes about "productivity", but specific workflows where you can target improvement, measure whether it holds and then iterate. If it works, choose the next low-hanging fruit. That's engineering. The rest is theatre.

The "democratisation" argument has a darker side. Yes, productivity gains for lower-skilled workers sound appealing. But are they actually learning anything? Probably not. They're not forming a deep understanding of the human behaviour, communication, relationships, networks and insights that your business and customers operate in. When you close the gap between skilled and unskilled workers by reducing the need for foundational competence, you're not helping people develop expertise - you're creating dependency on systems that hallucinate and make errors. Microsoft has received so much feedback about Copilot being intrusive that they had to add a disable button¹¹ - users recognise when "help" undermines their ability to work effectively.

What actually needs to happen

Here's the hilarious part: these focused AI systems have existed for years. Quality control detection in manufacturing production lines has been using computer vision and machine learning since well before the current AI hype cycle¹². Credit card fraud detection systems have been using machine learning algorithms to identify suspicious transactions in real-time for decades¹³. These aren't new applications - they're proven systems that work because they solve specific, well-defined problems.

Information and APIs aren't ordered enough to be useful unless they're built to consider actual human use cases. AI can't supplement poor information architecture. It can augment decision-making and influence decisions, but the human element needs to be there - deciding how systems should function, understanding competitive advantages and disadvantages, knowing when outputs are wrong.

The companies that will succeed with AI aren't the ones throwing it at everything. They're the ones who deeply understand a specific industry and identify a key bottleneck, build focused solutions that solve that specific problem, integrate AI where ambiguity doesn't completely suck, keep systems legible - you can see what the AI is doing and intervene when needed - and make the AI's contribution measurable and improvable.

AI can fill a gap really well where pattern recognition or handling ambiguity provides value. But it's not about cramming chatbots or complex new workflows down users' throats. It's about supplementing existing processes where it makes sense.

The throwing-AI-at-everything-and-seeing-what-sticks approach is ill-advised. It leads to half-baked features that don't provide real function and ultimately produce poor outcomes for users. Worse, it contaminates the AI label itself. When users encounter broken AI features repeatedly, they stop trusting AI capabilities even where they'd genuinely help.

Find a specific bottleneck and solve it

I'm not trying to dump on every AI solution here. Some will be brilliant at really specific things. But they're going to be built by people who deeply understand an industry and have identified a genuine bottleneck that AI can solve.

Current AI systems still can't include clinical data or compare to prior imaging in medical contexts¹⁴. They operate in isolation, unable to integrate seamlessly into multidisciplinary care¹⁴. They hallucinate, generate wrong responses and provide misleading information at rates between 3% and 10%¹⁵. These aren't problems that get solved by deploying everywhere and hoping for the best.

The real opportunity for AI isn't in ubiquity. It's in precision. It's in understanding exactly where these systems add value and building implementations that are specific to a problem, legible in their operation, integrated into existing workflows, measured for actual impact and designed with human oversight.

When you put AI everywhere, you make it less valuable where it could genuinely help. You train users to distrust AI outputs and you create a landscape where useful implementations get drowned out by AI-washing.

AI can be transformative when it's done right. But "done right" means focused, measured and integrated solutions built by people who understand both the technology and the domain. Not copilots bolted onto every product whether they make sense or not.

Once we have these focused systems clearly defined and working, and only then, might we see some evolution towards the ubiquity everyone's chasing. But we're not there yet. Not even close.

---

References