The AI Experiment: What We're Learning From the Quiet Retreat

Something interesting is happening with AI adoption. Small businesses in the UK went from 42% adoption to 28% in just over a year. Not because AI stopped working, but possibly because something about the approach wasn't quite right.

The data tells a curious story. While AI spending increased six-fold—from £2.3 billion to £13.8 billion—project abandonment rates jumped from 17% to 42%. More investment, fewer successes. It's a pattern worth understanding.

Perhaps the issue wasn't the technology itself, but how we framed it. Implementation suggests certainty—a clear path from A to B. But AI rarely works that way. The organisations seeing results seem to treat their work differently. They're running experiments, not implementations.

CAST, a charity focused on social impact, maintains a library of AI experiments. The language matters. An experiment can succeed by teaching you something, even if it doesn't deliver the result you expected. An implementation either works or fails.

The businesses quietly succeeding with AI share some common patterns. They start with a specific problem, not a general ambition. They work in small increments—proving value before scaling. They accept that data will be imperfect and plan accordingly. And crucially, they build human oversight into the process from the start, not as an afterthought.

There's also a cost reality emerging. What begins as a £1,200 annual subscription often becomes £3,500-4,000 once integration, training, and maintenance are factored in. The organisations that budget for this reality fare better than those surprised by it.

The retreat from 42% to 28% might not be a failure of AI. It might be a necessary recalibration. A shift from "we must do AI" to "here's a specific experiment worth trying."

If there's a lesson in the data, it's this: AI works best when treated as a tool for specific problems, not a general solution for everything. The question isn't "should we use AI?" but rather "what's worth experimenting with, and what would we learn either way?"

That shift in thinking—from implementation to experiment—seems to make all the difference.

---

Seven things to avoid

Looking across the projects that stalled or were abandoned, some patterns emerge:

Starting too broad. "We need an AI strategy" is harder to test than "can AI help with invoice processing?"

Underestimating the real costs. Software licences are typically 30-40% of total spend. Integration, training, and ongoing maintenance make up the rest.

Expecting perfect data. Waiting for clean, structured data before starting often means never starting. The organisations making progress work with what they have.

Treating it as set-and-forget. AI tools need adjustment, monitoring, and refinement. The initial setup is just the beginning.

Skipping the human layer. When AI makes mistakes—and it will—someone needs to catch them. Building oversight in from the start costs less than adding it later.

Forcing adoption. Tools that teams don't trust or understand tend to be quietly abandoned, regardless of how much was spent on them.

Measuring the wrong things. "Are we using AI?" matters less than "is this solving the problem we identified?"

The businesses seeing results tend to avoid these patterns by keeping experiments small, specific, and honestly evaluated. Not every experiment needs to succeed. But every experiment should teach you something worth knowing.

---

*Cost estimates based on "The Cost of Implementing AI in a Business: A Comprehensive Analysis" (Walturn) and UK SME implementation data showing mid-sized AI projects typically range £80K-£400K, with 70% abandonment rate before completion.