An insurance company I work with automated 60% of its claims processing last year. No chatbots, no generative models, no transformer architecture. They taught a computer to read forms and check boxes. ROI: 400% in year one.

That company is in the 20%. Most aren't.

Last year, companies worldwide spent $154 billion on AI — more than the entire GDP of New Zealand. According to BCG, 74% of them have yet to see any tangible value from the investment. RAND Corporation puts the project failure rate even higher: by some estimates, over 80% of AI projects fail outright. That's double the failure rate of traditional IT projects.

These aren't moonshots. These are practical business applications with supposedly clear ROI — and they're cratering at a rate that should embarrass every board that signed off on them.

This isn't a technology problem. GPT-4 works fine. Claude delivers. The models do what they promise. This is a management execution problem, and in Britain it has its own special flavour: teams buy tools before fixing process, then hit GDPR, FCA, and legacy-system reality at rollout. Different accent, same failure mode.

What the 20% know that you don't

IBM says 42% of companies have "deployed" AI. Another 40% are "exploring." Sounds impressive until you look at what "deployed" actually means. Deloitte found that two-thirds of companies expect fewer than 30% of their GenAI pilots to reach full scale in the next six months. Companies are running demos that never reach customers.

The minority who succeed aren't using better models or secret techniques. They've worked out something that should be obvious but apparently isn't: AI is 10% model, 20% data plumbing, and 70% getting humans to change how they work.

Most companies do the exact opposite. They spend 70% of their time on model selection and tuning, 20% on data if they're lucky, and treat adoption as an afterthought. Then they're shocked when nobody uses their amazing system.

Research by Brynjolfsson, Li, and Raymond showed this clearly. They studied AI assistants in customer support and found an average 14% productivity gain — good but not earth-shattering. The surprise was where the gains landed. Novice workers saw massive improvements, suddenly performing like employees with years more experience. AI didn't replace expertise. It democratised it.

Your newest hire performing like a veteran. Your B-players performing like A-players. That's practical value you can measure in pounds. But only if someone actually uses the thing — which brings us to why most don't.

The five ways smart companies fail at AI

I've watched dozens of implementations crater. Smart people make the same five mistakes, in the same order, every time.

They choose the wrong problems

Tuesday morning. The CEO has just read about AutoGPT in the Financial Times. By Tuesday afternoon, there's a mandate: "We need an AI strategy." Six months and £3 million later, you've built a customer service chatbot for a company whose real problem is that fulfilment takes three weeks.

A financial services firm I know burned £2 million on an AI-powered financial adviser. Impressive demos, genuinely good technology. One problem: their clients didn't want AI advice. They wanted their human advisers to return calls faster.

The best implementations start with a brutally simple question: what business process causes us the most pain and has measurable outcomes? Not "what would be cool?" Not "what would make a good press release?" What actually hurts?

They live in data fantasy land

Every AI pitch contains the same beautiful lie: "We'll clean up the data as part of the implementation."

No. You won't. You never do.

Here's a test. Pick your most important business metric. Can you pull clean data for the last quarter in under an hour? Would finance, operations, and sales pull the same number? If you just laughed bitterly, you're not ready for AI. You're ready for a data infrastructure project.

A retail chain showed me their "AI readiness assessment." It was gorgeous. Their actual inventory data was spread across seven systems, three of which were Excel files on someone's desktop. The last time all seven systems agreed on anything was never.

No data product, no AI product. This is physics, not philosophy.

They start with tech instead of humans

Most AI projects begin the same way. An engineer builds something impressive. Executives watch a demo. Everyone applauds. Six months later, it's gathering digital dust.

Why? Because nobody asked who would use it every day, how it fits their actual workflow, what happens when it gets things wrong (and it will get things wrong), whether it'll survive the security audit, or whether the least technical person on the team can figure it out.

A manufacturing company built an AI quality control system — state-of-the-art computer vision, 94% accuracy. It failed completely. The factory workers had to log into three different systems to use it. The old paper checklist took 30 seconds.

Starting with technology instead of users is like designing a restaurant by starting with the kitchen equipment. You might end up with a brilliant kitchen. Shame nobody can find the door.

They forget about reality

Your prototype is beautiful. It runs flawlessly on clean data with direct API access and no security requirements. Congratulations — you've built something that works perfectly in a universe that doesn't exist.

I know a healthcare company that built an AI diagnosis assistant. Brilliant technology, genuinely helpful. Time from prototype to production? Three years. The AI development took three months. The other 33 months went on enterprise security requirements, compliance approvals that needed seventeen signatures, integration with systems older than the junior developers, and data governance policies written by someone who'd never seen a neural network.

Production is where prototypes go to discover what they're actually made of.

They overestimate what AI can actually do

Just because ChatGPT can write Shakespeare doesn't mean it should run your supply chain.

Companies read breathless articles about AGI and assume current AI can handle any business problem. It's like seeing a Tesla and assuming all cars can drive themselves.

The implementations that work are narrow, focused, and pragmatic. Reduce invoice processing time by 40%. Flag 90% of fraudulent transactions. Answer 60% of support tickets without human intervention. Not "transform our business model with AI" or "become an AI-first company." Those are strategies. Implementations need specifics.

The boring winners

Want to know where AI actually makes money? Document processing. Quality control. Fraud detection. Support ticket routing. Invoice matching.

Not sexy. Not going to get you on the cover of Wired. Just profitable.

That insurance company I mentioned at the top? Basic document extraction. They didn't need a large language model. They needed a computer that could read forms reliably and a team that was willing to redesign the process around it. The technology was the easy part. The hard part was getting 200 claims handlers to change how they'd worked for a decade.

The best AI strategist I know has a simple rule: three things with AI, everything else the old way. Because AI excellence requires focus. Better to be spectacular at three critical processes than mediocre at thirty.

Before you approve another AI project

Answer these honestly. Is this problem going to exist, largely unchanged, for the next 12 months? Process instability kills AI faster than bad data. Can you access clean, relevant data today — not tomorrow, not once you implement the new system? Is there a specific person whose bonus depends on this working? ("The team" is not a person.) Can you measure success in business terms within two quarters — not "improved accuracy" but "reduced costs by £2 million"? Do you have a believable path from prototype to production that includes security, compliance, and integration?

More than two "no" answers means you should stop and fix the foundations first.

And if your pilot doesn't have success metrics, an owner with skin in the game, and a path to scale, it's not a pilot. It's procrastination with a budget. I've seen companies with 50-plus pilots and zero production systems. Everyone knows within 90 days whether a pilot is going to work. Act on that knowledge.

The uncomfortable truth

Vendors won't tell you this: poorly implemented AI makes things worse. It creates shadow processes, doubles work, and erodes trust. I watched a company implement AI-generated reports that nobody trusted. The result was that every AI report required a human-generated verification report alongside it. They doubled their reporting workload.

The 20% of companies succeeding with AI aren't doing anything glamorous. They pick focused problems, fix their data, redesign processes around people, manage change properly, and measure what matters.

The question isn't whether AI can deliver value. The evidence is overwhelming that it can. The question is whether your organisation can deliver AI.

AI strategy without execution discipline is expensive software tourism.