For a decade it was the safest advice in the economy. Governments funded coding bootcamps. Schools rewrote curricula. Parents repeated it at dinner tables: learn to code, and the future will take care of itself.

​

Then the future arrived, and it codes.

​

The numbers are stark. Software development job postings sit more than a third below their pre-pandemic levels. Employment among 22–25-year-olds in the most AI-exposed roles has fallen 16%,  with young developers down nearly a fifth since 2024 alone. At the largest technology companies, recent graduates now make up just 7% of new hires, down from over 9% two years ago. Microsoft's chief executive has confirmed that 30% of the company's code is now written by AI, while software engineers made up more than 40% of its most recent major round of layoffs.

​

​

The wrong variable

The lesson is not that coding was a bad bet. It is that an entire economy assessed the wrong variable. The question everyone asked was "is this skill in demand?" The question that mattered was "what does this work actually consist of?"

​

Coding was hit first for structural reasons, not cyclical ones. The work is purely digital. It has verifiable right answers, code either runs or it doesn't, which makes it ideal training ground for machines. And it carries no physical asset, no licence, no client relationship that a substitute would have to rebuild.

​

Any work sharing those three properties is next in the queue, regardless of how in-demand it looks today. Which means the useful tool is not a list of safe sectors. It is a filter you can run any business through.

​

The Four Questions

 

01 Atoms or bits?

Does delivering the value require physical presence, physical assets, or physical trust? AI can schedule care work, price a logistics run and market a restaurant, but it cannot lift a patient, drive the last mile in the rain, or cook the meal. Businesses whose value is delivered in atoms have a moat that improves as everything around them digitises.

​

02 Is the demand legislated or fashionable?

Compliance obligations, energy-transition mandates, an ageing population's need for care: demand written into law or driven by demographics does not disappear when a model improves. Demand driven by trend, the fastest-growing UK small businesses by insurance applications are currently Reformer Pilates studios, cleaning services and cake makers,  can be real and still be low-moat. Growth is not the same as durability.

​

03 Do you own the accountability, or just the output?

Clients pay AI for output. They pay humans to be responsible when the output is wrong. Audit, legal advice, security, clinical judgement: in each case the signature is the product, and the signature cannot be automated because its entire value is that a person carries the consequences. If your business sells deliverables, ask what fraction of your fee is really paying for the deliverable and what fraction is paying for someone to blame.

​

04 Does AI cut your costs, or your prices?

This is the decisive question. If AI makes your inputs cheaper, your admin, marketing, research, first drafts, your margins widen and you win. If AI makes your output cheaper, if the thing you sell is the thing the machine now produces, then you are not the user of the tool. You are the input being substituted. Many businesses will discover which side they are on only when a client asks why the invoice hasn't fallen.

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​

AI-proof is not AI-free

One final distinction, because it is where the next decade's winners and losers divide. The best-positioned businesses will run AI aggressively on the cost line while selling something it cannot deliver, human costs falling, human-only value intact. The losers will do the opposite without noticing: human cost structures, machine-priced output.

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​

​