I have spent more than a decade building AI for defense. That experience taught me one lesson businesses keep ignoring. The most advanced AI is rarely the most cost-effective. Yet companies are pouring money into the newest, most complex models when simpler tools would do the job. Most of us will never sign an AI contract, but our pensions and savings are funding these choices. If we want to deflate the AI bubble before it bursts, that has to change.
The scale is hard to ignore. Data from the Kobeissi Letter shows that AI-linked stocks now account for 45 percent of the S&P 500’s total market cap. This share has risen by 20 percentage points since late 2022. One Los Angeles County pension fund estimates that 8-19 percent of its portfolio may be exposed to AI. A considerable amount of our savings now depends on whether AI spending delivers. Anyone with a pension should be paying attention.
Many organizations now choose AI to look impressive rather than to solve a real problem. One MIT study found that 95 percent of organizations saw no measurable return on generative AI, despite $30 billion to $40 billion in enterprise spending. I don’t think that gap reflects the technology. It reflects how the technology is chosen. Too often, the decision starts with the newest tool on the market rather than the job that needs doing.
Many leaders invest in AI capacity they don't need. The most advanced models do not always mean the best results. In fact, they can be more expensive, harder to control, and less predictable than the task requires. I have heard executives dismiss simpler systems as not real AI. That is marketing, not engineering. AI spans a wide range of techniques. For example, the UK’s latest guidance on dependable AI recognizes everything from linear regression to neuromorphic computing as part of the AI landscape. These include mature rule-based systems that can do the job reliably at far lower cost. Fear of falling behind is often what drives these choices.
AI Only Should Only Be as Complicated as the Job Requires
My principle when it comes to AI is that it should be as simple as possible to do the job, not as complex as possible to look good. On an air-defense research project about a decade ago, experts wanted neural networks. But there were only around a dozen real cases to learn from. My team used simple rules with a human making the final call. This proved cheaper, faster, and more reliable. It is a lesson I have carried ever since.
The people who build AI rarely understand the jobs it supports. Developers are not traders, managers, or soldiers. In defense, I had to work hard to make my developers understand how a young officer would actually use the system. When they didn’t, the officer stopped trusting it and went back to the old way. The fix has two parts. Involve real users from the start. And make sure the people deploying AI are as good at using the AI as the ultimate users need to be.
When organizations match AI to the task rather than the trend, the rewards are considerable. Businesses spend less, get dependable results, and stay resilient enough for the next shock. Employees use tools built around how they work. Customers get service that works, with a route to a real person when required. And for savers, the AI bubble will deflate gradually rather than burst, protecting hard-earned money.
Staying on this path carries real risk. I was working in strategy consulting when the dot-com bubble burst and saw firsthand how quickly crowd-driven investment unraveled. The 2008 crisis followed a similar pattern. A bubble that deflates can be managed, but one that bursts hurts people who had no say. Workers, customers, and pension savers end up paying for choices made in boardrooms.
Business leaders must ask whether their AI is as simple as it needs to be or as complex as it can be. They should judge it by results, not by how advanced it sounds. Employers need to involve end users from day one. Pension trustees should ask how the companies they fund choose their technology. The best AI is a system people can understand, trust, and challenge when it matters.
Dr. Darrell J. Jaya-Ratnam leads DIEManalytics, which advises governments and companies on operational analysis and AI, and is the author of a book on defense AI. All opinions are the author’s own.