What This Year's AI Budget Is Actually For

The companies already ahead with AI did the unglamorous work first. If yours is behind, this is the year to catch up, and it will not pay for itself yet.

When American factories first switched to electricity, most of them got nothing for it. A plant owner in the 1890s would pull out his central steam engine, set an electric dynamo in the same spot, and wait for the improvement. It never came. Electricity took two or three decades to register in productivity figures, and closer to forty years to deliver its full effect. The economist Paul David worked out why. Those early factories had wired electricity into a building designed for steam, where a single engine drove the shafts and belts that ran every machine at once. The gains arrived only when a later generation threw out that layout and rebuilt the factory around what electricity made possible: a small motor on each machine, and a floor arranged around the work rather than the power source. The technology was ready in 1900. The companies needed another generation to catch up to it.

The same adjustment is under way with AI, and you can already tell who is ahead of it.

A handful of companies spent the last two years on the unglamorous part. They cleaned up their data. They pushed their people past the novelty stage into real daily use. They took a few processes that mattered and rebuilt them so AI could do genuine work inside them. Those companies get more out of today's AI than their rivals do, and they will get more out of next year's too. The platform they chose was never what set them apart, since anyone can buy the same one. The work around it did.

The rest of this is for everyone else, the companies standing roughly where those factory owners stood in 1898. If you have not built that foundation yet, this is the year to build it, and you should begin knowing you will not see a dollar return on it any time soon.

There is a sound reason to spend anyway.

The technology will keep getting cheaper and better with no help from you. OpenAI, Anthropic, and Google are spending fortunes to see to it, none of it on your budget. The one thing they cannot sell you is a company that knows how to use what they ship. That has to be built inside your own walls, slowly, and the time you lose now cannot be bought back later. Treat this year's AI money as a down payment. It buys little productivity today. It buys the ability to use the far more capable AI already on the way.

The labs will tell you the same thing

Which is odd, given their incentives. At Davos in January, Satya Nadella warned that AI could turn into a bubble if its benefits stayed locked inside the tech industry. "For this not to be a bubble by definition, it requires that the benefits of this are much more evenly spread," he said, arguing that the technology would only "bend the productivity curve" once ordinary companies, not just the labs, learned to use it. Sam Altman made the same case from the opposite end: "Companies that are not set up to quickly adopt AI workers will be at a huge disadvantage," and getting there would take "a lot of work and some risk." Neither man is describing a problem with the technology. They are describing a problem with the buyer.

This pattern already has a name

If that still sounds like a story invented to soothe the CFO, economists have been documenting it for decades. Erik Brynjolfsson and his colleagues call it the Productivity J-curve. When a technology this broad arrives, measured productivity falls before it rises, because companies must first spend on things that do not yet count as output, like retraining people and reworking how a job gets done. That spending drags the numbers down before any of it pays off. Electricity followed the curve. So did computers. AI is following it now. The dip is real, and it is supposed to be there. The winners are the companies that spend the dip building something.

Why chasing a return this year backfires

Judge this year's AI spending by this year's savings and you will push your teams toward the small, safe wins. Summarize the meeting. Draft the email. Shave a few minutes off a chore nobody enjoyed. Each one yields a tidy figure for the finance review and nothing beneath it. You pay for the dip and never climb out of it.

Two weeks ago I wrote that a high AI bill is fine, so long as you can show what it bought. This is the other half of that argument. For a company starting from behind, what this year's money buys is capability, and you have to measure it as capability, on purpose, or you will never separate real foundation work from expensive noise. The divide is already visible in the results. BCG's latest research places only about 5% of companies in the group with genuine AI capability. Where those companies apply it, they are already growing faster and cutting costs deeper than everyone else. The distance between them and the rest is widening.

What the foundation actually is

Foundation is a soft word, so here is the hard version. If you are behind, this is the work.

• Fix your own knowledge first. The technology is becoming a commodity. Your data, and the knowledge carried around in your employees' heads, are not. Get it cleaned up, organized, and reachable, so the AI can actually draw on it. Nadella's blunter formulation: a company either builds its hard-won knowledge into systems it controls, or watches that value bleed out to whoever will.
• Put the tools in everyone's hands, not a pilot team's. Anyone whose job AI might touch should be using it on real work until they develop an instinct for where it shines and where it falls apart. Ethan Mollick's research keeps reaching the same conclusion: the limit is human skill, not the technology, and the surest sign a company will adopt AI is whether its leaders use it themselves.
• Put a few people in charge of trying everything. A small team that tests each new release as it lands, runs real experiments, decides what to build and what to buy, and turns the best discoveries into workflows the rest of the company can follow. This is what lets you absorb next spring's release in a couple of weeks rather than argue about it for two quarters.
• Rebuild the work, don't decorate it. Take two or three processes that matter and redesign them around AI from the ground up, the way the electrified factories eventually did. Lay a chatbot over a broken process and all you own is a faster broken process.
• Set the guardrails and keep an honest scorecard. Spending limits and access rules, plus a real count of what you are accumulating: the people who can now use the tools, the processes you have rebuilt, the experiments you have run. Measure it deliberately. The point was never to stop measuring, only to measure what matters this year.

The clock is already running

None of this can be rushed, which is exactly why you start now. A far more capable AI will land in the next several months, as one did this year and the year before. When it does, the companies that spent this year on the foundation will have it working almost at once. The ones still waiting for proof will be starting the job their competitors started a year ago, against a target that keeps moving. The labs will hold up their end. The rest is yours, and it takes longer than a fiscal year.