Don't confuse the AI bubble with AI's usefulness.
Every few weeks another story asks whether AI is a bubble. Valuations look stretched, companies borrow heavily to build data centers, and commentators line up on both sides. If you run a department, a clinic or a town office, it is fair to wonder whether you should slow down the pilot you started in the spring until the dust settles.
Two questions that sound like one.
The bubble debate is about prices. It asks whether the money going into chips, power and data centers will earn back what investors paid, at the valuations they paid. That is a real question, and nobody knows the answer yet. It is also a question for people who own shares in those companies or lend them money.
Your question is narrower and easier to check. Does the tool your team uses this month take work off someone's desk, catch an error, or help you answer a request faster than before? You can find that out by watching your own work for a few weeks. No analyst can answer it for you, and no market correction changes what the tool did last Tuesday.
The two get blurred because they share a vocabulary. "AI is overhyped" can mean the stock prices are too high, or it can mean the software doesn't do much. The first claim may well be true while the second is false for a given task in a given office. Keep them apart and both become easier to think about.
What a correction would change, and what it wouldn't.
Past infrastructure booms are a useful guide here, with the usual caution about history. Plenty of the companies that built railways or laid fiber in a rush did badly, and some disappeared. The track and the cable stayed in use, and people kept finding new jobs for them long after the original investors had taken their losses. When a market overpays for capacity, the overpayment tends to fall on the people who paid it.
For an organization that uses AI tools, a correction would show up in more ordinary ways. A vendor might raise prices, cut a free tier, merge with a competitor or shut down a product. A feature you rely on might be folded into a bigger plan you don't want. These are the same risks you already manage with payroll software or a scheduling system, and they call for the same precautions.
The bubble question is about whether investors paid too much. Your question is whether the tool did the work.
The practical response is portability. Keep the instructions your team has written for AI tools in a document you own, not only inside the tool. Know how to export your data and what format it comes out in. Prefer month-to-month terms while you are still learning what a tool is worth. None of this requires a view on the market.
Judge the tool by its work, not its press.
Hype distorts decisions in both directions. When the headlines are glowing, organizations buy seats for everyone, sign annual contracts and announce programs before anyone has found a use for them. When the headlines turn sour, the same organizations cancel the one tool that was saving their front desk an hour a day, because it now feels embarrassing to defend.
Both mistakes come from judging AI as a category instead of as a set of specific tasks. "Is AI worth it?" has no useful answer. "Is it worth having the intake assistant draft first replies to new client emails?" does, and the person who reviews those drafts can usually tell you within a month.
We see the steadiest progress in teams that keep a short list of tasks where a tool has earned its place, with a name next to each one. That list doesn't care what the market did this quarter. It gets longer when a new use proves itself and shorter when one stops paying off.
Where the hype does cost you.
There are two places where the bubble talk deserves a seat at your table. The first is contracts. A long commitment to a young vendor carries more risk when its funding depends on investor mood, so ask what happens to your data and your price if the company is sold or closes. If the vendor can't answer plainly, keep the term short.
The second is attention. A team that spends its meetings debating whether AI is real has less time for the dull work of trying it on an invoice backlog or a stack of permit applications. The debate feels productive because it is about something large, but it rarely changes what anyone does on Monday.
A usefulness test for any AI tool.
Run each tool your organization pays for, or is thinking about buying, through these questions. Answer them with your own records, not the vendor's demo.
- Name one task this tool did for us in the last two weeks, and the person it helped.
- If we switched it off tomorrow, who would notice first, and what would they have to do by hand?
- Would we still pay for it if the price rose sharply at renewal, or if the free tier we rely on went away?
- Can we get our data and our written instructions out in a usable form within a day?
- How long is our commitment, and what happens to it if the vendor is bought or closes?
- Did we adopt it because of a specific task, or because of what we were reading at the time?
A tool that gives clear answers to the first two is doing real work, whatever the markets decide. One that stalls on question four is the one to fix first, since that is where a correction would actually reach you.
