The fear of being late is making leaders buy the wrong things

Vague worry about AI turns into a specific deadline, and a deadline rewards any action that can be reported quickly. Signing a purchase feels like progress on the day it happens. IBM's 2025 survey of two thousand CEOs across thirty-three countries shows how common this is: sixty-four percent said the risk of falling behind drives them to invest in some technologies before they understand the value, yet only thirty-seven percent believe it is better to be fast and wrong than right and slow. Half said the pace of recent investment had left them with disconnected, piecemeal technology. Only twenty-five percent of their AI initiatives had delivered the expected return, and sixteen percent had scaled across the enterprise.

Anxiety tends to produce four kinds of purchase, all bought before anyone names the decision they are meant to improve. A platform before a use case, on the assumption that use cases will follow. Licences for everyone, which are easy to report and stall because nobody has a named task. An agent that turns out to be a relabelled chatbot or RPA product: Gartner estimates only about 130 of the thousands of vendors selling agentic AI offer real agentic capability, and forecasts more than forty percent of agentic projects will be cancelled by the end of 2027. And a pilot sized to be announced, which succeeds as a slide and teaches the organisation nothing.

Being late is cheaper than it feels. Model access prices have fallen steeply since 2023, late buyers get more mature products, and they can learn from the early buyers' cancellations. Buying wrong costs more, and most of that cost never shows up on the invoice: renewals nobody owns, integration work to unwind, data sitting on vendor terms nobody read, and credibility lost with staff who were told this tool would change their work. The real cost of lateness is somewhere else entirely. Clean data, team skills and governance that moves quickly take a year or more to build, and none of them can be bought at the last minute.

The protection is an order of operations, followed in sequence. Name the decision the purchase should improve, who makes it today, and the number that would show it got better. Check readiness across data, skills, process ownership and leadership attention. Check the data specifically: that it exists, has an owner, and can leave the vendor on your terms, which is where most failed purchases actually fail. Run a pilot with its exit written in advance. Then sign a contract you can leave, with a short term, a clear exit, data returned in usable form, and no price protection that only works if you stay. The first four steps happen before any money moves and add roughly a month.

This quarter, hold any pending AI purchase until its decision and success metric are written down. Give the board the sequence with dates rather than a product name or a licence count. For tools already bought, find the renewal dates and name an owner for each renewal decision. The exception worth taking seriously is where an early mover gains something a follower cannot buy later, such as exclusive data or network effects. Those cases exist, and they are rarer than the deadline makes them feel.