As AI spending grows, so does the pressure to prove it’s worth it. The danger isn’t that AI has no value — it’s that we measure the easy things and convince ourselves of returns that aren’t really there.

The vanity metrics trap

It’s tempting to count usage: prompts run, hours “saved”, documents generated. These feel like progress but rarely touch the bottom line. Time saved isn’t value unless it’s actually redeployed to something useful; an hour freed and then absorbed by other busywork is an hour that changed nothing on the books.

Measuring what matters

Honest measurement is harder but more useful:

• Tie it to a real outcome — faster cycle times, fewer errors, higher conversion, lower cost to serve.
• Set a baseline first — you can’t claim improvement you didn’t measure before.
• Count the full cost — licences, integration, oversight and maintenance, not just the subscription.
• Be willing to find zero — some pilots won’t pay off, and knowing that is valuable too.

The pragmatic takeaway

Measure AI the way you’d measure any investment: against outcomes, with a baseline, and with the full cost in view. The goal isn’t to justify the spend — it’s to learn what’s actually working, and do more of that.