Companies are being asked what their AI spending bought, and the answer most teams give, hours saved, is the one a finance department can do least with. Proving value is one of the top barriers for the people who run data and AI work, and the way out is to tie each project to one number the business already tracks.
Saved hours are not a result
For example, take a hospital group's billing team of eight. An AI assistant now drafts the appeal letters for insurance claims that were denied, and each letter takes ten minutes instead of forty. At 60 appeals a week, that gives the team 30 hours back. Payroll has not changed, so the finance director sees nothing, because time saved is not something the accounts record.
Now say the team spends those 30 hours appealing claims it used to write off, and gets paid on ten more claims a month worth $800 each. That is $8,000 a month in cash, and it shows up in a figure the finance director already watches: the share of denied claims that are eventually paid. The tool and the hours are the same in both versions. Only the second has a result, because someone decided what the hours were for.
The computer years had the same shape
Retail and wholesale had gone for years without productivity gains. Then Walmart and McKesson used technology to change how goods moved through their distribution centers, and other companies in both industries responded until productivity rose across the board.
The payoff did not arrive when the computers were bought. It arrived when someone pointed them at a job with a number attached, and that number let everyone else see what had been gained. I expect AI to follow this pattern. The company that aims a tool at one counted job will be easy to see, and the company that gives everyone the tool and counts nothing will have a lot of use and nothing to show for it.
Budget season sorts the projects
When money is tight, a finance director has to choose between projects. A project that reports how often people used the tool loses to one that reports a figure the director already tracks. That will cut some pilots that work well and keep some that are only average but well measured.
It also explains two problems we have written about. Most pilots never reach scale, and few firms know what their AI costs. Each is missing half of a ratio: a result with no cost, or a cost with no result, cannot be compared with anything.
If you run a team, three things can be done before the next tool goes live. Write down how long the job takes now and what it costs. Name the figure that should move and the person who watches it. Then decide what the freed hours are for.