Your project list says the quote reminder shipped weeks ago. The board is green. The status says done. The budget meeting wants a number for what it saved.
Here is the problem. Done is a label your team typed into a tracker. It is not proof that a customer ever felt the change. One record settles the argument, and your team probably never opened it.
The short answer: pick one AI project marked done in the last 90 days and find the record it should have left behind. A working reminder leaves sent mail with customer names on it. A working dashboard leaves a decision somebody made from it. No record means no result, whatever the board says.
- What the send history said
- Three records every finished project leaves
- Run the audit in one sitting
- Three verdicts, and no fourth option
- Budget season is why this cannot wait
What the send history said
Picture a small service company at budget time. The owner needs one number for what the last three months of AI work saved. The team points to three finished projects on the board.
One project moved some customers to email billing. One was a dashboard, designed across five meetings. One was a quote reminder, built to chase late quotes by email.
Before the owner counted the reminder, someone opened its send history. Three emails had gone out in total. All three went to coworkers who were testing the tool. The customer reminders were still switched off.
Three records every finished project leaves
Every finished project leaves a trace somewhere outside the project board. A sent email leaves a recipient list with real names. A cheaper process leaves one bill that got smaller. A used dashboard leaves a decision someone can name.
Ask three questions, in this order. Who received the email the tool was built to send? Which bill got smaller after the project shipped? What decision did someone make using the dashboard?
A project with a trace answers at least one question fast. A project without one answers none of them. The gap between those two answers is your real AI result for the quarter.
Vendor numbers skip this step on purpose. A claimed saving sounds better before anyone subtracts the checking time, which is the point made in Beacon’s 9.6 AI hours only count after you subtract review time. Your own records carry no such incentive. They just sit there, waiting to be opened.
Run the audit in one sitting
Pick the project your team is proudest of. Pride is the bias you most need to test. Then open the record that project should have left behind.
Read the recipient list, not the project page. Count sends to customers, not sends to the team. Check the setting that turns the feature on for real accounts. A tool tested by staff and never switched on has saved your customers nothing.
Do the same for the dashboard. Ask the named user for one decision it changed. If nobody can name one, the five meetings bought a screen, not a result. Write the answer down in one line, either way.
Repeat for every project marked done this quarter. The full pass takes one sitting for a small firm. It needs no new software and no consultant. It needs the send history, one bill, and one honest answer.
This is also the check that separates licences from use. Seats on a plan prove a purchase, nothing more, as Your team has AI licences. AI fluency is what turns them into saved hours argues from the skills side. The audit makes the same point from the records side. Usage leaves traces. Shelfware leaves invoices.
Three verdicts, and no fourth option
Every audited project lands in one of three buckets. It earned its keep, and the record proves it. It works but was never switched on, and one setting fixes it. Or it produced nothing, and the done label comes off today.
Fund the first bucket without apology. Fix the second bucket this week, then recheck the record in 30 days. Move the third bucket out of the savings column before the budget locks.
Notice what is missing from that list. There is no bucket for “probably working.” Probably is how a dead reminder survives three budget cycles. The record exists or it does not.
The habit pairs well with a written check on the work itself. Review AI-generated work with one signed line before your name goes on it covers the output before it leaves. This audit covers the project after it ships. Run both and the board starts telling the truth.
Budget season is why this cannot wait
Canadian firms are setting next year’s software budgets right now. Renewal dates do not ask whether the tool ever ran. They just charge the card again.
Outside money will not rescue a weak audit either. Public programs spread thin by design, as The Chamber AI Adoption Program averages $270 a business. Only 540 get hands-on cohorts shows in plain figures. Your own records are the only audit that is free, fast and already done.
In the service company example, one project of the three carried a figure worth taking to the boss. The other two were not failures of effort. They were failures of the last switch, and of nobody checking. That is the normal shape of a first AI year, and it is fixable in an afternoon.
So run the check before the renewal lands. One short prompt is all the method needs, and the version worth copying takes minutes on your own records. Your one action this week: pick the project your team calls finished, open its send history, and bring the one-line answer to the budget table.
