The two numbers
Hours returned. How many hours a week the task took before, and how many it takes now, measured the same way both times. This is the number that survives an argument, because the people doing the work can confirm it.
What those hours cost. The loaded cost of the people who did them, or the value of what they do instead. Hours times cost is the return. Everything else is a story about the return.
The three that do not matter yet
Tokens, calls, or minutes of model time. These are the cost of the tool, and they are usually small. Tracking them first is like tracking the electricity a dishwasher uses before asking whether anyone still washes up by hand.
Accuracy on a benchmark. A model's score on someone else's test says nothing about your task. Measure your task.
Adoption. Logins and prompts sent tell you people tried it. They do not tell you the work changed.
What to measure before you build
The before number is the one people skip, and without it there is no return to show. For the one task you are about to automate, write down: who does it, how many times a week, how long each time as a range, and where the copies of the information live. That page is the baseline, and it is also the scope: it is what we build against and what the demo on Friday is measured against.
This is the whole reason the audit exists. It produces exactly this page, with the manual cost as a range for the two or three tasks worth fixing, before anything is built.
What to measure after
The same page, a month later, the same way. The difference is the return. Then the second-order effects, which are real but should be written down as observations rather than numbers: the invoice that was chased before anyone noticed it was late, the answer found in a minute that used to take a morning.
When an agent should keep the numbers
Once the baseline exists, the measurement itself is a task with a trigger, a rule and an output, which means it can be automated. The agent reads the run logs, counts the tasks handled, applies the baseline, and writes the return into a monthly note nobody has to remember to produce. That is a small build, and it is the one that answers the question this page started with, every month, without anyone asking.
Start with the audit. It is the page you need before either number exists.