Time saved.
Once the work is done.

A clear before / after measurement

Choose a unit of work: a case processed, a request classified or a document accepted. Observe the usual workflow, then the assisted workflow on comparable cases. Keep the exceptions; removing them from the calculation flatters the result without improving the service.

Record active human time and total elapsed time separately. A machine waiting in the background does not have the same effect as a colleague stuck in front of an answer. Also note the operations that have to be started over.

An example of a calculation, not a promise of gain

In this example, the gain is twenty minutes, not thirty-five. Installation, subscription and maintenance costs are not yet included. If the assisted cases contain an additional significant error, its consequences must also be assessed.

A small set of cases that looks like the real work

Include clean, incomplete, contradictory and rare inputs. Keep what is used to improve the system separate from what is later used to compare it. Otherwise, the result may only show that it recognizes the examples already used.

Decide what must be exact: an amount, a reference or a recipient is not judged like a style preference. Have doubtful cases reviewed by someone who knows the business. Keep the size and limits of the sample with the conclusion.

Decide after the pilot

Continuing can mean increasing the volume, but also fixing the integration or simplifying the task given to the model. No gain on the scope tested is useful information. It avoids rolling out further a solution that shifts the workload instead of reducing it.