How do we tell whether AI is actually saving time? Time the work before and after, and count the checking as part of the job.
Short answer. Time the work, because impressions of speed can be badly wrong. In a randomised trial, experienced developers expected AI to cut their completion time by 24%, believed afterwards that it had cut it by 20%, and were measured taking 19% longer. So pick three to five repeating tasks, time them by hand for a couple of weeks as a baseline, then time the same tasks with AI, including the minutes spent checking and fixing the output. The saving is net minutes per task times tasks per month. Use usage logs to see who is using the tool, not to prove a benefit.
01
Why not just ask people?
METR’s randomised trial involved 16 experienced developers and 246 tasks. Developers forecast that AI would reduce completion time by 24% and, after the study, estimated that it had reduced it by 20%. Measured completion time actually increased by 19%. The study concerned software developers and early-2025 tools, so it does not predict your office, but it shows how far feeling and measurement can diverge.
In a study of 5,179 customer support agents, Brynjolfsson, Li and Raymond found issues resolved per hour rose 14% on average, by 34% for novice and low-skilled workers, with minimal impact on experienced and highly skilled workers. Effects differ by task and by person, so you need your own numbers.
02
A four-week measurement plan
Choose three to five tasks that repeat: minutes of a standard meeting, a first draft of a standard letter, finding a clause across files, formatting a report.
Weeks 1 and 2: time 10 to 15 occurrences of each task done the usual way, with a stopwatch or a simple log, and note the quality.
Weeks 3 and 4: the same tasks with AI. Time from start to finished-and-checked, not to first draft.
Compare like with like: similar tasks, similar people, and ideally a comparison group who carry on as before.
Sample the outputs for quality: have a second person score a few from each period against the same checklist.
Record what the saved time was used for. Minutes that disappear into other tasks are not a cash saving.
03
What to record
Measure
How
Trap
Minutes per task
Stopwatch or log, from start to finished-and-checked
Counting only the AI’s drafting time
Rework
Minutes spent fixing and re-asking
Ignoring the cost of errors found later
Quality
A second reader scores a sample against the same checklist
Faster but worse
Use
Logs of who used which app and how often
Mistaking use for benefit
Spread
Results by person and by task
Averages hide who gained and who lost
Opinion
A short survey: where did it help, where did it get in the way
Useful for diagnosis, not for proving a saving
04
Net, not gross: a worked example
These numbers are examples, not measurements. Suppose writing minutes by hand takes 45 minutes after a one-hour meeting. With AI, a draft appears in minutes, but a person spends 15 minutes checking the action items against the transcript. The net saving is 30 minutes. Across 20 such meetings a month that is 600 minutes, or 10 hours.
If checking takes 40 minutes instead, the saving falls to 5 minutes a meeting, or 100 minutes a month. If errors slip through and cost time later, it can turn negative. That is why the checking minutes belong in the measurement.
05
What ANVE says
ANVE logs every model call with the user, the app, the model and whether it ran locally or in the cloud, which shows adoption by person and by app. It does not time your work. For meeting minutes, every action item quotes the transcript line it came from and each owner confirmation is logged, so the checking step is built in and can be timed.
ANVE has no customers yet, so any time-saving figure for your office has to come from your own pilot.
06
Questions people ask
What counts as a good result?
Set the target before the pilot: a minimum net saving per task and a quality floor. Decide in advance what you will do if either is missed.
How long should we measure?
Long enough for novelty to fade and for each task to recur many times. Four weeks is a practical minimum for most office tasks; extend it for tasks that happen monthly.
Should we measure everyone?
No. A small group of volunteers plus a comparison group gives clearer results than a thin sample across the whole office.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
