Start here · Lesson 1.4
Measuring Voice to AI: The Metrics That Matter
It is easy to track activity with voice to AI: how many times you dictated, how many minutes you talked, how many drafts you made. None of that tells you whether it is working. A day with ten dictations and ten rewrites is worse than a day with two clean ones.
This post gives you five numbers to track weekly, all measured by you with a note app or a spreadsheet. No tool tracks them for you, and I am not quoting benchmarks, because I have none. Your own trend over four weeks is the only comparison that matters.
Lesson 1.4 walks you through it step by step: Talk to AI instead of typing.
Track outcomes, not activity: time to a sendable draft, edit load, errors caught, jobs where you used voice, and whether you would repeat it.
Why activity numbers mislead
Counting dictations rewards doing more. But the goal is less effort for the same quality. A person who dictates a lot and then spends twice as long fixing text is making things worse and the counter would not show it.
Good metrics answer a plain question: is my work getting done faster or better, with no new mistakes?
The five numbers
- Time to a sendable draftMinutes from starting to speak to a draft you are willing to send. Round to the nearest five. Do not time the typed version every day, just a few times to have a reference.
- Edit loadAfter the AI cleanup, how much did you change by hand? Score it none, light or heavy. Count how many of your last ten jobs were heavy.
- Errors caughtHow many wrong words, names or numbers you found when you read the draft. A rising number can mean the setup needs work, such as the microphone.
- Voice shareOut of the jobs you could have dictated this week, how many did you? This shows whether it is a habit or a novelty.
- Would repeatA yes or no after each job. Count the yeses. This is the most honest number you will track.
Setting up a simple log
Use a note or a sheet with one row per job and these columns: date, job, minutes, edit load (none, light, heavy), errors caught, would repeat (yes or no). Fill the row right after sending, while you remember. It takes under a minute.
| Column | Example entry |
|---|---|
| Date | Tue |
| Job | Reply to supplier |
| Minutes | 10 |
| Edit load | Light |
| Errors caught | 1 (a name) |
| Would repeat | Yes |
That table row is a made up example of the format, not a real result.
How to read the numbers weekly
Look once a week, at the end of the week, for five minutes. Compare with the previous week rather than with a target.
- Minutes falling and edit load light: the routine is working. Keep it.
- Minutes flat and edit load heavy: the prompt or the aim is weak. Revisit the cleanup prompt.
- Errors caught rising: check the Microphone source and your environment. In System Settings, Keyboard, Dictation, you can choose the device Mac listens to, per Apple's guide.
- Voice share low: the shortcut may be awkward, or the jobs may not suit voice. Try another job type.
- Would repeat is mostly no: believe it. Drop voice for that job type.
Metrics that sound useful but are not
- Word count dictated. More words is not better work.
- Number of AI prompts. Many prompts for one job can signal a bad brief.
- Raw transcription accuracy as a percentage. You cannot compute it reliably by eye, and what matters is whether the final draft was right.
The one number that protects you
Errors that reached a customer. If a wrong price, name or date ever goes out, that is the metric that outweighs all time savings. Keep it at zero by reading names, numbers and dates before sending. If it is not zero, slow down before you try to speed up.
Setting a review rhythm
Put a recurring ten minute slot in your calendar, for example Friday afternoon. Read the log, answer three questions and write one change for next week:
- Which job type had the best numbers?
- Which had the worst?
- What single change will I test next week?
One change at a time is the rule. If you change the microphone, the prompt and the shortcut together, you will not know what helped. The routine in the Talk framework for voice to AI fits well with this review, since its last step is to keep what works.
Privacy and your log
Your log should hold job types and numbers, not client details. If you do want to note specifics, keep them minimal. Also remember that Apple says you can check Keyboard settings to see whether your general text dictation is processed on your device. That is useful context if privacy is one of the things you track. For a trial that uses these numbers, see a 7 day voice to AI plan.
Questions people ask
What is a good time saving to aim for?
I will not give a target, because it depends on you and the job. Compare against your own earlier weeks.
Do I need a spreadsheet?
No. A note with one line per job is enough.
How long should I track?
Several weeks is better than one. Stop once you know which jobs suit voice, then check again now and then.
Can software track this for me?
Some of it, perhaps, but I have not verified any tool for this. A manual log is more reliable for judging whether the output was good.
Keeping the measuring light
Do not let tracking become a second job. If logging takes longer than a minute per task, cut columns until it does not. The two most useful, if you keep only two, are edit load and would repeat. Together they tell you whether the drafts need work and whether you want to keep going. Add the rest only when a question comes up that they cannot answer. A small log kept for several weeks is worth far more than an elaborate one abandoned after three days.
Where to go next
Start the log on your next job. Lesson 1.4, Talk to AI instead of typing, sets up the habit that gives you something to measure, covering voice typing, ready to send drafts and replies by talking.