Zovda
Building a labeling team looks cheaper than buying data, because the comparison is a salary against an invoice. The invoice includes everything; the salary includes one person.
The rest is spread across people who were hired for something else.
The four costs that do not appear
- Guideline writing and its revisions, which is senior work
- Review, which needs a second person per row at some rate
- Ramp-up, which is weeks before the first usable row
- Idle time between projects, which is most of the calendar
A team of five producing steadily for a year is a rare achievement. A team of five producing for four months and waiting for scope for eight is normal.
Task shape decides the answer
Narrow, repetitive, high-volume work with stable rules favors building. Ambiguous, changing, low-volume work favors buying, because the setup cost never gets amortized.
Anything that needs a specialist to judge favors buying the specialist's hours, not hiring them.
Count the switching cost
Every new task needs a guideline, a calibration round, and a pilot. Building means paying that cost for each task; buying spreads it across customers.
If you have three tasks a year, the arithmetic rarely favors building.
Compare per usable row
Divide the annual cost, including idle time and review, by the rows that passed your checks. That number is comparable to a quote and almost never to a salary.
Zovda runs these programs through Datuva, where the guideline and the review record travel with the data. The corpus catalogue shows what each program produces.
Tell us the task and the volume. We reply within one business day.