Who does the work, and where?
Some vendors do the work in-house, some pass it to a crowd, and some pass it to another vendor. None of those is automatically wrong. What matters is that you can find out which one applies to you, and that the answer isn't vague.
For us, it's a network of 30,000+ people worldwide, working in tools we built and run ourselves.
How is quality measured?
Be wary of a flat promise of near-perfect accuracy with no explanation. Ask how quality is checked: how many review layers there are, how disagreement between annotators is handled, and whether you can see the numbers.
Every batch of ours goes through multi-level review before delivery. If you want independent figures, our evaluation service can measure accuracy against targets you set.
Can they work in your languages?
A team that labels English well may not label Tamil or Bengali well, and dialect matters even within one language. Ask who would actually do the work, and whether they're native speakers of the language and region you care about.
We work in languages from all over the world and go deep on Indian ones. The page on Indian language data has more.
What happens to your data?
Ask for the answers in plain terms. Is there an NDA before any data moves? Who can see the data? Can it be reused for other clients or other projects? What happens to it when the work ends?
Here's ours. Every project starts with an NDA, before we touch your data. We use it only for what we agreed, nothing gets reused or resold, and it's deleted or returned the moment you ask. Only the people working on your project can see it.
Who will you talk to?
When something goes wrong, you want a person who knows your project, not a ticket number. Ask whether you'll have a named contact, and whether that person is the one you speak to before signing.
One named person owns your engagement from scoping to delivery, and you'll hear from Krishsinh first.
Do they build their own tools or rent them?
This affects how quickly a vendor can adapt a workflow to your project, and how much of your data passes through third-party software. Neither answer is wrong, but you should know which one you're getting.
We've already built our own tools for audio prompts, conversation capture, transcription, translation, image annotation, and image, video and media collection.
Can one vendor cover what you need?
If your project needs collection, labeling and evaluation, splitting it across three vendors means three sets of instructions and three places where things get lost. Check whether the vendor can carry the whole chain, even if you start with only one part.
We offer eight services, from collection and annotation to transcription, translation, evaluation and synthetic data.
Warning signs
A vendor is worth a second look if they won't explain their process, can't say who does the work, are vague about what happens to your data, or can't name a person for you to talk to. A very low price with no explanation of how it's possible deserves the same suspicion.