Skip to content

Managed labelling vs platforms: who owns the quality problem?

The real difference between a managed service and a software platform is where accountability for label quality sits. Everything else follows from that.

Read the guide
Define the task → Apply expert judgment → Review the resultThree connected layers illustrate how a defined task becomes expert annotations and a reviewed reference.
Define the taskApply expert judgmentReview the result

Three operating models

ModelYou supplyVendor suppliesQuality owner
Self-hosted open-source toolWorkforce, hosting, QA designSoftware onlyYou
Commercial platformWorkforce or vendor marketplace, QA designHosted software, workflow featuresMostly you
Managed serviceGuidelines, acceptance criteriaWorkforce, tooling, QA layersShared, by contract

How to tell which model a provider actually runs

Provider marketing often blurs these categories. Ask three things in writing: who employs the people doing the labelling, whose infrastructure the data sits on during work, and who signs off quality before delivery. The answers place a provider in one model regardless of how it describes itself, and they tell you where contractual responsibility for errors and rights will sit.

Decision questions

  • Can your data leave your environment? If not, self-hosted tooling may be the only option.
  • Do you have someone who can write and maintain labelling guidelines? Managed services still need them.
  • Is the task stable for months, or changing weekly? Rapidly changing tasks favour in-house loops.
  • Do you need a contractual quality guarantee, or is internal measurement enough?

Compare the full cost of delivery

Compare a scoped quote rather than a headline rate. Structurally, platforms tend to charge for seats or usage, while managed services price per item, per hour or per project. Ask any vendor to show the total cost including rework of rejected batches.

Have an annotated dataset or specialist expertise?

Tell us what you own, how the labels were reviewed and which rights you can license. A short overview of your dataset or annotation business is a useful place to start.

Prepare for a licensing conversation