Annotation cost is meaningful only when the unit of work and the quality review are defined. A quote for labeling an image is not directly comparable with a quote per object or reviewed batch. This guide explains how buyers can scope a pilot, compare estimates and agree acceptance without relying on an unexplained accuracy claim or a universal market rate.
Updated
On this page
Define the unit behind the price
| Possible unit | What the buyer needs to clarify |
|---|---|
| Image | Whether several objects, classes or attributes are included |
| Object or marked region | Which geometry and review rules apply |
| Text item | Length, language and the decisions required |
| Hour or delivery capacity | How completed work and review are recorded |
| Accepted batch | Which checks determine acceptance and what happens to rejected items |
A unit should represent the task you actually need. If the dataset contains a wide range of complexity, ask how difficult items are identified and priced. A low rate can be misleading when it applies only to the simplest portion of the collection.
Understand what changes the effort
Label count, ambiguous cases, specialist terminology, tool behavior and the required output can all affect the work. Review is another workstream, not a free property of every completed label. Ask who checks the result and what that review is expected to find.
Guidelines also affect cost. A clear rule with examples can reduce repeated questions. A definition that changes after production may require earlier items to be reviewed again. The proposal should explain how instruction changes and their resulting rework are handled.
Compare two illustrative image tasks
Consider two hypothetical collections of one thousand images each. The first typically contains two objects to mark per image; the second typically contains eight. The image count is identical, but the second task involves roughly four times as many object decisions before considering complexity or review.
This example is not a productivity estimate or a quotation. Objects can vary in difficulty, and some work happens once per image or batch. Its purpose is to show why a buyer needs representative samples rather than a price comparison based only on file counts.
| Quote item | Question for both suppliers |
|---|---|
| Pilot | Does it include guideline clarification and a checked output file? |
| Production | What task and complexity assumptions does the price cover? |
| Review | How is the review sample selected and who resolves disagreements? |
| Corrections | Which errors are corrected within scope and which changes require another estimate? |
Agree quality checks before production
Define the errors that matter to the receiving workflow. A missing object, an incorrect class and a geometry error are different issues. An item that is genuinely ambiguous may need a review flag rather than a forced answer.
Keep an instruction version and a review record. If reviewers disagree, determine whether the rule is missing, the task is subjective or an instruction was applied incorrectly. A single quality percentage without those definitions does not tell a buyer what the dataset can be used for.
Use the pilot to improve the purchasing decision
- Choose a representative sampleInclude normal items and the difficult cases present in the larger collection.
- Review the task definitionsResolve recurring questions and record accepted examples.
- Test the delivery formatConfirm that the output can be used by the receiving team or system.
- Compare the full estimatesUse the same unit, review depth, acceptance process and change assumptions.
A pilot should leave the buyer with better information even before a production engagement begins. From Ethiopia or any other delivery location, the relevant comparison is the actual work and review arrangement. The supplier’s country alone does not establish quality, capacity or savings.
For an annotation proposal from Zoha, the next step is a representative sample and a shared task definition. We use those to discuss the work unit, review approach, exception handling and delivery format. The estimate should explain the scope behind its price and how corrections are handled, so your team can assess the service before extending the work to a larger dataset.