Short answer. In careful usage, labelling assigns a class or value to a whole record — this message is a billing enquiry, this image is indoors — while annotation is the broader term covering labels plus structure: where the signal is, what attributes it has, when it occurs, how it relates to something else, and why the judgement was made. In commercial usage the two words are interchangeable, and vendors use them inconsistently enough that neither tells you what you are buying. The practical consequence is that the terminology is not a specification and cannot be compared across quotes. What can be compared is the output schema: the geometry, the attribute set, the edge-case rules and the review standard. Write those down and the vocabulary question disappears.
Two providers describing the same workflow will call it different things, and the same provider will use both words on the same page. This is not evasion — it is genuine terminological drift across modalities, tools and company histories. It becomes a commercial problem only when a buyer uses the word as though it defined the deliverable, which is when quotes stop being comparable and scope arguments start.
Why are the two terms used interchangeably?
Both describe adding supervision to raw data, and every labelling task is trivially also an annotation task. A team marking images as "cat" or "dog" is annotating them. The distinction that survives across most usage is one of granularity:
- Labelling tends to describe a decision about a whole record, drawn from a fixed set of options.
- Annotation tends to describe a decision about a part of a record, or a decision carrying additional structure — a span, a polygon, a timestamp, an attribute, a relation, a rationale.
Note also that the American spelling labeling and the British labelling are the same word, and neither carries a technical distinction. Vendors in the same market use both.
Where the terms diverge in practice is by modality, and largely for historical reasons:
| Modality | Word usually used | Why |
|---|---|---|
| Image and video | Annotation | The output almost always includes location, not just category |
| Speech and audio | Transcription, or annotation | Output is a text stream plus timings, not a class |
| Text | Both, roughly evenly | Document classification is labelling; entity extraction is annotation |
| Tabular | Labelling | Record-level targets by construction |
| LLM output review | Evaluation, rating, or annotation | The target is a judgement, not a property of the input |
Where the distinction does carry real information
There is one place the difference is not cosmetic, and it is the reason it keeps mattering commercially: the two imply different billable units.
A record-level label is priced per record, and the price is stable because every record takes roughly the same time. A sub-record annotation is priced per object, per span or per hour, and the cost is driven by density — how many objects are in the image, how many entities are in the document — which varies enormously between assets that look identical in a file listing.
This is why a vendor who quotes per image without asking about your average objects per image has priced an assumption rather than your work. Anyone comparing quotes should read how to compare annotation vendor quotes and annotation pricing models before normalising anything, because the unit is where most of the variance between quotes lives.
What a specification has to say instead
Replace the term with eight fields. If all eight are answered, no one needs to agree on vocabulary; if any are blank, the word on the contract will be filled in differently by each party.
- Output schema. The literal structure of one delivered item — field names, types, permitted values. Not a description of it: the structure.
- Geometry or granularity. Whole record, span, box, polygon, mask, keypoint, cuboid, timestamp range, or a rank over candidates.
- Taxonomy. The class list, with an inclusion and exclusion test per class rather than a one-line description.
- Attributes. The optional properties attached to an instance — occlusion, truncation, visibility, confidence, dialect, speaker role. Attributes are how you record a real distinction that annotators cannot apply consistently as a class.
- Edge-case rules. What happens with partially visible objects, overlapping speech, mixed-language text, ambiguous intent — and, critically, what an annotator does when genuinely unsure. A skip-and-flag path with adjudication beats a forced guess, which manufactures noise and hides it.
- Density expectation. Average and range of objects, entities or turns per asset. This is the number that determines cost.
- Quality standard. The metric, the threshold, the overlap rate for double-labelled work, and who adjudicates disagreement.
- Format and delivery. File format, coordinate convention, encoding, split structure, and what accompanies the data — guideline version, agreement figures, provenance.
Two rules of thumb sit behind that list. Do not collect detail the model will never use — every additional field costs money and adds an opportunity for inconsistency. And prototype the schema on a small, deliberately diverse sample before scaling: if trained reviewers cannot apply the instructions consistently, the taxonomy needs revision, and finding that out after fifty items is a morning's work rather than a re-adjudication project.
Which does your project actually need?
Work backwards from what the model must predict.
| If the model must... | You need | Priced by |
|---|---|---|
| Choose one category for a whole record | Record-level labelling | Per record |
| Find something and say where it is | Localisation annotation | Per object |
| Track something across time | Temporal annotation with persistent identity | Per object per sequence, or per hour |
| Relate two things to each other | Relation annotation | Per document, usually per hour |
| Judge whether an output is good | Rubric scoring or preference ranking | Per comparison |
| Explain why a judgement was made | Annotation with rationales | Per item, at the highest rate |
The bottom two rows are where most enterprise spend has moved, and they are also where the word "labelling" is least useful — nobody is assigning a class to anything. A rater is applying a rubric to open-ended output, and the quality question is whether independent raters agree, not whether a label is correct. That work is covered separately in what to buy: RLHF, SFT or distillation.
How to compare vendors when the vocabulary differs
Ignore the headline term and compare the operating method. Six requests settle it, and none of them require the vendor to use your words:
- A sample delivered file from a comparable project, so the schema is visible rather than described.
- The reviewer instructions for that project, including the edge-case section — which is the part that grows during a project and is the highest-value artefact it produces.
- The quality method: overlap rate, metric, threshold, adjudication path, gold-set injection.
- The pre-labelling position: what is automated, at what confidence, and whether unassisted control batches are kept.
- The escalation route for uncertain and out-of-scope items.
- Turnaround on a guideline change, because production annotation is iterative and the responsiveness matters more than the initial rate.
Red flag: a proposal that answers your specification by restating your terminology back to you. It means the vendor has agreed to the word rather than to the deliverable, and the disagreement has been deferred to the first disputed batch rather than avoided.
How Lifewood approaches this
Lifewood scopes work from the output schema rather than from the service name — geometry, taxonomy with inclusion and exclusion tests, attribute set, edge-case rulings and adjudication path agreed before the first batch, then versioned as rulings accumulate. The quality standard is stated per task type rather than as a single figure, with dual-layer human-in-the-loop review held to a 95%+ accuracy threshold.
Scope spans classification and record-level labelling through to segmentation, tracking, transcription, relation extraction and LLM response evaluation, across 50+ languages and 40+ delivery centres in 30+ countries with 56,788 registered contributors — which matters for the tasks where the judgement depends on hearing or reading the material as a native speaker would. The AI-data heritage runs to 2004, with the current company established in 2018.
See global AI data, AI data services, the QA process and AI data validation.
Sources and further reading
- NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0, January 2023) — on documenting data decisions across the lifecycle.
- Companion guides: How to Compare Annotation Vendor Quotes and Data Annotation Pricing Models.