Data annotation pricing isn’t one neat number, it’s a shape: a per-unit or hourly rate plus quality assurance, security and project management costs layered on top. The single most important thing you can do as a buyer? Ask for a scoped quote that spells out QA, a pilot phase and DPA or SCC alignment before you sign anything. Some providers build GDPR-aware annotation into that quote from day one.
What actually drives annotation price
Right, let’s get into the detail. Every annotation quote is really answering five questions: how hard is the task, how much data is there, how accurate does it need to be, who’s allowed to touch it, and how fast do you need it back? Get clear on these and you’ll spot a padded quote from a mile off.
Task type is the biggest lever. Simple binary classification (is this a cat or not) is cheap and fast. Bounding boxes and polygon segmentation take longer per item because annotators are drawing precise shapes, not just picking a label. Transcription and audio labelling add a time cost because someone has to listen, pause, type, and check. Complex tasks like semantic segmentation or multi-object tracking in video sit at the top end, simply because there’s more surface area for error.
Data modality matters just as much as task type. A single image with one object is a different job to an image with fifteen overlapping objects that each need their own class. Video multiplies this again, since every frame (or every few frames) might need re-annotation. Text tends to be cheaper unless you’re asking for nuanced sentiment, entity relationships or domain-specific tagging that needs someone who actually understands the subject matter.
Then there’s quality. This is where a lot of quotes get vague, so push for specifics.
A few other cost drivers worth flagging:
- Specialist language or subject expertise: annotating legal, medical or financial data for regulated sectors needs annotators who understand the domain, not just the language, and that pushes rates up.
- Volume and throughput: a rush job with a tight deadline usually carries a premium, while larger, steady volumes can unlock better per-unit rates.
- Security requirements: if your dataset includes personal data, health records or anything sensitive, you’ll likely pay for isolated environments, restricted-access reviewers and encryption, and that should show up as its own line item.
- Onboarding and tooling: setting up the annotation platform, writing the guidelines, and running a pilot all cost something before a single unit gets labelled.
- Project management: someone has to coordinate annotators, track progress and manage revisions, and that overhead is usually baked into the rate rather than itemised, which is worth asking about.
Pro Tip: Ask every provider to break their quote into per-unit rate, QA percentage, and fixed setup costs separately, so you’re comparing the same three numbers across every bid.
How providers price work, and worked budgeting examples
There are basically three pricing shapes in this market, and picking the wrong one for your project is how budgets blow out.
- Per-label or per-unit pricing charges a fixed amount per image, per audio minute, or per text record. It’s efficient for large, repetitive, well-defined tasks (think product categorisation at scale), but it can hide costs if QA, adjudication and rework aren’t clearly included in that per-unit figure.
- Hourly or resource-based pricing works better for unpredictable or creative labelling, where task complexity varies item to item and a flat per-unit rate would either overpay on easy items or underpay on hard ones.
- Hybrid and milestone pricing combines a per-unit base rate with separate line items for QA rounds, adjudication and project management, or breaks a large project into paid milestones (pilot, scaling phase, final delivery). This tends to be the easiest model to compare across vendors, because nothing’s buried inside a single number.
Let’s imagine some numbers to see how the maths adds up (these are not market rates, just a way to picture the shape of a quote). For example, a pilot batch of images at a low unit rate might come to a modest sum for annotation alone. Add a QA pass at a moderate percentage of annotation cost and the total grows accordingly. Scaling that rate to a larger project increases the base annotation cost plus QA, adjudication for disputed labels, and a project management fee. A video transcription example might use a per-minute rate and result in a cost before QA.
A well-specified data processing agreement that ties payment items to compliance tasks can reduce downstream rework costs, according to EDPB guidance on linking contractual clarity to operational outcomes. In practice, that means a contract clause about deletion timelines or subprocessor use isn’t just legal box-ticking, it can save you money later.
To normalise quotes properly, ask every provider the same three questions: does this rate include QA, and at what sampling percentage? Is adjudication (resolving disagreements between annotators) charged separately or bundled? What’s the rework policy if accuracy falls below the agreed threshold? Without answers to these, you’re comparing apples to a vague fruit basket.
Contract and GDPR items that must be reflected in price and the offer
Here’s the bit most procurement teams skip, and it’s the bit that bites hardest later. If your dataset contains any personal data (names, faces, voices, health information, anything that identifies a person), GDPR applies, and that has a direct effect on price.
First, work out who’s the controller and who’s the processor. If you’re sending the annotation provider personal data to label, you’re very likely the controller and they’re the processor, which means you need a proper data processing agreement, not just a service contract. This isn’t optional, and it’s not free: implementing the right security measures, subprocessor management and audit trails all cost the provider something, and that cost should appear in your quote.
The Regulation (EU) 2016/679 (GDPR) is the legal foundation here, and it applies to personal data processing across the EU regardless of where the annotation work physically happens. If data moves outside the EU/EEA, you’ll need standard contractual clauses, and the European Commission’s Commission Implementing Decision (EU) 2021/915 sets out the baseline template, which must be completed with the specifics of your processing activity rather than used as a generic tick-box.
Ask your provider to itemise or at least confirm these elements before you accept a quote:
- A signed DPA referencing the standard contractual clauses where data crosses borders, completed with your specific purposes and categories of data.
- A subprocessors list, so you know exactly who else might touch your data and where they’re based.
- Storage and deletion terms, spelling out how long data is retained and what happens to it at project end.
- Encryption and access controls, particularly for sensitive categories like health or biometric data.
- Audit rights, so you can verify compliance claims rather than take them on faith.
The EDPB standard contractual clauses appendix details exactly what needs completing here: purpose, categories of data, retention periods and technical measures. A vague “we’re GDPR-compliant” line in a proposal isn’t enough. Ask for specifics, because EDPB guidance explicitly recommends concrete operational descriptions over generic compliance statements, and those specifics are exactly what let you compare compliance-related costs fairly across providers.
A practical RFP / RFQ checklist to get comparable, GDPR-ready quotes
Want quotes you can actually compare side by side? Send every provider the same brief, and ask for the same structure back. Here’s what that looks like.
- Share a real data sample along with your labels or ontology, so annotators know exactly what “correct” looks like.
- Define the expected output format (JSON, CSV, COCO format, whatever your pipeline needs) so there’s no rework translating between formats later.
- Set acceptance criteria upfront: what accuracy percentage counts as a pass, and what happens if a batch fails it.
- Require a DPA and SCCs (where relevant) as a condition of the quote, not an afterthought negotiated after signing.
- Ask for the subprocessors list, deletion policy and audit rights in writing, before you compare pricing.
- Request pricing broken into three parts: price excluding VAT, the VAT rate and amount, and price including VAT, which mirrors the format used in standard procurement price annexes.
- Confirm payment terms (deposit, milestone payments, net terms) alongside the pilot cost.
- Ask for a paid pilot with its own defined acceptance criteria, separate from the main project quote.
Pro Tip: When you get quotes back, line them up in one spreadsheet with columns for base rate, QA inclusion, adjudication, and rework policy. If a column’s empty, that’s your next question to the vendor, not a gap you fill in with assumptions.
A paid pilot is genuinely the single most effective way to validate the assumptions behind a unit price and surface hidden complexity before you commit to the full volume.
Copyable cost-breakdown template and short example
Here’s a template you can drop straight into a tender document or an email to a shortlist of providers. Ask each one to fill in every row, in your currency, with VAT shown separately.
| Line item | What it covers | Example basis |
|---|---|---|
| Pilot | Small test batch to validate assumptions | Fixed fee or per-unit rate on sample volume |
| Per-unit annotation | Core labelling cost | Rate per image, minute, or record |
| QA rounds | Review and accuracy checking | Percentage of annotation cost or per-unit add-on |
| Adjudication | Resolving disagreements between annotators | Per disputed item or hourly |
| Project management | Coordination and reporting | Fixed fee or percentage of project total |
| Tooling and licence | Annotation platform access | Fixed monthly or per-project fee |
| Security and compliance | Isolated environment, encryption, restricted access | Fixed uplift or per-unit premium |
| Reporting | Progress and quality reports | Fixed fee or included in project management |
Ask providers to add price excluding VAT, VAT rate and amount, and price including VAT as final rows, which matches the formatting guidance in structured data readiness and security audits, which recommend showing labour and resource costs clearly within the final price.
A few things to check before you accept the filled-in template:
- Every row has a real number, not “varies” or “TBC”, since a blank row is a hidden cost waiting to surface later.
- Payment terms are stated, including deposit percentage and milestone triggers.
- The pilot row is priced separately, so you’re not locked into full-volume pricing based on an unvalidated rate.
glocco® proof points: quality controls and pricing transparency
Some established providers have spent years building workflows that make a quote trustworthy rather than just cheap. Some language service providers offer data annotation alongside other text and video services, applying quality discipline similar to that used in translation to annotation and labelling work.
Take video captions as an example: our two-human-review workflow means every caption gets checked twice before it’s considered audit-ready, not once. That kind of layered QA is exactly what should be showing up as a distinct line in any annotation quote, ours included, because it’s the difference between “labelled” and “labelled correctly.”
We also run structured quality audits on annotation output, so buyers aren’t just taking our word for accuracy, they can see the process behind it. When personal data is involved, our approach to anonymisation feeds directly into how we scope contracts, because the technique used changes what the DPA needs to cover.
The goal is to ensure the price quoted reflects the work needed to get trustworthy data, highlighting the importance of asking for an itemised quote.
Four common procurement mistakes and short corrective actions
Here’s an honest take: most annotation budgets don’t blow out because of one dramatic failure, they blow out because of four small, avoidable mistakes made at the quoting stage.
The first is chasing the lowest unit price without checking what QA scope it includes. The second is vague acceptance criteria: if you haven’t defined what “accurate enough” means before work starts, you’ve got no basis to reject a bad batch. The third is skipping the subprocessors list, which sounds like paperwork until a data protection query lands on your desk and nobody can tell you who touched the dataset. The fourth is skipping the pilot entirely and going straight to full volume, which is the single fastest way to discover a labelling ambiguity after it’s already been repeated ten thousand times.
- Don’t decide on unit price alone: ask what QA percentage and rework policy are baked into that number.
- Write acceptance criteria into the RFP, not into a follow-up email once work has already started.
- Demand a subprocessors list and deletion policy before data leaves your hands.
- Always run a paid pilot before committing to the full project volume.
How glocco® can help with your next annotation project
If you’ve read this far, you already know the drill: get a scoped quote, insist on QA line items, and make sure the DPA actually says something specific. That’s exactly how we work.
Our Data Annotation service is built around the same transparency this whole guide has been arguing for. You’ll get a quote with QA, adjudication and project management shown as separate items, not folded into one mystery number. Data processing terms are aligned to GDPR and SCC requirements from the start, so there’s no scramble to retrofit compliance once the contract’s signed. And if you’re not ready to commit to full volume, we’ll scope a pilot first, with its own acceptance criteria, so you can see exactly how the work holds up before scaling.
Whether you’re labelling text for a fintech model, annotating video for a compliance-sensitive tool, or validating a dataset that touches personal data, the process should feel the same: clear, itemised, and grounded in a contract that actually protects you. Head to our Data Annotation page to request a scoped quote or a pilot, and see the line items for yourself.
This article is general information, not a substitute for advice from a qualified lawyer. Consult a qualified legal professional about your own circumstances before acting on anything here.
Sources
FAQ
How much does data annotation typically cost?
Pricing shapes vary by task complexity, data modality and required accuracy, so there’s no single universal rate. Ask providers for a quote broken into per-unit rate, QA percentage and setup costs so you can compare like for like rather than relying on a headline number.
What should a GDPR-ready annotation quote include?
It should reference a signed data processing agreement, list any subprocessors, state deletion and retention terms, and confirm standard contractual clauses where data crosses borders, as outlined by EDPB guidance. A quote without these specifics hasn’t accounted for compliance costs properly.
Should I always run a pilot before a full annotation project?
Yes, a paid pilot is the most effective way to validate a provider’s unit price and surface hidden complexity in your dataset before you commit to full volume. Define acceptance criteria for the pilot separately from the main project’s criteria.
Does glocco® offer a pilot for data annotation projects?
Glocco® scopes pilots as part of its Data Annotation service, letting buyers test acceptance criteria and quality before scaling to full volume. Pricing for pilots and full projects is available on request through that service page.
What’s the difference between per-unit and hourly annotation pricing?
Per-unit pricing charges a fixed rate per labelled item, which suits large, repetitive tasks, while hourly pricing suits unpredictable or creative labelling where task difficulty varies. Hybrid models combine both, adding separate QA and adjudication lines for easier comparison across quotes.
