Data Labelling and Annotation Services: Offshore, EU and Expert Teams
Nov 7, 2026
Most AI projects need examples that people have checked and described. Is this email a complaint or a question? Which line on this invoice is the VAT? Is the chatbot’s answer correct? Attaching those answers to examples is called data labelling (data labeling in American spelling), or data annotation. Good labels are what let you train a model and check whether it works.
It’s also a step that’s easy to underestimate. This guide explains what data labelling involves, how to keep quality high, and how our data annotation services are staffed: offshore teams for volume, EU-based students when data has to stay in Europe, and professionals when the judgement needs real expertise.
What data can be labelled?
Data | Typical label | Used for |
|---|---|---|
Emails and tickets | Topic, urgency, positive or negative tone | Sending requests to the right team |
Invoices and contracts | Fields, amounts, clauses | Document extraction |
Images and video | Objects, defects and where they are | Quality control, inspection |
Chatbot answers | Correct, partly correct or wrong, and why | Testing and training language models |
Search results | Relevant or not, in order | Better search and document Q&A |
Data labelling for LLM evaluation
LLM evaluation uses reviewed chatbot answers and search results to test whether a system performs well on the cases that matter to your business. Teams building with language models need labelled examples less to train from scratch and more to test. A set of carefully reviewed answers, covering the cases that matter to you, is how you decide whether a model is good enough to ship. Reviewed examples also help train custom models with reinforcement learning, where a model learns from scored attempts.
How we keep quality high
A label is an answer attached to an example, such as complaint for an email. Labels are only useful if they’re consistent, so we build quality checks into every project:
Write the guidelines. A short rulebook with examples of each label and, more importantly, the tricky cases in between. It grows as questions come up.
Run a pilot batch. A few hundred items, each labelled by several people, so we can measure how often they agree (inter-annotator agreement). Where they disagree, we look at why. Usually the guidelines need clarifying, sometimes an expert has to decide, and sometimes more than one answer is reasonable.
Hide test items. Where the task allows, items with reviewed answers are mixed into the work, so we can track each labeller’s accuracy.
Double-label a sample. A share of items is labelled twice, and disagreements go to a senior reviewer.
Pre-label with AI. For many tasks a model suggests the label and people confirm or correct it, a human-in-the-loop setup. That saves time, but suggestions can sway reviewers, so we also check a sample labelled without the suggestion.
We agree up front which quality measures you’ll receive, and report them with every delivery, so you can see the quality instead of taking our word for it.
Who does the data annotation?
The right team depends on the data. Through our partners, we work with three kinds of teams, and most projects mix them.
Offshore teams: volume at a lower cost
For large volumes of non-sensitive data with clear rules, such as product images, public web text or anonymised tickets, we work with established offshore labelling partners. They can usually scale up quickly and cost well below European rates. Before any offshore work, we check whether the data can be properly anonymised for the task. Removing names alone isn’t always enough, so if personal data would remain, the work goes to the EU team instead.
EU-based students: when data has to stay in Europe
Some contracts require data to stay in the EU, and sending personal data outside Europe needs a legal basis and safeguards that many organisations would rather avoid. For these cases we work with a large pool of university students based in the EU. It costs more than offshore. Before work starts, we agree with you where the data is hosted, who can access it and how it’s processed. Students can also be picked for the languages in your data, such as French, Dutch, German or English.
Expert labelling: when it takes a professional
Some labels need real expertise. Is this contract clause unusual? Does this scan show a fracture? Is this answer to a tax question correct under Belgian rules? A student can’t judge that reliably, so we bring in professionals such as lawyers, doctors, engineers or accountants, depending on your field. They cost more per item, so we use them where it matters most. They write the guidelines, label the test set and review the cases other labellers flag.
Team | Offshore | EU students | Experts |
|---|---|---|---|
Best for | Large volume, clear rules | Personal or sensitive data | Judgement calls, test sets |
Where data goes | Outside the EU, only if properly anonymised | Stays in the EU | Stays in the EU when personal data remains or residency is required |
Relative cost | Lowest | Medium | Highest |
Speed to scale | Fast | Fast | Slower |
Keeping your data safe
Least access. Labellers work inside a labelling tool on the items assigned to them, not in your source systems.
Personal data handled deliberately. Offshore work only gets data that is properly anonymised for the task. Anything else stays with the EU team.
Contracts in place. Confidentiality obligations for everyone involved and, where personal data is processed, data processing terms covering each partner, agreed before work starts.
Audit trail. The labelling tool records who labelled what and when.
These are the same rules you’d apply anywhere else in your organisation. If they aren’t written down yet, that’s data governance work worth doing first.
The tools we use
We use open-source labelling platforms such as Label Studio, Argilla and CVAT, hosted in the EU or on your own infrastructure, the same way we run models for sovereign AI. We agree the export format and label definitions with you, so your team can use the dataset in its own tools.
How much does data labelling cost?
Most labelling projects, from guidelines and a pilot batch to the final labelled dataset, land between €15,000 and €100,000 excluding VAT. Where yours falls depends on the volume, how complex the labels are, the quality checks, the deadline and how much needs experts. We quote after a first call, and the pilot batch confirms the effort per item before you commit to the full volume.
Frequently asked questions
How many labelled examples do we need?
It depends on the task, the variety in your data and how confident you need to be. As a rough guide, testing a language model often starts with a few hundred reviewed examples, while training a classifier or an image model usually takes thousands or more. The pilot batch gives a real estimate for your case.
Can’t AI just label the data?
Partly. A model can pre-label and people correct it, which cuts the cost. But AI-generated labels can contain systematic mistakes, so they need checking. A test set in particular needs independently reviewed answers, or you risk grading a model with its own judgement.
How long does it take?
It depends on the volume, the complexity and the team. Plan for two weeks for a pilot batch, and we agree a schedule for the full volume once we’ve seen how long each item really takes.
Which languages can you cover?
For Belgian data, French, Dutch and English are the usual mix, often with German. Other languages depend on partner availability, which we confirm before quoting.
Can our own staff do the labelling?
Yes, and for small expert test sets it’s often the best option. We provide the guidelines, tools and quality checks, and your team brings the judgement. Many projects combine your experts with our larger teams.
Do offshore partners ever see personal data?
Offshore teams only work on data that has been properly anonymised for the task. If that isn’t possible, the work goes to the EU-based team.
Need labelled data for a model, a test set or a search project? Tell us about it with what the data is, roughly how much there is, and whether it can leave the EU. We’ll suggest the right mix of teams and a pilot to start. New to AI? Our interactive guide to how LLMs work shows why examples matter so much.
CONTACT
