Data labelling for AI is the work of marking the correct answer on the examples a model will learn from: drawing a box around a defect in a photo, tagging an email as "invoice" or "complaint", or transcribing a call recording. The model generalises from those labels, so label quality directly sets the quality of its decisions in production.
Why data labelling for AI becomes the bottleneck
Many teams discover only after a custom AI project starts that most of the time goes into data rather than the model. You may have hundreds of thousands of product photos, years of customer correspondence or call recordings. None of it is marked up in a way a model understands, such as "there is a scratch in this photo, and it is here".
The first instinct is often to hand the job to a few interns. Two weeks later you have a dataset in which one person calls a defect a "stain" and another a "paint fault", while some small defects are not marked at all. The model then learns exactly that inconsistency.
We covered when to train your own model rather than use an off-the-shelf one in our guide to custom AI model training. This article focuses on what comes after that decision: preparing the labelled data the model will be trained on.
The cost of poor labels
Research shows how heavily data issues weigh on AI projects:
A survey of 565 data and analytics professionals by Drexel University's LeBow College of Business and Precisely, a data integrity software company, found that a lack of data governance (62%) is the top data challenge holding back AI initiatives, and only 12% say their data is AI-ready. (Drexel LeBow & Precisely, 2025 Outlook: Data Integrity Trends and Insights)
In the same survey, 64% cited data quality as the top challenge to data integrity. For a sense of scale, data quality expert Thomas C. Redman, writing in MIT Sloan Management Review, Seizing Opportunity in Data Quality (2017), estimates the cost of bad data at 15% to 25% of revenue for most companies.
The barrier looks similar in Türkiye. According to TurkStat's AI Statistics 2025 bulletin, 74.2% of enterprises that considered but did not use AI cited a lack of expertise. Labelling is one of the places where that expertise gap is most visible: the task looks simple, but doing it well takes method. Labelled training data is part of the same picture: a wrong or inconsistent label enters the model directly as bad data.
What gets labelled, and how?
The label type is chosen according to the decision the model will make. Labels that are more detailed than needed waste time; labels that are too coarse leave a gap the model cannot learn across.
| Data type | Label type | Example use |
|---|---|---|
| Image | Classification (one label per image) | "Pass / fail" inspection |
| Image | Bounding box | Locating a defect or a hard hat |
| Image | Segmentation (pixel-level area) | Measuring the area of a stain or tear |
| Text | Classification | Sorting emails into invoice, complaint, quote |
| Text | Entity tagging (NER) | Tax number, date and amount in a document |
| Audio | Transcription and speaker separation | Turning call centre recordings into text |
| Time series | Event marking | Marking the period before a failure in sensor data |
We explain how image labels are used in computer vision quality inspection, and how they apply in textiles in AI fabric defect detection.
The data labelling process in seven steps
- Define the decision. What question will the model answer in production? "Is there a defect?" and "which defect, where and how large?" are very different labelling jobs.
- Write the label set and guidelines. Prepare annotation guidelines with a definition, edge cases and example images for every label. The answer to "what do I do if I'm unsure?" belongs in the guidelines.
- Pick a representative sample. The data should reflect real conditions: different shifts, lighting, suppliers and product types. Rare but critical cases are deliberately over-sampled.
- Build a gold set. A small reference set labelled jointly by experts is used both to train labellers and to measure quality.
- Measure agreement. Have two different people label part of the same data and measure inter-annotator agreement (Cohen's kappa, for example). When agreement is low, the problem is usually the guidelines, not the people.
- Speed up with the model. Once a first model exists, it pre-labels new data and people only correct it. Prioritising the examples the model is least sure about (active learning) yields more improvement for the same effort.
- Version and record. Keep a record of which model was trained on which dataset and which version of the guidelines. When the guidelines change, it is then clear which old labels need reviewing. That record is also the foundation of MLOps and the model lifecycle, where models are monitored and retrained.
Managing labelled datasets cannot be separated from wider data discipline in the organisation; who may change which dataset, and which version is current, should be written down. We cover master data, definitions and ownership in our guide to data quality and duplicate records.
Common data labelling mistakes
The most common mistake is starting to label before writing guidelines. Every undefined edge case is left to the labeller's interpretation on the day, and inconsistency quietly enters the dataset. The second is labelling only the "easy" examples: a model trained on sharp, well-lit, typical cases fails on the blurred or borderline images it meets on the line.
The third is ignoring class imbalance. If defective items are a small share of production, the model can learn to say "everything is fine" and report an accuracy that looks high but is useless. The fourth is losing context in document and audio projects. We show how field labelling works in OCR invoice processing and how speaker and topic labels are used in call centre speech analytics.
Who should do the labelling: in-house, outsourced or model-assisted?
| Option | Strength | Weakness | When it fits |
|---|---|---|---|
| In-house experts | Deep domain knowledge; data stays inside | Expert time is costly and limited | Defects, medical images, legal documents and other specialist work |
| Outsourced labelling | Scales quickly | Little domain knowledge; sharing data carries privacy and confidentiality risk | Generic objects, simple classification, datasets without personal data |
| Model-assisted (pre-label + human review) | Cuts effort markedly | Needs a labelled core set to start | Every scaling stage after the first model is built |
In practice the most efficient set-up is a mix: in-house experts write the guidelines and gold set, a model-assisted process carries the volume, and experts only review uncertain and critical examples. That way expensive expert time goes on the hard cases the model genuinely needs to learn from, and labelling speed does not drop as the project grows.
Data protection and security
If the data to be labelled contains personal data, labelling is itself personal data processing. Camera footage showing employees' faces, voice recordings, customer emails and documents with identity details all fall into this category. Under KVKK, Türkiye's data protection law, the purpose, legal basis and retention period should be set in advance, and faces and identity details masked before labelling wherever possible.
If labelling is handed to a provider abroad, or to a tool hosted outside Türkiye, this will usually count as a cross-border transfer and should be assessed under Article 9 of KVKK. We explain how the 2024 amendment reshaped those rules in KVKK cross-border data transfer, and set out the broader obligations for AI projects in AI and KVKK personal data.
How we run data labelling at Digital Bridge
We treat labelling not as a separate task but as the first and most decisive stage of a custom AI model training project. We start with a needs analysis and set out scope and phases in a written proposal:
- Discovery: together we define the decision the model will make, the data available and who knows that data best. For image projects we also assess camera and lighting conditions at this stage; for computer vision projects we can also install cameras, lighting and processing units on site where needed.
- Guidelines and gold set: we write label definitions and edge cases with your experts and prepare the reference set.
- Pilot: we train a first model on a limited dataset and show its errors class by class. We measure accuracy on an independent test set the model has not seen during training, and those errors decide where the labelling budget goes next.
- Governance and compliance: we set up a data governance and data quality framework for dataset versioning, ownership and access rights, and for projects involving personal data we settle masking and transfer rules through our data protection compliance consultancy.
Keeping the guidelines and example images in one place with their versions also matters. In SmartFiles document management, part of the Smart360 family, every file uploaded under the same name is stored as a new version and earlier versions can be restored, which makes it easier to trace which guideline version was used in which labelling round. For more guides on the topic see Artificial Intelligence, and for choosing a first project read AI in business: where to start.
Next step
Write down in one sentence the decision the model will make, then ask two different people to label the same 100 random examples separately. How far they agree tells you where the project should begin. To review the result with us, get in touch through our contact page.