When you're signing a data annotation agreement, you're not just buying labor—you're defining the quality, speed, and legal boundaries of your training dataset. A poorly worded contract can leave you with mislabeled images, missed deadlines, or unclear ownership of your annotated data. This guide walks you through the critical clauses you need to understand before signing.
Scope of Work and Annotation Guidelines
Your contract must spell out exactly what you're asking annotators to label. Don't assume "label images" is enough. Specify:
- Image types and formats (RGB photos, medical scans, video frames, etc.)
- Annotation method (bounding boxes, polygons, semantic segmentation, keypoint labeling)
- Label taxonomy (the exact classes and definitions annotators will use)
- Quality thresholds (what constitutes acceptable work)
A real example: instead of "annotate objects in images," write "draw bounding boxes around vehicles in street-view photos using Pascal VOC format, with a minimum of 95% accuracy measured against a hold-out test set." Request that the provider supply detailed annotation guidelines in an appendix—these become enforceable contract terms.
Quality Assurance and Accuracy Standards
This is where vague language costs you money. Establish measurable quality metrics upfront:
- Consensus thresholds: Require 2–3 independent annotators per image, with disputes resolved by a senior reviewer
- Inter-rater agreement (IRA): Specify a minimum Cohen's Kappa or Fleiss' Kappa score (typically 0.75–0.85 for good quality)
- Audit procedures: Define how many images you'll spot-check and who pays if errors exceed the agreed threshold
- Rework clauses: State that the provider must re-annotate failed batches at no additional cost
Most reputable annotation vendors expect these terms. If a provider balks at quality metrics, that's a red flag.
Pricing, Timelines, and Payment Terms
Data annotation costs vary wildly depending on complexity and volume. Know what you're paying for:
- Per-image pricing: Typically ranges from $0.10 (simple classification) to $5+ per image (medical or 3D annotation)
- Volume discounts: Negotiate 10–20% off for batches exceeding 100K images
- Rush fees: Expect 20–50% premiums for 48-hour turnarounds
- Milestone payments: Don't pay 100% upfront. Use a 30% deposit, 50% upon delivery, 20% after QA acceptance
Set a firm delivery schedule tied to batch size. A vendor handling 10,000 images should commit to a specific completion date; include late-delivery penalties (e.g., 5% per week overdue) or a termination clause if they exceed 30 days.
Data Ownership and Confidentiality
This clause determines who owns the annotated dataset and how it's protected:
- IP ownership: Confirm that you own the annotated data and all deliverables, not the vendor
- Data handling: Require the vendor to use encryption at rest and in transit, limit employee access, and sign individual NDAs
- Deletion terms: Specify that all raw data and intermediate files are deleted within 30 days of final payment
- Subcontracting restrictions: Prohibit the vendor from outsourcing to third parties without written approval
- Audit rights: Reserve the right to audit their security practices annually
If your dataset includes sensitive information (medical records, faces, proprietary images), add explicit clauses prohibiting any use outside the contract scope.
Liability, Dispute Resolution, and Exit Clauses
Protect yourself with clear consequences:
- Liability caps: Limit damages to the total contract value (standard for service agreements)
- Indemnification: The vendor indemnifies you against IP infringement claims from their annotators
- Termination for convenience: Allow you to cancel with 14 days' notice and only pay for completed, accepted work
- Dispute resolution: Choose arbitration or small claims court (faster than litigation) and specify jurisdiction
- Force majeure: Clarify pandemic or infrastructure failures don't trigger penalties
Common Red Flags
Avoid contracts that:
- Lack specific accuracy metrics or rely solely on subjective "quality"
- Lock you in for 12+ months with no termination option
- Claim ownership of your annotated data or allow reuse for other clients
- Include no SLA (Service Level Agreement) for delivery timelines
- Offer only one point of contact for issues
Using a platform like Mercoly, you can compare multiple annotation vendors side-by-side, review their sample contracts, and see how others have negotiated key terms—saving time and reducing legal friction.
Frequently Asked Questions
Q: What accuracy score should I demand in my contract? A: For most computer vision tasks, aim for 90–95% accuracy; medical imaging and autonomous driving require 98%+. Define how accuracy is measured (confusion matrix, IOU, or human spot-check consensus) in the contract.
Q: Can I require the vendor to sign an NDA separate from the main contract? A: Yes—in fact, insist on it if your data is proprietary or sensitive. A standalone NDA gives you additional legal standing and often covers employees the main contract doesn't explicitly name.
Q: Who pays for rework if the vendor misunderstands my annotation guidelines? A: The vendor should absorb the first round of corrections if the error stems from ambiguous guidelines; you cover costs if you change requirements mid-project. Lock this in writing to avoid disputes.
Ready to review contracts with confidence? Compare trusted annotation providers and their standard terms on Mercoly today.