When an AI team licenses data, from a vendor, a publisher or a business that created it, the data licensing agreement decides what the team can actually do with it. A weak agreement can leave a model built on data the team isn't entitled to use. This guide sets out the key terms of a data licensing agreement for AI training, evaluation and retrieval, and the questions each side should ask. It is not legal advice, so always involve qualified lawyers.
Parties and authority
The agreement should identify the licensor and confirm that it has authority to grant the licence. For business records, authority usually sits with the company acting through its directors, or with an insolvency office-holder where the company is in an insolvency process.
Definition of the licensed data
Be precise:
- the systems and record types included
- the date ranges
- the formats and delivery method
- exclusions, such as HR, special category data or client-confidential material
- any later updates or additional deliveries
A schedule describing the dataset, with a data dictionary, avoids disputes.
Permitted use
This is the heart of the agreement. Consider whether the licence covers:
- training, fine-tuning and pre-training
- evaluation and benchmarking
- retrieval and grounding at run time
- building RL environments (see RL environments)
- annotation and the creation of derived datasets
- internal use only, or use in commercial products
- sharing with affiliates, contractors or cloud providers
Prohibited uses should be explicit, for example re-identification, resale of raw data, or use in particular sectors.
Exclusivity
Exclusive, non-exclusive, or exclusive within a field, time or territory. Exclusivity affects price and the licensor's freedom.
Term and termination
How long the licence lasts, and what happens at the end. A key question for AI is the status of models already trained. Most agreements state whether trained models may continue to be used after termination, and whether retraining or deletion is required.
Derived data and outputs
Who owns annotations, cleaned versions, embeddings and synthetic data generated from the licensed data? Can the licensee keep them after termination?
Privacy and data protection
- Who is controller and who is processor, if personal data is involved?
- What anonymisation or pseudonymisation is applied before transfer?
- Obligations not to re-identify
- Security measures
- International transfer mechanisms
- Breach notification
Warranties
Typical licensor warranties cover ownership or authority, lawful collection, and the absence of known infringement. Licensees may want warranties on accuracy or completeness, which licensors often limit. Realistic, specific warranties are better than broad ones that will be qualified away.
Indemnities and liability
Allocation of risk for third-party claims, such as copyright or privacy, with caps and exclusions.
Provenance and audit
Rights to receive provenance documentation, and to audit compliance on reasonable notice. See data provenance.
Payment
Fixed fee, staged payments, per-record pricing, revenue share or a combination. Define acceptance criteria for deliveries.
Confidentiality
Both the data and the commercial terms are usually confidential.
Deletion and return
What must be deleted at the end, how deletion is certified, and how this interacts with trained models.
A checklist
- Authority confirmed
- Dataset defined in a schedule
- Permitted and prohibited uses listed
- Exclusivity defined
- Model and derived data rights after termination
- Privacy measures and roles
- Warranties, indemnities and caps
- Audit and provenance
- Payment and acceptance
- Deletion and certification
Related reading
How Maevar helps
Maevar introduces record owners to vetted buyers. Before any introduction, we look at fit, who has authority, what restrictions apply and what the permitted use needs to cover. The buyer pays our introduction fee. The owner and the buyer agree the licence terms directly. Discuss your use case.
Special issues for AI
Memorisation and output
Models can sometimes reproduce training data. The agreement may need to address what happens if outputs closely reproduce licensed content, and what technical measures the licensee must take.
Benchmark publication
If the data is used for evaluation, can results be published? Can examples be quoted in papers?
Fine-tuned models for third parties
If the licensee fine-tunes models for its customers, does the licence extend to those customers' use?
Future models
Does the licence cover use in future model generations, or only those trained during the term?
Negotiation tips
- Agree the permitted use before discussing price, because it drives value.
- Use schedules for technical detail, so the main terms stay readable.
- Tie payments to accepted deliveries.
- Agree a process for removing records later, for example if a rights issue emerges.
- Keep the provenance documentation as a contractual deliverable.
Governing law and disputes
Agree which law governs the contract and how disputes are resolved. Cross-border licences especially need to say this clearly.
FAQ
What is a data licensing agreement?
A contract granting permission to use data for defined purposes, while ownership stays with the licensor.
Does a licence cover models trained on the data?
Only if it says so. Check the termination and derived-data clauses carefully.
Can licensed data be shared with contractors?
Only where the agreement allows it, usually with flow-down obligations.
What happens if the licensor didn't have the rights?
Warranties and indemnities set out the remedies, but prevention through due diligence is better.
Do I need a lawyer?
Yes. Data licensing for AI raises complex IP and privacy issues.