AI Agreements: A Practical Risk Guide for Tech Businesses
Artificial intelligence is now embedded in customer platforms, internal tools, analytics engines and decision systems.
When AI is procured, licensed or deployed, the legal risk sits primarily in the contract.
Traditional software terms are not enough. AI introduces specific risks around data, IP, explainability, bias and liability. These must be addressed expressly in the agreement.
This article sets out the key contractual risk areas and what to negotiate.
1. Intellectual Property: Who Owns What?
AI systems typically involve:
The underlying algorithm
Documentation and interface
Training data
Customer input data
Output generated by the system
Each layer must be addressed separately.
Key risks
Ownership of the AI system
The vendor usually owns the algorithm and platform. Confirm they have sufficient rights to license all third-party components (including large language models).
Customer input data
Customer input data is typically pre-existing IP owned by the customer. The contract must:
Confirm customer ownership
Limit vendor licence rights
Align licence scope with privacy and confidentiality terms
AI outputs
Ownership of AI-generated output is legally uncertain in many jurisdictions because copyright usually requires human authorship.
Practical position:
Allocate contractual ownership regardless of copyright uncertainty
Consider confidentiality protections if IP protection is unclear
Require substantial human modification for critical deliverables
Improvements and “learning”
AI systems evolve. The contract must address who owns improvements resulting from:
Customer data
Customer training
Ongoing system use
Common models:
Vendor owns general improvements
Customer retains rights in high-value data-driven improvements
Layered ownership structure
Action: Do not rely on generic IP clauses. AI requires layered IP allocation.
2. Data Use, Privacy and Confidentiality
AI systems rely on large datasets. This raises privacy and confidentiality exposure
Customer training data
Contract should address:
What data the AI can access
Permitted purposes of use
Whether vendor can retain data
Whether vendor can use it to improve models
Vendors often seek broad rights to use data for product improvement. This must be assessed against:
Customer contractual obligations to its own clients
Privacy law compliance
Commercial sensitivity
Historical / third-party training data
Seek warranties that training data:
Was lawfully obtained
Complies with privacy obligations
Does not breach confidentiality
Generative AI use
If using third-party generative AI tools internally:
Avoid uploading confidential or personal information without clear contractual and technical protections
Confirm vendor does not store or reuse prompts beyond service delivery
Implement manual review of outputs
Action: Align licence rights, privacy compliance and confidentiality obligations. They must operate together.
3. Explainability and Audit Rights
AI systems can operate as “black boxes”.
If AI is used in decisions that:
Affect individuals
Require regulatory justification
Must be explained to customers
then transparency provisions are critical.
The contract should address:
Whether decisions are explainable
Access rights to explanations
Cooperation with regulatory investigations
Record-keeping obligations (especially in cloud environments)
Suppliers may resist full transparency due to trade secrets. A balanced position is often:
Outcome explainability
Audit rights limited to compliance purposes
Action: If you must explain decisions, your contract must allow you to.
4. Acceptance, Testing and “Learning” Risk
AI evolves. Acceptance testing is not always a single event.
We suggest considering:
Proof of concept or trial phases
Staged acceptance as the system evolves
Performance metrics for evolving systems
“Circuit breakers” allowing suspension if bias or non-compliance occurs
Action: Define measurable objectives. Tie payment and continuation rights to performance.
5. Warranties and Indemnities
AI systems often include “as-is” disclaimers, particularly for outputs
That may not be commercially acceptable.
Vendor warranties to consider
Performance warranties
Algorithm design assurances
Lawful use of training data
Non-infringement of third-party IP
Steps taken to mitigate bias
Customer warranties
Vendors typically require:
Customer has rights to input data
Customer data does not infringe third-party rights
Indemnities
An indemnity means one party agrees to pay if certain losses or claims happen.
Customers commonly seek indemnities for:
Privacy breaches
Confidentiality breaches
IP infringement
Vendors often seek indemnities relating to:
Customer misuse
Third-party claims arising from customer data
Action: Match indemnity structure to actual AI risk profile.
6. Bias, Automated Decision-Making and Risk Allocation
AI risk profile depends on use case.
Risk increases where AI:
Automates decisions affecting individuals
Operates without human review
Generates public-facing content
Processes sensitive data
It may be helpful to identify:
Primary operator
Obligations for monitoring and maintenance
Liability allocation for misuse or flawed training data
Action: If automated decisions create real-world consequences, liability caps should reflect that.
7. Limitation of Liability
A limitation of liability sets a cap on how much a party can be responsible for if something goes wrong. i.e. this is the maximum amount you can claim from me.
Standard liability caps may not reflect AI risks.
Unique AI risks include:
Hallucinated or inaccurate outputs
Third-party IP claims
Privacy breaches
Model degradation from training data
We suggest reassessing traditional caps in light of:
Damage to third parties
Damage arising from evolved systems
Action: Review caps against realistic exposure, not standard SaaS templates.
8. Termination and Exit
AI systems embed deeply into operations.
On termination, consider:
Ongoing rights to input data
Ongoing rights to output data
Portability to a new supplier
Whether customer data can be removed from the system
Escrow viability
AI vendor lock-in risk is higher than standard software due to model training and data entrenchment.
Action: Address portability and data extraction before signing.
9. Governance and Regulatory Change
AI regulation is evolving.
We recommend considering governance or review mechanisms if law changes materially.
Suppliers may resist blanket compliance with future laws, but structured review clauses are commercially reasonable.
Action: Include review mechanisms tied to regulatory change.
Final Observations
AI risks can be managed by the terms in the contract.
Organisations relying on AI should:
Separate IP layers clearly
Align data licence with privacy obligations
Secure explainability where required
Tie performance to measurable outcomes
Reassess liability caps
Address bias and automated decision-making risk
Plan for exit and portability
AI contracts require deliberate drafting. Standard SaaS templates are rarely sufficient.
For small, mid-sized and enterprise organisations operating in technology environments, these issues should be assessed before procurement, not after deployment.
Free AI Contract Checklist
Want a free copy of the AI Contract Checklist? Email hello@pixellegal.com.au
Further information
If you need support with drafting or reviewing an agreement that contains AI related services, email hello@pixellegal.com.au or organise a free consultation.
Disclaimer
This article is provided for general information purposes only and does not constitute legal advice. It does not take into account your organisation’s specific circumstances, systems, or regulatory obligations. You should obtain tailored legal advice before entering into a contract for AI related services.
