AI is increasingly being embedded across almost every part of an organisation. From software development and customer service to marketing and product innovation, organisations are investing heavily in, and becoming heavily reliant on, AI tools and systems.

But as AI adoption grows, so does a familiar pattern: the most valuable assets are often not the technology itself rather the real value lies in the underlying data, know-how, business intelligence and outputs that flow into and from AI systems. However, many AI systems are still being procured using contractual frameworks designed for traditional software or SaaS. As a result, organisations can overlook provisions or issues that have a significant impact on the value they ultimately derive from their AI investments. 

In that context, key areas that deserve particular attention when reviewing and negotiating AI contracts include: 

Who owns the outputs? 

One of the first questions any organisation looking to procure an AI system will ask is: who owns the AI-generated output? It is also one of the most misunderstood.

Many organisations will assume that because they pay for the service and provide the prompts, they automatically own everything the AI produces. In practice though, ownership and usage rights are determined by the underlying contract. Depending on the supplier’s terms, a customer may receive anything from full ownership to a limited right to use the output.

Intellectual property ownership provisions in traditional software and SaaS contracts are not always an easy fit in this context as those usually reflect a binary approach that the supplier owns their software, which is then used by a customer under license. As such they fail to address the nuances of AI system use, which often involve multiple strands (input, output, prompts, training etc.) each of which require separate contractual consideration as to ownership and usage rights.

Even where ownership sits with the customer, it is important to understand whether the supplier retains any rights that could restrict commercial exploitation. This becomes particularly important where AI-generated content forms part of a product, service or wider intellectual property strategy. It is also worth looking beyond the initial output. Rights relating to modifications, derivative works and future developments can ultimately become more valuable than the original AI-generated material.

Who checks the outputs?

Very few AI suppliers will be willing to provide unconditional guarantees that the outputs provided by their system are correct, legally compliant etc. On the flip side, every regulator has made clear that it will not accept use of AI as an excuse for non-compliance.

It is therefore critical for customers to consider what contractual comfort is provided as to the measures a supplier has put in place to mitigate the risk of misleading, incorrect or non-compliant output as far as is reasonably possible – for example, if an AI powered agent is being procured to interact with consumers, has that been trained to respect mandatory consumer protection laws in the territories in which the relevant consumers are based? If an AI system will be used to collate and summarise customer reviews, has that been trained to understand the rules on ensuring those are genuine and not misleading?

Equally important to consider is what audit measures will be put in place (whether by the customer and/or supplier) to identify misleading, incorrect or non-compliant output and what corrective measures will be undertaken where that is identified.

Protecting inputs 

While much of the discussion around AI ownership focuses on outputs, equally (if not more) valuable intellectual property and other rights often sit in the inputs.

Inputs are more than the end user instructions and often include proprietary datasets, internal documents, technical specifications, customer information, business processes and carefully developed prompts.

For many organisations, these assets represent years of investment and form a key source of competitive advantage. Contracts should make it clear that ownership of customer input remains with the customer and that the supplier receives only those rights needed to provide the service. 

Particular care should be taken with contractual provisions that permit suppliers to retain, analyse or repurpose such customer assets. As AI capabilities become increasingly accessible, exclusive access to valuable data may become one of the most important differentiators between competitors.

Do not assume any and all data can be inputted into an AI system

Equally as important as protecting inputs is the need for customers to assure themselves that they hold the necessary rights to input data into any AI system.

Customers will usually be expected to provide a supplier with contractual comfort that they have the right to do so, which should not be assumed particularly in the case of data and information obtained from third party or public sources where there may be restrictions placed on what the customer can do with that data or information that may include specific prohibitions on inputting that into an AI system.

Risk is also particularly heightened in the case of any personal data where customers will need to assure themselves (and have an audit trail that they have done so) that they have a lawful basis for inputting and processing that personal data via an AI system and that any required disclosures have been made to relevant individuals.

Understanding training rights 

Related to protection of customer inputs is the question of to what extent customer assets can be used to train or improve the supplier’s general AI models.

Suppliers will often view the ability to use customer data for general AI training purposes as critical to product development, many customers though will see it quite differently.

Interactions with an AI system can reveal product and service roadmaps, pricing strategies, internal processes, customer insights and industry expertise, which customers would not wish to see competitors benefit from.

Customers should therefore understand precisely what training rights a supplier is seeking, whether information will be anonymised or aggregated, how long it will be retained for and whether opt-out mechanisms or dedicated environments are available. 

Confidentiality and data protection are not tick boxes

AI systems frequently process some of an organisation’s most sensitive information (including personal data), making confidentiality and information security more important than ever.

Traditional confidentiality provisions may not always address AI-specific risks, particularly where prompts, uploaded documents and datasets are retained for analytics, product improvement or model training.

Organisations should understand how suppliers handle confidential information, trade secrets and personal data, including retention practices, access controls, security measures and deletion commitments. 

In the case of personal data, it is important that customers and suppliers do not assume that the standard “controller to processor” provisions seen in many traditional software and SaaS contracts (which assume a supplier does nothing more than process a customer’s personal data in accordance with that customer’s instructions solely for service provision for that customer) are sufficient/appropriate. Dependent on the rights granted to a supplier the relationship may be a more complex “controller to controller” or “joint controller” one and, as required under data protection laws, a customer should always undertake a fact specific assessment of what personal data will be shared and how that may be used and ensure that appropriate and mandatory contractual assurances relating to personal data use are put in place.

Managing change in evolving systems 

Unlike much traditional software, AI systems do not stand still. Models constantly evolve, functionality changes and outputs can differ significantly over time. While updates may improve performance, they can also alter compliance outcomes, operational processes or risk profiles.

Organisations should consider whether contracts include appropriate governance mechanisms, such as advance notice of significant changes, testing opportunities, change management procedures and the ability to roll-back if unexpected consequences should materialise. 

Don’t leave exit planning until the end 

Exit provisions are often overlooked during procurement, despite becoming crucial when a relationship comes to an end.

As for any comprehensive software or SaaS contract, AI contracts should address data export rights, retrieval of outputs, retention periods, deletion obligations and transition support at the outset when the relationship is new and the parties keen to reach agreement. A well-designed and clear up-front exit strategy helps protect both business continuity and the value created throughout the relationship as well as mitigating the risk of disputes at the end of a relationship when the commercial motivation for reaching agreement may be less. 

Final thoughts 

AI contracting is about far more than procuring technology. For many organisations, the real commercial value lies in the data, outputs and intellectual property generated through AI systems. By focusing on output ownership, customer inputs, training rights, confidentiality, change management and exit arrangements, organisations can better protect their commercial interests and maximise the value of their AI investments. In the AI economy, protecting the value around the technology may prove just as important as procuring the technology itself.