The integration of AI into third-party service delivery represents a significant shift for the financial services sector. While attention has focused on FS businesses deploying AI directly, its use by suppliers in the background is discussed less often, yet is an important topic for practitioners advising banks, insurers, investment firms and other regulated businesses.
The nature of the customer-supplier relationship changes when a supplier relies heavily on AI. The FS business still purchases an outcome, but the mechanism by which it is achieved carries risks that traditional services contracts were not necessarily designed to address.
A further development is agentic AI, which is capable of autonomously executing multi-step business processes previously performed by human teams. Where an AI agent reasons, plans and acts across a workflow without human intervention, the supplier relationship is neither a straightforward software-as-a-service arrangement nor a conventional managed service, but a hybrid sitting between these established contracting frameworks.
This note provides a practical framework for identifying AI use in supplier delivery, conducting tailored due diligence, and addressing critical contractual provisions.
Step 1: classify the AI: tool or service component?
The first question is whether AI is an incidental tool of the supplier's trade or a core component of service delivery. Where AI is merely a tool and the FS business is comfortable with the supplier's overall operational resilience, it may not need to understand which tools the supplier uses to deliver services. However, where reliance on AI is so fundamental that the AI provider would be considered a subcontractor, or where AI is used both to execute work and to provide quality assurance and oversight, contractual commitments are called for. The FS business can no longer assume that the supplier's checking processes provide an independent safeguard.
Practitioners should advise clients to apply a risk-based scale when classifying AI, ranging from incidental use through to agentic AI to guide the level of contractual intervention required. This classification must be ongoing, not fixed at a single point in time. Suppliers may change how they use AI over a contract's term, particularly in long-term arrangements where new services added via statements of work may receive less governance scrutiny.
Step 2: conduct AI-specific due diligence
The usual sourcing due diligence may not be sufficient where AI is material to service delivery. Pre-contract due diligence must move beyond technical validation towards checking governance, policy and human oversight. It is unlikely to be necessary to test AI model accuracy where AI is simply part of the supplier’s solution; the FS business should instead focus on testing the supplier’s own quality assurance, for example, how it verifies AI outputs.
Generic assurances of "responsible AI" practices are insufficient where the supplier’s services relate to a critical part of the FS business. The due diligence process should generate specific, verifiable information to inform both the contracting decision and ongoing governance.
Step 3: negotiate key contractual provisions
Performance and Service Levels
A supplier's use of AI can impact service levels. While AI may outperform humans in speed and consistency, there is a risk of unpredictable behaviour if training data changes, which may lead FS businesses to seek AI-specific service levels for accuracy and bias metrics.
Practitioners acting for FS businesses should consider warranties framed around "intended or foreseeable use" rather than detailed technical specifications. The contract should also include a model change management framework: a minimum notice period before model or version changes, access to a staging environment for revalidation, and the right to defer go-live where the change presents unacceptable risk.
Data Use and Protection
If an FS business’s data contributes to improving a model that subsequently serves competitors, it has in effect subsidised the erosion of its competitive position. The use of customer data for training also creates data protection issues where identifiable individuals are concerned. The contractual default should be an explicit prohibition on using customer data for AI model training, whether by the supplier or any third-party provider in the supply chain.
The contract should require the supplier to use only approved AI tools operating under enterprise agreements that exclude customer data from model training and limit retention of prompts and outputs. A blanket prohibition on inputting the institution's data into consumer-facing AI interfaces should be explicit. The FS business should also have the right to approve or reject specific AI tools and to receive notification of changes to the supplier's AI tooling or policies.
Security and Incident Notification
Where the supplier's AI operates outside the direct visibility of the FS business, the contract should define specific notification obligations covering material AI model failures, instances where incorrect AI-generated outputs have been incorporated into deliverables, suspected exposure of customer data through AI systems, and any change to the supplier's AI tooling or oversight model that could affect service quality or risk profile.
Notification timelines should reflect the FS business’s own operational resilience reporting obligations and the expectation that regulators may require it to account for AI-related incidents in outsourced functions.
Audit Rights and Governance
Audit rights are often the subject of negotiation in services arrangements. Much like the early days of cloud services contracts, practitioners are negotiating with suppliers on the extent of oversight and access. The key principle is that contracts where AI is used by the supplier to deliver services need to provide the FS business with appropriate oversight rights to satisfy regulatory requirements.
Oversight in this context often requires access to model governance artefacts, validation summaries, or drift and monitoring results, rather than model code or full datasets. The FS business may also require rights to inspect how the supplier supervises, tests and updates AI tooling and transparency regarding human-in-the-loop controls. Suppliers will typically resist disclosure of proprietary algorithms, but that may be acceptable since what matters is assurance of safe and controlled AI operation in the context of the service delivery.
Contractual audit rights must extend beyond standard information security reviews to include confirmation that customer data has not been incorporated into the supplier's or any third party's AI model.
Liability
The case for enhanced liability provisions is based on the risk that a failure in the underlying AI system could be systemic rather than isolated, could be harder to detect and could intensify data-related liability. Where the contract imposes a strict ban on using customer personal data for model training, a breach may justify a higher liability regime with a reduced range of exclusions and/or a higher or uncapped liability limit.
Suppliers working with frontier AI models will themselves be subject to terms that frequently disclaim liability for output accuracy, IP infringement and consequential loss. FS businesses should require transparency on the supplier's upstream terms and should not accept a position where the supplier's liability to the FS business is narrower than the upstream protections the supplier has secured. If the supplier cannot accept appropriate liability provisions due to its upstream contracts, the FS business may need to consider whether it is the right supplier for them.
Operational resilience and exit
Operational resilience is a key focus area for FS regulators. FS businesses should consider whether the supplier has defined a fallback operating model that enables continued service delivery within agreed impact tolerances when the AI element is offline, whether through deliberate suspension or uncontrolled outage.
FS regulators also focus on exit, seeking to ensure that FS businesses can facilitate a managed exit without excessive disruption to end customers. Practitioners negotiating exit provisions should address what the FS business needs to transition effectively, including access to prompt libraries, workflow configurations, evaluation datasets and documentation of how AI-supported processes operated, even where the underlying model remains within the supplier's control.
The contract should also require the supplier to retain decision logs for a defined period following termination, to support any regulatory inquiry or look-back exercise.
FS businesses should consider whether the additional complexity of migrating from a supplier in an AI context requires an extended exit period. Many would rather pay for enhanced exit support than risk a chaotic exit threatening their ability to service end customers.
Conclusion
Traditional services contracts assume human-scale performance and traceable decision-making. AI disrupts these assumptions, requiring a rethink of pre-contractual due diligence, contract structure, and ongoing oversight.
The regulatory overlay compounds the challenge. FS businesses face compliance obligations across multiple overlapping frameworks, including PRA SS2/21, the FCA operational resilience framework, SM&CR, UK/EU GDPR and, for those with EU-facing operations, the EU AI Act. These regimes intersect with the supplier relationship, and the services contract is the primary mechanism for managing that multi-regime exposure.
Practitioners advising regulated businesses should treat the contractual framework not merely as a commercial document but as the principal risk management instrument for AI-enabled supplier relationships. Getting it right at the outset and keeping it under review as the supplier's AI use evolves is now an essential discipline in financial services outsourcing practice.
.jpg&w=3840&q=75)
Copyright & permissions
.jpg&w=3840&q=60)
