An artificial intelligence purchasing agent tasked with finding the lowest price could agree with agents working for other buyers to pay suppliers no more than a set amount. That could expose the company deploying the agent to antitrust scrutiny and leave it in a separate fight with its AI supplier over who should have prevented the agreement.
The hypothetical scenario illustrates what is at stake as companies give AI agents more authority over everyday business decisions, according to a Tuesday (Nov. 17) advisory by law firm Mayer Brown. Agents can execute transactions, change records and communicate with customers and suppliers without a person reviewing each step. Regulators are expected to look first to the company that deployed the agent. Whether that company can recover losses from its supplier may depend on what the contract assigns to each party and what the company can prove about how the agent acted.
The advisory offered a retail example in which a sales agent could encourage New York shoppers to buy scooters that exceed the applicable speed threshold. The retailer could face enforcement action even if it argued that the vendor should have built a safeguard against prohibited sales. The supplier might point to the retailer’s sales instructions, its configuration of the agent or its control over which transactions require approval.
Before rollout, companies can use these scenarios to agree with suppliers on who runs each safeguard, the advisory said. For a retail agent, the agreement could specify who maintains jurisdiction-specific product restrictions, who tests that the agent blocks prohibited transactions and who can suspend it when a control fails. Procurement teams can set limits on whom their agents contact and decide which commitments require staff approval. The contract can spell out who covers losses if an agreed safeguard fails and whether limits on payments apply.
Companies may also have to explain an agent’s decisions to a regulator, according to the advisory. Audit rights, verification tools and logs that document decision steps are key safeguards. Companies should agree with vendors in advance on which records to keep and how quickly teams can get them, including instructions, configuration changes, approvals and transactions. These records could help establish whether the incident followed a supplier update, a change made by the customer or a failure of a control both parties expected to work.
Performance disputes may turn on what a vendor promised and how the contract measures it, the advisory said. An agent that repeatedly places the same order could prompt a gross negligence claim aimed at recovering more than a damages cap allows. A customer might seek early termination if a customer relationship management agent routinely misses meetings or spams contacts. A supplier update that reduces accuracy could lead to a dispute over good faith. Companies can set performance benchmarks before launch, test vendor updates and agree on when repeated failures justify ending the contract.
Third-party injuries can also give rise to disputes, according to the advisory. Third parties may sue over misleading statements to customers or disparaging claims about competitors. If an agent visits a site that bars automated access, the parties may dispute who should have blocked it. Similar allocation disputes may follow a privacy or security incident if an agent discloses sensitive information. The parties may disagree over whether supplier safeguards or the customer’s deployment controls should have prevented it.
Intellectual property disputes can arise during an agent’s acquisition of third-party material, ownership of new code it produces, or a supplier’s alleged use of customer data to train a competing product, the advisory said. A customer may struggle to gather enough evidence of unauthorized training to bring a claim and may still have trouble identifying its data in a training set. Contract terms on ownership and data use are stronger when paired with a way to verify compliance.
Before launch, operations and compliance teams should walk through a simulated incident with the supplier, according to the advisory. The exercise can show who can stop the agent, how quickly the company can reconstruct what happened and whether it has the records it would need to answer a regulator.
For all PYMNTS AI coverage, subscribe to the daily AI newsletter.