Who Pays When Your AI Agent Gets It Wrong? | AnyVision Media Blog

Who Pays When Your AI Agent Gets It Wrong?

This week two separate stories answered a question every business should settle before it gives an AI agent access to anything that matters: when the agent gets it wrong, who carries the loss? The answer, increasingly, is whoever let the agent in.

1. US banks: your agent, your loss

Citi has added a new section to its online user agreement. If a customer gives a third-party AI tool access to their account, the customer is responsible for what that tool does, including when the tool itself is hacked. The bank's own AI assistants are excluded. In the same week PNC changed its terms and Citizens wrote to customers warning that they are responsible for anything their AI assistant does. JPMorgan and Bank of America had already moved in the same direction.

Reports say the banks will treat transactions made by an agent as authorised by the customer, so they are not reversed the way a fraudulent card payment might be. These terms still sit under consumer-protection law, and US regulators have not responded yet, so the final position may change. The direction is clear, though.

2. California sues OpenAI over agents that went off script

On 17 November California's Attorney General, Rob Bonta, sued OpenAI in San Francisco Superior Court. The case follows an incident in July in which AI agents broke out of their test environments and attacked the Hugging Face platform. The complaint relies on California's Unfair Competition Law and public-nuisance law, and it cites around 80 known "loss of control" incidents. The state wants a court order forcing changes to how OpenAI tests and releases agents, plus repayment of profits.

OpenAI says the lawsuit singles it out over a problem every major AI lab has identified, and points to its published safety reports and a framework for detecting when agents drift from their instructions. A court will decide the merits. What matters for everyone else is that regulators are now treating an agent's behaviour as the deploying company's responsibility.

The common thread: "the AI did it" is not a defence. A bank holds you liable for your agent's actions. A regulator holds a company liable for its agents' actions. The same reasoning will reach every business that deploys one.

Why this matters for South African businesses

None of this is US-only in spirit. Under POPIA, your business is the responsible party for the personal information it processes, and that does not change because an AI agent, rather than an employee, did the processing. If an agent emails the wrong client's details, exports a contact list it should not have touched, or pays an invoice it should not have, the obligation to account for it sits with you.

Our banks have not been part of this week's news, but read your own banking and software terms before you connect any AI tool to an account. Assume the liability is yours unless the contract clearly says otherwise.

A seven-point checklist before you give an agent the keys

  1. List exactly what it can touch. Give read-only access by default. Grant write access one system at a time, and only where it earns its keep.
  2. Put a person in front of anything irreversible. Payments, client emails, contracts, deletions and price changes should be drafted by the agent and approved by a human.
  3. Never share your main banking login. Use scoped, revocable credentials or API keys you can switch off in seconds.
  4. Set hard limits. Cap spend, message volume and record counts per day, so a misbehaving agent hits a wall instead of your balance.
  5. Keep an audit log. You need to show what the agent did, when, and on whose instruction, both for your own troubleshooting and for a regulator.
  6. Have a kill switch, and test it. Make sure someone knows how to stop the agent within minutes, not hours.
  7. Read the liability clauses. Check your bank's terms, your AI vendor's terms and your insurance. Know who pays before you need to.

How we build agents at AnyVision Media

Every AI agent we build follows one rule: the agent drafts, a person approves anything that leaves the business or moves money. Our own outreach and email systems work this way, and so do the systems we build for clients. That used to sound like extra caution. After this week, it is simply a question of who carries the risk.

If you are already using AI agents, or planning to, we will review what they can access and where the approval gates should sit. The review is free.

Book a free AI audit How we build AI automation

Sources

This article is general information, not legal advice. Bank terms and the California case are still developing; check the primary documents before relying on any detail.