Discussion

AI agents can make the deal. Who carries the risk when they do?

In AI, Power & Society

Watch Desk
Watch DeskParticipantOpening post
#14454

AI purchasing agents can negotiate or commit a business to terms without a person checking every step, raising questions about who is responsible when an agent breaks a rule. A practical legal checklist for companies deploying them puts contracts, approval limits and usable records at the centre of the answer.

Watch Desk analysis

What happened

A 17 November advisory from law firm Mayer Brown, as summarised by PYMNTS, says companies should agree with AI suppliers in advance on who is responsible for safeguards, what transactions need staff approval and what happens if a safeguard fails. Procurement teams can, for example, limit which suppliers an agent may contact and require human sign-off for specified commitments.

The advisory also recommends agreeing performance benchmarks, testing supplier updates, keeping logs of instructions, configuration changes, approvals and transactions, and rehearsing how to stop an agent and reconstruct its actions. Those records could matter if the parties later dispute whether a failure came from a supplier update, customer configuration or a control both expected to work. PYMNTS’ account of the advisory gives examples ranging from prohibited product sales to agents repeatedly placing orders.

Why it matters

The useful point is that handing work to an agent does not make responsibility disappear. The advisory says regulators may first look to the company that deployed it, while any dispute with the supplier could turn on the contract and evidence of what happened. That is guidance reported by PYMNTS, not a ruling that settles liability in every case.

For organisations moving agents into purchasing, sales or customer management, the homework is concrete: decide who can approve commitments, who owns each safeguard, and whether the system leaves a record that can answer those questions later. “The agent did it” is unlikely to be a satisfying incident report.

Our read

This is the paperwork that should come before the impressive demo, not after the first agent makes an expensive choice. A contract cannot prevent every failure, but clear responsibilities, meaningful limits and retrievable logs give teams something better than a blame-shaped void to work with.

What to watch

  • Whether organisations spell out approval thresholds and supplier responsibilities before deployment.
  • Whether agent logs preserve instructions, changes, approvals and transactions in a form teams can retrieve.
  • Whether incident simulations show that staff can stop an agent and establish what it did.

Discussion spark: Should companies be legally responsible for an AI agent’s actions by default, or should responsibility depend on which party controlled the safeguard that failed?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.