Agents with accountable authority

Machines will trade with machines

Digital tools have long helped people compare prices, place orders and manage suppliers. Newer intelligent agents may undertake several steps on a person’s or organisation’s behalf: discover opportunities, evaluate alternatives, request information, negotiate within limits and initiate transactions.

That capability promises to reduce routine work. But it creates two linked questions: Can an agent work beyond a single proprietary network? And who is answerable for what it does?

An agent restricted to its provider’s ecosystem may reproduce the same gatekeeping that trapped earlier generations of commercial software. An agent free to act anywhere without clear authority, however, could make damaging commitments or expose information. Free Commerce therefore joins interoperability with governance.

The mandate

Imagine a buyer assigns an agent to source replacement office equipment. It can compare qualified suppliers and request quotes, but it may not spend more than an approved amount, share confidential employee records or accept terms outside the organisation’s policy. A human must approve any purchase beyond the threshold.

That is a mandate: a description of who the agent acts for, which operations it may perform, the limits of those powers and when it must stop or escalate. Technical systems may represent the mandate differently, but the underlying obligations remain.

A practical mandate should specify its principal, purpose, permitted actions, budget or value limits, eligible counterparties, data permissions, duration, approval triggers and revocation process. Without those details, ‘the AI decided’ can become an excuse instead of an explanation.

The audit trail

Every important act should be traceable: which authority was presented, what information was used, which version of the policy applied, who approved an exception, what was sent to the other party and what outcome occurred. Records must themselves respect security, privacy and retention obligations.

Traceability does not mean every model’s internal reasoning is perfectly interpretable. It means organisations should be able to reconstruct consequential actions and identify who authorised them, within the limits of the technologies involved.

What happens when rules collide?

A supplier’s agent offers a discount if the buyer confirms an order immediately. The buyer’s agent estimates that the deal is attractive, but the offer requires agreeing to unfamiliar liability terms. The buyer’s policy forbids that. A properly constrained agent does not quietly trade policy compliance for apparent savings. It pauses, declines or requests human approval under the approved process.

The same principle applies to uncertain identity, expired credentials, ambiguous instructions and attempts to manipulate an agent with data embedded in documents. The content of an external quote is evidence for commercial evaluation, not an instruction that outranks the principal’s mandate.

Open agents and open markets

Interoperability means the supplier should not need to run its agent on the buyer’s preferred platform in order to negotiate. Both parties should be able to establish identity, communicate offers and exchange records through agreed protocols while retaining their own systems and policies.

Shared standards create the possibility of those interactions. They do not by themselves guarantee safety. The surrounding implementation must enforce access controls, transaction constraints, dispute handling and applicable law.

Try it: write a safe mandate

Draft a procurement agent mandate in eight lines: principal; task; authorised information; allowed suppliers; maximum spend; actions allowed without approval; conditions for escalation; expiration/revocation. Then add a sample event log entry showing who authorised an action and what the agent actually did. Review the mandate for one loophole and close it.

Key terms

• Agent: Software acting towards a defined goal on behalf of a principal.
• Principal: Person or organisation whose authority the agent exercises.
• Mandate: Boundaries of delegated authority.
• Escalation: Transfer of a decision to an authorised human or process.
• Audit trail: Records that support reconstruction and accountability for an action.

Remember

The future is not simply autonomous commerce. It is accountable, interoperable commerce in which people and organisations remain responsible for the authority they delegate.