Risk · R-061 · Agentic & Autonomous Systems
Unauthorised commitments by agents
An agent makes a statement, offer or transaction that binds the organisation, without the authority a person in the same position would need.
- Severity
- High
- Audience
- Board · Heads-of
- Frameworks
- Article 14 · PRIN 2A.5 · ASI03
- Responsible AI domain
- Lifecycle stage
- Reviewed
- 2026-11-10
What it is
Organisations control who may commit them: signing limits, delegated authorities, approved wording. An agent that converses with customers or transacts with counterparties can make a promise, quote a price, accept terms or waive a fee. Whether the organisation is legally bound depends on the facts and the jurisdiction; the practical exposure is that customers rely on what they were told and the organisation is held to it.
How it happens
- The agent's instructions describe its task and say nothing of its limits.
- A customer negotiates and the agent agrees in order to be helpful.
- The agent applies a policy that is out of date or that it invented.
- Authority limits that bind staff were never translated into limits on the agent.
Why it matters
The exposure is financial, legal and reputational, and it scales with the number of conversations. A disclaimer that the agent may be wrong carries little weight where the customer had no means of checking.
Warning signs
- No statement exists of what the agent may and may not agree to.
- Transcripts contain undertakings such as refunds, rates or deadlines.
- Complaints cite what the assistant said.
- The agent can complete a transaction that a member of staff at that level could not.
Who owns it
The business owner of the channel is accountable, as for a team of staff. Legal and compliance set the limits. Engineering enforces them in the agent's permissions and approval steps.
In practice
An insurer's assistant, pressed by a customer about a late renewal, states that cover will be backdated. No such concession exists in the policy, and the customer relies on it when a claim arises. The same pattern generalises to any agent that speaks or transacts for an organisation.
Mitigating controls
Primary controls
- Approval gates for high-impact agent actions Preventive
- Requires human approval before an agent makes a binding statement or transaction. Before an agent takes an action above a defined impact threshold, or one that cannot be reversed, a designated person reviews the exact action and approves or rejects it.
- Intended-use definition and use restrictions Directive
- Defines what the agent has authority to do. The approved purpose, users, population and conditions of use of each AI system are stated precisely, and any use outside them requires reassessment before it begins.
Supporting controls
- AI contract terms Directive
- Terms of use can limit reliance on agent statements. Contracts with AI suppliers cover data use, change notification, information and audit rights, sub-contracting, incident notification, liability and exit.
- Review of AI-generated communications Preventive
- Reviews what agents tell customers. Customer communications produced by AI are checked against content standards before use, in full or by sampling according to risk, and required wording is enforced.
Framework and clause references
| Framework | Clause | Title |
|---|---|---|
| EU AI Act | Article 14 | Human oversight Oversight of consequential actions. |
| FCA Consumer Duty (PRIN 2A) | PRIN 2A.5 | Consumer understanding outcome Customers rely on what they are told. |
| OWASP Top 10 for Agentic Applications | ASI03 | Identity and Privilege Abuse Authority exceeded. |
Sources
- EU AI Act: Regulation (EU) 2024/1689 European Union (EUR-Lex) · official text
How to cite
Axiom Verity (2026). Unauthorised commitments by agents. Reference Library, R-061. https://axiom-verity.com/risks/unauthorised-commitments-by-agents. Reviewed 2026-11-10.
Found an error or a better source? Report a correction.