Agentic AI: When Technology Starts Making Decisions
AI is moving beyond providing answers. Increasingly, it's being given permission to act.
It also exposes a governance question: are the controls designed for systems that advise us sufficient for systems authorised to act?
An AI assistant might recommend a course of action. An AI agent can potentially carry it out: accessing information, using software, initiating transactions or interacting with other systems.
The opportunity is considerable. So is the responsibility that comes with it.
When a mistake becomes an action
Traditional AI risks often centre on inaccurate or misleading outputs. With agentic AI, an incorrect response may become the starting point for a series of actions.
An agent could misunderstand an instruction, act on unreliable information or access systems beyond what its task reasonably requires.
Connect several agents and the consequences can become harder to anticipate. A problem in one part of a workflow may affect another before anyone notices.
The UK's National Cyber Security Centre has highlighted the need to think carefully about agentic AI adoption, including the security implications of greater autonomy.
The concern isn't that every AI agent is inherently unsafe. It's that the consequences of a failure depend on what the agent has been allowed to do.
Governance must extend beyond the model
Governing agentic AI means looking beyond the individual model. Organisations must consider the entire workflow: the tools an agent can use, the data it can reach, the actions it can take and the consequences when systems interact.
An agent preparing a report does not need the same permissions as one authorised to update records, execute code or communicate externally.
Before deployment, organisations need clear answers:
- What can the agent access and change?
- Which actions require human approval?
- How will unexpected behaviour be detected?
- Who can intervene and stop it?
These decisions need clear ownership, enforceable limits and ongoing monitoring. A governance policy is of little value if it cannot be applied to what an agent actually does.
Human oversight also needs to be proportionate. Requiring approval for every routine action could defeat the purpose of automation. Allowing consequential actions to proceed without effective safeguards creates a different problem.
The balance depends on the potential impact of getting something wrong.
Accountability cannot be automated away
I've spent more than 30 years working with senior technology, cyber security and risk executives. Whatever the technology, clear responsibility remains fundamental to effective leadership.
Agentic AI doesn't change that principle. It makes governance more demanding: accountability must remain clear even when actions are automated.
Security guidance from OWASP reinforces the need to limit agent permissions, monitor behaviour and establish controls for high-impact actions.
The challenge for senior executives is not simply deciding where AI could improve efficiency. It is ensuring governance develops alongside the technology, rather than after an incident.
It's deciding how much autonomy the organisation is prepared to grant, what controls must accompany it and who remains answerable for the outcome.
As AI gains the freedom to act, governance must keep pace with the authority we give it.
Sean Farrell
Founder, MERCERBRIDGE
Sean founded MERCERBRIDGE in 2006 and brings more than three decades of specialist Executive Search experience.
Discussing leadership and technology risk?
MERCERBRIDGE works with organisations appointing senior leaders across Technology, Cyber Security, Risk and Data & AI Governance.