Agentic AI describes systems that pursue a goal across multiple steps, choosing their own actions and using tools along the way, rather than answering a single prompt. The governance question changes with it: instead of reviewing one output, you are deciding in advance which actions an agent may take unsupervised and where a human must approve.
Why it matters for a UK business
It is the difference between a tool that answers when asked and one that acts on its own initiative. That is useful for work that varies each time, and risky for work where a wrong step is expensive and nobody was watching.
What it looks like in practice
An agent is given a goal and a set of tools, and works out the steps itself. It might read the inbox, draft replies, update the CRM and book the follow-up without being told each step. The value is in the autonomy. So is the risk.
What to do about it
Draw the boundary before building anything: what it may do unattended, what needs a person before it proceeds, what it must never touch. The boundary is the design decision. The rest is configuration.
Related terms
Read next
The AI value gap, and where agents fit → · All 22 terms in the glossary · The resource library
If this is the term that has come up in your business and you want it worked through against your own situation rather than in the abstract, that is a conversation, not a page.
Speak to ChrisWritten by Chris Duffy. Last reviewed .
