The one sentence that changes the AI conversation
TBS published the Guide on the Use of Agentic Artificial Intelligence in May. Buried in the overview is a sentence that should reframe how we all talk about this:
Generative AI may describe what should be done. Agentic AI can attempt to do it.
That is not a technical footnote. It is a governance dividing line.
Here is the part I think gets missed. Agentic AI does not swap one set of risks for another. It inherits every generative concern — accuracy, privacy, fairness, security — and then adds a second set on top.
The new set is about actions rather than content. What did it touch? What permissions did it have? Can we see what it did? Can we undo it? Most of us are far less practised at these, because until recently we did not need to be.
So the honest framing is not "agentic AI is the next thing to assess." It is "agentic AI is everything we were already assessing, plus a category we have barely started on."
The guide's answer is two principles layered on top of FASTER. Bounded autonomy, meaning tight explicit limits on data, tools and scope, with read-only as the default. And recoverability, meaning the thing can be paused, stopped and returned to a safe state, with every action logged somewhere the agent cannot edit.
Both of those are design decisions, not procurement decisions. They have to be made before anyone builds anything, by people who understand the program. Which means the governance conversation cannot wait until there is a tool to assess.
If your department is planning agentic pilots this year, the question is not whether the AI is good. It is whether you can explain, in one sentence, what it is allowed to do and what happens when it does something you did not expect.
Can you?
Guide on the Use of Agentic Artificial Intelligence, TBS:
lnkd.in/gTDH663Q
Views my own.
#GCDigital #AIGovernance #PublicService
ALT Diagram showing that agentic AI inherits generative AI's concerns and adds new ones. A dashed boundary labelled "Agentic AI — attempts to do it, not just describe it" contains two boxes. The first, "Content questions," is inherited from generative AI: accurate, biased, secure. The second, "Action questions," is new with agentic AI: permitted, logged, reversible. Both feed into two principles below: bounded autonomy, meaning read-only by default, and recoverability, meaning pause, log, return safely.
