What it means
The useful line is between reading and writing. Generative AI summarises, drafts and explains, and a person still decides and acts. Agentic AI is handed a goal and carries out the steps itself: it creates the task, updates the deal, sends the reminder. Most products sit somewhere between the two, with an assistant that proposes a change and a person who approves it.
The term is also used loosely for any chatbot that can call a tool. What matters to a buyer is not the label but whether the software changes records without a person approving each change.
Why it matters when you choose software
A wrong answer costs a moment of attention. A wrong action changes a record that other people and systems rely on, sometimes in a way nobody notices for weeks. So the question to ask of an agent is less how clever it is than what stands underneath it.
Three things decide whether its actions can be trusted: whose permissions it runs under, whether every action is recorded in one place, and whether each one can be reversed.
Questions to ask a vendor
- Does the agent act as the person who asked, with exactly their permissions?
- Is every agent action in the same audit log as human actions, and marked as an agent’s?
- Can a single action, or a whole request, be undone?
- Does it show a plan and wait for approval before it changes anything?
- What does it do when a request is unclear, or outside what the person asking may do?
How Cyril approaches it
Cyril has three operating modes. In Platform mode nothing happens that the user did not do. In AI mode, which is in the product, Ask Cyril plans, shows the plan, asks for approval and acts where you can watch, and every change lands in one audit log with an undo. Agentic mode, which hands a goal over entirely, has a first version built and switched off. It opens after launch, not at early access.