TerminologytermStep 5: Assist vs actAllRiskOperations

Term: Agent / agentic AI

~7 min read

Estimated time: ~7 min read — for the in-app brief plus opening the primary source.

What this is

An agent is AI that can plan and take multi-step actions across tools (email, tickets, CRM) — not only draft text in a chat box.

Everyday example

Not just 'draft a reply,' but: read the inbox, open the CRM, update the opportunity, schedule a meeting, and email the customer — with or without asking you at each step.

An agent takes multi-step actions toward a goal — browse, call tools, update systems — not just draft in a box.

  • Higher upside and higher blast radius than a copilot.
  • Needs permissions, logs, spend limits, and a kill switch.
  • “Agentic” on a slide often means a scripted workflow with an LLM in the loop — ask what it can actually touch.
  • Customer-facing agents inherit your brand and your liability.

Next action: Do not approve an agent that can write to a system of record without a named owner and a rollback.

What changes in how you lead

How decision rights, process, and ownership should change.

  • Treat agents as operators with credentials, not as chat windows.
  • Risk, Legal, and the process owner sign the permission set together.

Compare related ideas

Agent / agentic AI vs Copilot

If a human must approve each external action, you are closer to copilot. If the system executes a chain of tools autonomously, treat it as an agent and govern it like automation, not like a spellchecker.

Open Copilot

Deep dive

Opportunity: automate routine multi-step workflows.

Risk: wrong action at speed; over-permissioned access; hard-to-audit chains.

Start with reversible internal tasks; require step-up approval for money, external messages, or deletes.

Labs and vendors increasingly brand features as 'agents' — demand a clear action list and permission model, not a buzzword.

Related terms

Related weekly lessons

terminologyagentsautomation