Assistant, agent, copilot: sorting out the labels

'Agent' is 2026's most fashionable word, so products of every shape wear it. Underneath, the three terms describe real differences on two axes: how much autonomy the software has, and how wide its scope is. Getting the decode right matters because the label on the box changes what questions you should ask before trusting it.

Working definitions that survive contact with products:

• Copilot — narrow scope, low autonomy. Embedded inside one tool (a code editor, a spreadsheet, a CRM), suggesting and drafting within it. You stay in the driver's seat; it can't act outside its host application.
• Assistant — broad scope, low-to-medium autonomy. Works across your email, calendar, documents, and questions; remembers context; drafts and prepares. Acts beyond advice only with your approval.
• Agent — variable scope, high autonomy. Given a goal, it plans and executes multiple steps — searching, calling tools, retrying failures — with checkpoints rather than per-step approval. An 'AI employee' is an agent packaged around a job role.

For a copilot, ask about quality inside its niche: how good are the suggestions, and does it learn your patterns? Risk is low; the tool can't reach anything else.

For an assistant, ask about memory and privacy: what does it retain about you, where does that live, and who else can see it? Its value and its risk both come from context.

For an agent, ask about permissions and audit: what may it do without asking, what stops it mid-mistake, and where's the log? Autonomy without an audit trail is the single biggest red flag in this market.

Products earn trust in that order, and so should your rollout: let a copilot prove useful, graduate to an assistant with memory, and extend agent-level autonomy only over workflows whose rules you've watched work. Skipping steps is how AI ends up quietly turned off after the first incident.