An AI assistant helps a person: it answers, summarises or drafts, and the person decides what to do. An AI agent acts: it takes steps in your systems toward a goal, sometimes without a person approving each one. The technology overlaps heavily. The difference is who performs the action, and that changes accountability, controls and what an incident looks like.
What does each one actually do?
An assistant works turn by turn with a person. Ask it to summarise a contract and it returns a summary. Ask it to draft a reply and it drafts. Every effect on the outside world passes through the person, who copies, edits, sends or ignores.
An agent receives a goal and works toward it. Asked to resolve a customer's complaint, it reads the case, checks the account, drafts a response, perhaps issues a credit and logs the outcome. It may finish before any person looks.
Why does the difference matter so much?
Because control changes hands. With an assistant, the person is the control: they read the output and decide. If the assistant is wrong, the person is the last line of defence, and usually catches it.
With an agent, that natural control is gone for any step the agent performs on its own. Something else has to decide whether each action may happen, and something has to record it. That is why an agent that is technically similar to an assistant needs a quite different set of controls.
Where does the line actually fall?
The boundary is not the product label. It is what the system can do without a person in between.
| Capability | Assistant-like | Agent-like |
|---|---|---|
| Answers questions | Yes | Yes |
| Drafts for a person to send | Yes | Yes, at depth 2 |
| Sends messages on its own | No | Yes, at depth 3 within bounds |
| Changes records | No | Yes, at depth 3 within bounds |
| Moves money or takes hard-to-undo actions | No | Only if certified at depth 4 |
| Coordinates other agents | No | Orchestrators, at depth 5 |
An agent held at depth 2 behaves much like an assistant: it drafts, a person acts. Many agents start that way and earn higher depth as evidence builds.
What does this look like in practice?
Take a relationship manager at a private bank. With an assistant, they ask for a summary of a client's portfolio changes and a draft email explaining them. They read both, correct a figure, and send the email themselves.
With an agent, the same request becomes a task: "prepare and send the quarterly update to these twenty clients." The agent gathers the data, drafts each email and, if certified at depth 3 for routine updates, sends them. One email contains a figure from a stale data feed. With the assistant, the relationship manager would have caught it. With the agent, the controls have to: the platform re-reads action-critical figures just before sending, and updates containing figures outside a tolerance are held for the relationship manager.
Assistant incidents are usually about content: a wrong answer, a poor draft, a leak in a response. A person was in the loop, so the harm is limited by what they let through.
Agent incidents are about actions: a refund issued, a record changed, a message sent, often across many cases before anyone notices. They need a different response, which is discussed in agent incidents look different from system incidents. The two neighbouring distinctions, from rules-based automation and from designed workflows, are covered in how an AI agent differs from workflow automation and how an agentic workflow differs from an AI agent.
What stays with a person?
With an assistant, almost everything. With an agent, whatever the bounds say: actions above its depth, above the business's limits, resting on unverified claims, or irreversible. A person also owns each agent's role and answers for its outcomes.
Where Pryme fits
Pryme governs agents, not assistants. Each agent is certified on your own cases and given a depth from 1 (read) to 5 (coordinate other agents). An agent at depth 2 drafts and a person sends; higher depths are earned, expire after 90 days, and drop the same day if a certification run clears no depth. The full design is in the governed agentic operating system guide.
Questions people ask
Can an assistant become an agent?
Yes, by giving it tools that act and removing the person between output and effect. That step is where governance requirements change, so treat it as a new deployment, not a feature update.
Is an agent always better than an assistant?
No. For work where a person must exercise judgement on every case, an assistant may be exactly right. Agents pay off where routine actions can be bounded and evidenced.
Do customers need to be told which one they are dealing with?
Where people interact with an AI system, transparency rules such as the EU AI Act's Article 50 can require telling them. That applies to assistants and agents alike.