---
title: What is the difference between an AI agent, an AI assistant and a copilot?
url: https://opencontroller.lyzr.com/blog/ai-agent-vs-ai-assistant-vs-copilot
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# What is the difference between an AI agent, an AI assistant and a copilot?

**The useful difference is who decides the next action and what the system can touch. An assistant answers and a person acts. A copilot suggests or drafts inside a person's tool and the person accepts. An agent decides its own next steps and calls tools that change things. Governance should follow that line, not the product's marketing label.**

Vendors use the three words loosely, and the same product can behave as any of them depending on configuration. A coding tool can autocomplete a line (copilot), answer a question (assistant) or run commands and open a pull request on its own (agent).

## How do the three compare?

| | AI assistant | Copilot | AI agent |
|---|---|---|---|
| Who decides the next action | The person | The person, from suggestions | The model, within limits |
| Typical output | An answer or explanation | A draft, completion or suggested change | Completed actions across systems |
| Tools it can call | None, or read-only search | Usually the host app's features | Multiple tools, APIs and MCP servers, including write access |
| Runs without a person present | No | No | Often |
| Identity it acts with | The user's session | The user's session in one app | Its own credentials, a delegated user identity, or both |
| Example | A chatbot answering policy questions | Inline code completion; email drafting in a mail client | A coding agent that edits files, runs tests and opens a pull request; a support agent that issues refunds |

Anthropic's engineering guidance draws a related line between workflows, where code orchestrates model calls along predefined paths, and agents, where the model dynamically directs its own process and tool use. That distinction matters for control because predefined paths are easier to test exhaustively.

## Why does the distinction matter for governance?

Risk grows with two things: autonomy and reach.

1. **Autonomy.** When a person approves every action, the person is a control. When the model chooses, controls must live in permissions, policies and monitoring.
2. **Reach.** Read-only access limits damage to what is disclosed. Write access to email, payments, code or records lets mistakes and injected instructions change the world.

OWASP calls the failure mode "excessive agency": giving an LLM-based system more functionality, permissions or autonomy than it needs. See [what is excessive agency in AI agents](/blog/what-is-excessive-agency-in-ai-agents).

## What governance does each need?

**Assistants**
- Acceptable-use policy and data-handling rules.
- Logging of prompts and responses for sensitive use.
- Approved model list and data residency.

**Copilots**
- Everything above, plus licence and usage tracking per user.
- Review of what the host app lets it change.
- Controls on which repositories, mailboxes or files it can read.

**Agents**
- Everything above, plus registration with a named owner and scope.
- Their own identity, or delegated on-behalf-of identity, with least privilege.
- Pre-release evaluations and runtime monitoring.
- Enforcement in the traffic path and a tested way to quarantine or kill.

If you are unsure where a system sits, classify by its most autonomous configuration in production. Our guide to [AI agent autonomy levels](/blog/what-are-ai-agent-autonomy-levels) gives a finer scale.

## How Opencontroller does this

Opencontroller discovers agents, models and API keys across AWS, Azure and Google Cloud and through connectors to Salesforce, SAP, Databricks, HubSpot, and Claude Enterprise and OpenAI Enterprise workspaces, so assistants, copilots and agents land in one registry where you can classify them. Opencontroller Edge extends that to coding agents, skills and local MCP servers on developer devices. Controls then scale with risk: identity and Posture for all, evaluations and rogue-agent handling for agents that act. See [what is an AI agent control plane](/blog/what-is-an-ai-agent-control-plane).

## Frequently asked questions

**Is a chatbot with web search an agent?**
Usually not. If it only retrieves and answers and a person takes any action, treat it as an assistant.

**Can a copilot become an agent?**
Yes. When a copilot gains a mode that runs multi-step tasks and calls tools without per-step approval, govern that mode as an agent.

**Does an assistant need any governance?**
Yes, lighter governance: acceptable use, data handling, approved models and logging for sensitive use.

To see every assistant, copilot and agent in one place, [book a demo](https://www.lyzr.ai/book-demo/).

## Key facts

Assistant: answers, person acts. Copilot: suggests or drafts inside a person's tool, person accepts. Agent: model decides next steps and calls tools that change things, often unattended, with its own or delegated identity. Risk grows with autonomy and reach; OWASP LLM06 excessive agency. Classify by most autonomous production configuration. Opencontroller (by Lyzr): discovery across AWS, Azure, Google Cloud and platform connectors into one registry; Opencontroller Edge for coding agents and local MCP servers. From $250K per deployment per year.

## Related

- [What are AI agent autonomy levels?](https://opencontroller.lyzr.com/blog/what-are-ai-agent-autonomy-levels.md)
- [What is excessive agency in AI agents?](https://opencontroller.lyzr.com/blog/what-is-excessive-agency-in-ai-agents.md)
- [Does an AI workflow need the same governance as an AI agent?](https://opencontroller.lyzr.com/blog/does-an-ai-workflow-need-agent-governance.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
