In brief
- A chat window does not reveal whether the underlying system is a chatbot, copilot or agent.
- A copilot augments a person's work; an agent has more freedom to choose how a task is completed.
- As freedom to act increases, permissions, approvals, logs and stop conditions become more important.
- For many business tasks, a fixed workflow with bounded AI assistance is the most robust design.
What are the three underlying patterns?
A chatbot holds a conversation and provides information or guides a user through a narrowly defined flow. A copilot sits closer to the work context: it may know the open document or current case and suggest wording, analysis or next steps. The person remains in control.
An agent receives a goal rather than only one question. It can assess intermediate results, choose a next step and use available tools. Product names do not always respect these boundaries, so organisations should compare concrete capabilities rather than labels.
| Pattern | Who chooses the next step? | Typical system access | Control |
|---|---|---|---|
| Chatbot | The person asks the next question | Usually read-only access to approved sources | Ground answers and constrain sensitive topics |
| Copilot | The person decides from a suggestion | Context from a document, case or business application | Draft stays visibly editable and is adopted deliberately |
| AI agent | The system chooses steps within a defined objective | Several narrowly bounded read or action tools | Permissions, limits, stop rules and approvals per action |
| Fixed workflow | Software rules define sequence and branches in advance | Predefined interfaces | Deterministic rules, tests and exception paths |
Which pattern fits which task?
Start with the least freedom the task requires. A chatbot with reliable sources is often enough for recurring knowledge questions. A copilot suits demanding professional work in which people review drafts and make decisions. An agent becomes useful only when cases vary substantially and the choice of information or tools cannot be fully predefined.
- Step 1
Clarify the need
Is the aim to answer questions, assist an existing task or coordinate several steps?
- Step 2
Assess variation
Are the route and exceptions known, or must the system choose a route for each case?
- Step 3
Assess consequences
Is an error easy to correct, or can an action have financial, legal or operational impact?
- Step 4
Define control
Which outputs or actions must a person see, change or explicitly approve?
How does accountability change with autonomy?
Accountability does not transfer to the system. The organisation defines its purpose, permitted data, tools and limits. For a copilot, people need to know which suggestions require checking. An agent additionally needs technical limits on actions, amounts, recipients, runtime and spend.
A visible approval is meaningful only if the reviewer has enough context, time and authority. Consequential actions should sit in a separate, auditable approval step rather than being hidden behind a general confirmation.
- Derive rights from the real user or service identity
- Separate reading, drafting and execution technically
- Route ambiguous or risky cases to an accountable person
- Keep auditable records of tool calls, approvals and results
How should the three patterns be evaluated?
A chatbot is assessed for correct, grounded and useful answers. For a copilot, the organisation should also measure whether suggestions genuinely speed up professional work and how often they are corrected. For an agent, the path matters too: Did it choose suitable tools, respect permissions, avoid unnecessary steps and stop when uncertain?
- Test normal cases, exceptions and deliberately unanswerable cases
- Measure output quality and human correction effort together
- Simulate missing tools, system outages and manipulated content
- Retest after changes to models, prompts, data or tools
Example: customer service support
A chatbot answers simple status questions from approved sources. A copilot shows service staff relevant contract passages in the current ticket and drafts a response for editing before it is sent. A bounded agent can also look for missing details in two systems and prepare a callback task. It cannot cancel contracts or issue credits itself. One contact channel can therefore contain three clearly separated forms of assistance.
What to remember
Choose the smallest degree of freedom that measurably improves the task, not the most fashionable label. More autonomy is valuable only when real variation requires it and control grows with it.
Sources and further reading
These primary sources provide further detail on definitions, technical foundations or responsible use.
Content reviewed