An assistant answers. An agent acts.
The two words are used as if they meant the same thing. They do not, and the difference decides what a wrong answer costs you.
Most of the confusion comes from the fact that the same products now do both. A tool you chat with in the morning may be sending emails on your behalf by the afternoon.
So the useful distinction is not the product or the model behind it. It is what happens to the output, and who reads it before it counts.
1. An assistant answers
A person asks a question, the assistant replies, and the person reads the reply and decides what to do with it. The assistant’s work ends at the answer.
ChatGPT, Claude and Gemini answering a question are assistants in this sense. So is the tool your team uses to look up a policy before replying to a customer.
The person reading is a filter. They can notice that a price looks wrong, check it and choose not to repeat it.
2. An agent acts
An agent is given a goal rather than a question. It decides on a step, takes it, looks at the result and decides on the next one.
Its steps use tools: it reads files, drafts a page, sends an email, updates a ticket. What comes out at the end is not an answer but something done.
Nobody reads every step, and often nobody reads the result before it reaches someone. That is the point of an agent, and it is also the risk.
3. The model is not the difference
The same model can sit behind an assistant and an agent. What changes is the arrangement around it: who starts it, what it is allowed to touch and where the person stands.
Where the person stands
DiagramWith an assistant, the person stands after the output and decides whether to use it. With an agent, the person stands before it, setting the goal and the limits, and afterwards reads only some of what it did.
4. What a mistake costs
Suppose both get your price wrong. The assistant tells one person, who may check it against your site before repeating it.
The agent puts the wrong price in every follow-up it sends that afternoon. By the time someone notices, the mistake has a list of recipients.
The error is the same in both cases. What differs is how many places it reaches before a person sees it.
5. What each needs from your brand
Because the person stands somewhere different, each needs something different from you.
- Assistants you do not run, such as a customer asking ChatGPT about you, read whatever they can find. You cannot choose what that is, but you can make sure the best material about you is on your own domain, current and plainly written.
- Assistants you do run, such as your team’s internal tools, can read your record every time they answer. The person reading still catches what the record misses.
- Agents need the record and more. They need to know what they may do as well as what is true: which claims they may make, what they must refuse, and the named person they hand the rest to.
The checks a reader used to make in their head have to be written down before an agent starts, because no reader is there to make them.
6. When to use which
Use an assistant when the output needs judgement before it is used; the person reading is part of the job, not a delay in it. Use an agent when the task comes round again and again and the right answer can be written down in advance.
Between the two sits the sensitive work: anything that goes to many people at once, cannot be taken back or is about a named customer. There, let the agent make it and a person read it and send it.
Assistant or agent, task by task
| Task | Use | Why |
|---|---|---|
| Checking which discount is approved before a sales call | Assistant | The rep decides what to offer; the assistant makes sure they decide from the current terms. |
| Drafting a reply to a difficult customer email | Assistant | The situation is new, the tone matters, and a person edits and sends it. |
| Replying in public to a complaint on social | Assistant | It is under your name, in front of everyone, to someone already upset; a person should write the last version. |
| Researching how a competitor describes itself | Assistant | The question changes every time, and the answer feeds a decision rather than a page. |
| Wording a claim in a regulated industry | Assistant | Low volume, high stakes: every sentence should pass a person. |
| Sending the monthly newsletter | Agent, then a person | The agent assembles it from approved claims and the month’s news; a person reads the final send, because it reaches every subscriber at once and cannot be recalled. |
| Building a sales deck for a named prospect | Agent, then a person | The agent builds it from the approved library and current pricing; the rep checks every claim about that customer before it is shown. |
| Making a launch set from approved claims | Agent, then a person | The page, the emails and every ad size come from one set of rules; a person reviews the set before it ships. |
| Watching what assistants say about you each week | Agent | It checks rather than decides, flags what changed since the last run, and nobody is doing it today. |
| Answering “where is my order” | Agent | One correct answer, from your own systems; anything outside the list goes to a named person. |
| Tagging and routing support tickets | Agent | Clear rules and high volume, and a wrong tag costs a minute rather than a customer. |
Most agents should start life as assistants. Run the task with a person reading every output, write down the checks they keep making, and let the agent take over once those checks are rules it can follow.
7. A test for the one you have
Whatever a tool is called, ask one question: does a person read every output before it matters? If so, it is working as an assistant, and the person is your safeguard.
If not, it is an agent, and it should be held to an agent’s standard. That means a record to read from, rules for what it may do and someone named for what it may not.
Start with the audit. It is free, it is one pass, and you keep what it finds whether or not anything follows it.
