Skip to content
Back to all articles

Agents

Chatbot, assistant or agent: what your business actually needs

A plain guide to chatbots, assistants and agents, when each one fits, and why an agent should ask for approval before acting on its own.

October 21, 2026

5 min read

Hands typing on a laptop on a wooden table

Vendors use the words chatbot, assistant and agent loosely, and sometimes as if they meant the same thing. That makes it hard to know what you are buying or what you should build. The differences are practical, though, and they affect both what you spend and how much can go wrong.

Here is how we separate them, in plain terms.

The short version

TypeWhat it doesWhat starts itDoes it act on its own?
ChatbotAnswers questions in a conversationA visitor or customer sends a messageRarely, it mostly replies
AssistantHelps a team member with a taskA person opens it and asksNo, the person decides what happens next
AgentPicks up incoming work and handles stepsAn event, such as a new email or formYes, within limits you set

The same underlying model, whether it comes from OpenAI, Anthropic or another provider, can power all three. What separates them is where the tool sits, what triggers it and what it is allowed to do.

Chatbots: answers where people ask

A chatbot lives in a conversation window, usually on your website and sometimes on a messaging channel such as WhatsApp. Someone types a question and it replies. Older chatbots followed scripted decision trees. Current ones use a language model plus your own content, such as FAQ pages, policies and product details, to write answers in natural language.

A chatbot fits when:

  • You get the same questions over and over, and the answers already exist in writing.
  • Visitors ask outside business hours and would otherwise wait until morning.
  • You want to collect contact details or pass serious inquiries to a person.

The limits are worth knowing. A chatbot is only as accurate as the content it draws from, so gaps or outdated pages turn into wrong answers. It should know when to say "I'm not sure, here is how to reach us" instead of guessing. It also should not make commitments nobody approved, such as special discounts or delivery dates.

Assistants: help for your team, on demand

An assistant works with a person, not in place of one. Think of a custom GPT or a Claude project set up with your company's documents, tone of voice and instructions. A team member opens it and asks it to draft a proposal, summarize a long contract, rewrite a product description or find an answer in the internal handbook.

The person decides when to use it and what to do with the output. Nothing leaves the business unless they send it.

Assistants fit when:

  • The work needs judgment, but a good first draft saves time.
  • Your team keeps answering the same internal questions, like "where is the template for this?" or "what is our refund policy for that?"
  • You want consistent tone and format across different people.

Because a person reviews everything, assistants carry the least risk of the three. They also save the least time per task, since someone still has to start each one and check the result. For many small teams, that tradeoff is the right one.

Agents: work that starts without you

An agent is triggered by an event rather than by someone typing. A new email arrives, a form is submitted, an invoice lands in a folder. The agent reads it, works out what kind of item it is and takes steps: labels the email, drafts a reply, adds a row to a spreadsheet, creates a contact in your CRM or posts a note in your team chat.

That is where the time savings can get larger, because nobody has to remember to start the task. It is also where the risk goes up, because the agent acts inside your systems, often while nobody is watching.

Agents fit when:

  • Work arrives continuously through a channel you control, such as an inbox, a form or a shared folder.
  • The categories and next steps are clear enough to write down.
  • Delays cost you something, like a lead that cools off overnight.

Take a hypothetical home renovation company that receives quote requests, supplier invoices, scheduling emails and spam at the same Gmail address. An inbox agent could read each message, label it by type, copy the key details from quote requests into a sheet and draft a first reply for the owner. The owner still decides what gets sent. The agent takes over the sorting and the copying.

Why human approval matters at the start

When an agent is new, you do not yet know how it behaves on your real data. Test examples rarely cover everything. The first weeks will turn up the odd email, the client who writes in two languages, the invoice with a strange layout.

So we recommend starting every agent in a mode where it prepares actions and a person approves them. In practice, that means:

  1. Drafts, not sends. The agent writes replies into the drafts folder, and a person reviews and sends them.
  2. Labels before moves. It tags and sorts, but does not delete or archive anything important.
  3. A log of every decision. Each action is recorded with the reason, so you can spot patterns in its mistakes.
  4. A clear fallback. Anything it is unsure about goes to a "needs review" label instead of being forced into a category.

After a few weeks, you will know which actions it handles reliably. Those can be allowed to run on their own, one at a time. Actions with real consequences, like sending money, confirming orders or replying to an unhappy client, may stay on approval permanently, and that is fine.

An agent you supervise and trust a little is more useful than one you trust completely and never check.

Approval also protects your client relationships. Someone who reads the drafts every day notices when the tone is off or a fact is wrong, long before a client does.

Choosing between them

A few rules of thumb:

  • If the people asking are outside the business and need answers, start with a chatbot.
  • If the people asking are inside the business and need help with their own work, start with an assistant.
  • If the work arrives on its own and follows clear steps, consider an agent, with approval switched on.

You may well end up with more than one. A knowledge base written for an internal assistant can later feed a website chatbot, and an inbox agent can draw on the same documents when it drafts replies. Getting one well-scoped tool right first makes the next one easier to build.

Where to start

Pick the channel where repetitive work piles up: website questions, internal questions from your team or your inbox. For one week, note every item that comes through it and what you did with it. If those items fall into a few clear categories with predictable next steps, an agent with approval is worth testing. If the work is mostly answering questions, a chatbot or an assistant is the simpler first step.

If the inbox is where the work piles up, our Inbox Triage Agent is built for that, running inside your own email account with labels and drafts you approve.

Inbox Triage Agent, $597 one-time

An agent that reads incoming emails or tickets, labels them by topic and urgency, and drafts replies for your team to approve.

See what is included

Get started with Ninelabs OS

Pick a deliverable, pay once and send your brief. If you are not sure which one fits, write to us first and we will tell you honestly.