All writing

Chatbot or agent?

Guide5 min read

Chatbot or agent?

Guide · November 20, 2026 · 5 min read

Most businesses that ask us for a chatbot need something else, and a few that ask for an agent need far less. The short answer: a chatbot answers questions; an agent does work. Choose by the job you want taken off someone's desk, not by the word on the vendor's page.

This guide is for an owner or manager deciding what to buy or build. It is drawn from systems we have built and run for businesses of ten to a few hundred people.

What a chatbot is

A chatbot is a conversation. Someone types a question, and it answers from what it has been given: your website, your policies, your price list, your opening hours.

It waits to be asked, it speaks only to the person in front of it, and it changes nothing. When it works, a customer gets an answer at ten at night instead of the next morning. When it is out of its depth, the good ones say so and hand over to a person.

What an agent is

An agent is a worker. It is given a goal and access to the systems that goal touches, and it takes steps until the work is done: it reads what comes in, decides what it is, files it, updates a record, drafts a reply, and chases what is missing.

It does not need to be asked, and most of what it does happens where no customer sees it. One we run reads incoming requests, files each against the right job, and drafts the reply for a person to approve before anything is sent.

The difference that matters: a chatbot can be wrong in a sentence. An agent can be wrong in your records.

When a chatbot is enough

A chatbot earns its place when the same questions arrive again and again, and the answers already exist somewhere.

  1. Customers ask the same twenty questions, and someone answers them by hand
  2. The answers are settled: hours, prices, policies, how to book, where an order is
  3. Nobody needs anything changed, only explained
  4. A wrong answer is an inconvenience, not a cost

If that describes the problem, buy one rather than build one. Good ones are inexpensive, and building your own rarely pays for itself.

When an agent pays off

An agent earns its place when someone's day goes to moving information between places, and the rules for doing it can be written down.

  1. An inbox that is really a job queue, read and sorted by hand every morning
  2. Orders, forms or documents re-typed from one system into another
  3. Paperwork that has to be chased: missing signatures, overdue confirmations, unanswered quotes
  4. Replies that follow a pattern but still need to be right for this customer

The test is simple. If you can describe the job to a new employee in a page, and checking their work is quicker than doing it, an agent can probably do it. If the job depends on judgement nobody can write down, it cannot, yet.

What each costs to run

A chatbot's cost is mostly the subscription. An agent's cost is mostly the work of connecting it to your systems and keeping it right as they change.

ChatbotAgent
Setting it upDays: load your content and policiesWeeks: connect each system it reads or writes, and write down the rules
Running itA monthly fee, often per conversationModel usage per task, plus hosting
Keeping it rightUpdate the content when your answers changeReview its work, adjust rules, follow changes in the systems it touches
Who has to be involvedWhoever owns the websiteWhoever does the job today, for a few hours at the start
What it savesTime answering the same questionsHours of reading, re-typing and chasing, every week

The honest comparison is not chatbot against agent. It is an agent against the hours the job takes now. If those hours are small, the agent will not pay for itself.

What goes wrong, and how to avoid it

Most failures come from giving the software more freedom than the job needs, or less information than it requires.

  1. A chatbot that invents an answer. Limit it to your own content, and have it hand over to a person when it is unsure.
  2. An agent that acts without anyone noticing. Start with it drafting, not sending. A person approves what matters until its record earns more trust.
  3. An agent with access to everything. Give it only the systems and permissions its one job requires.
  4. Nobody owns it. Someone inside the business should answer for what it does, even when we maintain it.
  5. Automating a process nobody agrees on. If two people do the job two ways, write down the right way first. The software will only make the disagreement faster.

Every agent we build keeps a person in the loop for anything that leaves the building or changes money.

How to decide

Answer these five questions about the job you want to hand over.

  1. Is the work answering questions, or moving information between places? Answering points to a chatbot; moving points to an agent.
  2. Do the answers or rules already exist in writing? If not, writing them down is the first step either way.
  3. How many hours a week does the job take today? A few hours rarely justifies an agent.
  4. What does a mistake cost? The higher the cost, the more a person should approve before anything is final.
  5. Who will own it? If nobody, wait.

If the answers point both ways, start with the agent behind the scenes. The chatbot can come later, drawing on the same records.

Start with one job

The businesses that get the most from either one start small. They pick the single job that costs the most hours, automate that properly, and watch it for a month before adding the next.

If you have a job like that in mind, bring it to us as you would describe it to a colleague. We will tell you which of the two it needs, or whether it needs neither.

Start with one problem, or bring us the whole operation.