Chatbot vs AI agent: which one does your company need?

Part 01

Chatbot vs AI agent: the difference in plain words

A chatbot is a conversation. You ask, it answers. The answer can come from a fixed script, from the model's general knowledge or from your own documents. When the conversation ends, nothing in your systems has changed.

An AI agent does a task. Take a supplier email asking why an invoice is still unpaid. A chatbot can explain your payment terms. An agent reads the email, finds the invoice in the ERP, sees that the delivery note is missing, drafts a reply asking for it and puts the draft in front of the accounts payable clerk. The language model can be the same in both cases. The job is different.

Four things separate the two:

• Tools. The agent can call your systems: read the inbox, query the ERP, write to the CRM. A chatbot only returns text.
• Permissions. Every tool runs under an account with limits. Read-only on the ledger, drafts only in email, no payments. Most of the security work sits here.
• Memory. The agent keeps the state of the task: what it has checked, what it found, what is still open. A chatbot keeps the conversation at most.
• A decision loop. The agent looks at the result of each step and chooses the next one, until the task is done or it hands over to a person.

Anthropic, the company behind the Claude models, draws the same line in its engineering guide Building effective agents. Workflows are "systems where LLMs and tools are orchestrated through predefined code paths". Agents are systems where the models "dynamically direct their own processes and tool usage". In a workflow, your developers set the order of the steps. In an agent, the model sets it. The second is more flexible, and harder to test and to price.

Our guide to AI agents for business covers what agents do inside a company and how their price is set.

"When building applications with LLMs, we recommend finding the simplest solution possible, and only increasing complexity when needed."

Anthropic, Building effective agents, 2024

Part 02

Four levels, from FAQ bot to agent

The market talks about chatbots and agents as if there were two options. In practice there are four. Each step up adds capability, and each one adds work: more integrations, more permissions, more cases to test.

1. FAQ bot

It answers a fixed set of questions from a script or a short knowledge base. Opening hours, the return policy, where to find the invoice portal. Cheap and predictable, and useless once a question falls outside the list.

2. Assistant over your documents

The model searches your contracts, manuals and policies and answers with a link to the source. The technique is called retrieval-augmented generation, or RAG. It answers "what notice period do we have with this supplier?" without anyone opening the folder. It reads. It does not change anything.

3. Workflow with one model call

A fixed sequence, written in code or in a tool such as n8n or Make. A new email arrives, the model classifies it and pulls out the fields, and the workflow files it and opens the ticket. The model does one job inside steps you defined. Most document intake and inbox triage belongs here.

4. AI agent

The model receives a goal and a set of tools and chooses the steps. It fits work where the path changes from case to case: a supplier dispute that needs the contract, the delivery history and two earlier emails before anyone can reply.

Many problems sold as agents are better solved with simpler automation. If your process fits on a whiteboard as a fixed sequence of steps, level 3 does the job with fewer moving parts, lower running costs and results you can check case by case. Gartner has a name for the opposite habit: "agent washing", the rebranding of chatbots, RPA and AI assistants as agents. It estimates that only about 130 of the thousands of vendors selling agentic AI are real, and it predicts that over 40% of agentic AI projects will be canceled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls (Gartner, June 2025).

European companies mostly work at the lower levels today. In 2025, 19.95% of EU enterprises with 10 or more employees used at least one AI technology, according to Eurostat. The most common was analysis of written language, at 11.75%. Technologies that automate workflows or assist in decision making, the category closest to agents, reached 5.35%. The categories do not map one to one onto chatbots and agents, but they show that most companies start by reading and writing text.

Part 03

When a chatbot is enough, and when an agent pays back

Two questions place most use cases. How much does each case differ from the last one? And what does a mistake cost?

A chatbot is enough when

• The questions repeat and the answers live in documents you control.
• The output is information, and a person takes the action.
• A wrong answer is cheap to correct: the user asks again or picks up the phone.

A chatbot is a waste when

• People ask it to do things it cannot do: change a booking, cancel an order, update an address. Each request ends in "please contact support", and the customer has now waited twice.
• Your documents are outdated or contradict each other. The assistant repeats the contradiction with confidence.
• Nobody reads the logs. The questions people type are the best list you will get of what to automate next.

An agent pays back when

• Cases vary, and each one needs several steps across two or more systems.
• A trained person spends most of the time gathering information, not deciding.
• Volume is high enough to cover the setup and testing, and the exceptions can go to a person.

An agent is overkill when

• The steps are always the same. A workflow does it for less.
• Volume is low. A handful of cases a month does not justify the evaluation work.
• Permissions are a mess. If the agent would need admin rights on your ERP to function, fix the access model first.

Part 04

Costs and risks of each, side by side

Three things drive the cost at every level: the number of systems involved, the number of case types, and what an error costs. What changes between levels is where the cost sits. A FAQ bot costs little to run and a lot in frustrated users. An agent costs more per case, most of it in testing and permissions.

Anthropic's guide states the trade-off plainly: "Agentic systems often trade latency and cost for better task performance, and you should consider when this tradeoff makes sense." An agent calls the model several times on one case, where a workflow calls it once. Each extra call adds cost, waiting time and another chance to be wrong.

The risks differ in kind. A chatbot that gets it wrong gives a bad answer, which a person can ignore. An agent that gets it wrong takes a bad action: a reply sent, a record changed, a ticket closed. That is why agents need narrow permissions, a log of every step and a confidence threshold below which a person decides.

Part 05

Three worked examples

A support inbox

An online shop gets emails about order status, address changes, returns and complaints. An assistant on the website answers policy questions with a link to the right page. A workflow handles routine email: the model classifies each message, and the workflow pulls the order status and drafts the reply. Complaints are different. Each one needs the order history, the carrier's tracking and a decision about goodwill. That is where an agent fits: it gathers everything, proposes a resolution and a person approves any refund.

Verdict: an assistant and a workflow for most of the inbox, an agent only for complaints.

An accounting team

Supplier invoices arrive by email as PDFs. Matching them against purchase orders and delivery notes is a fixed sequence: extract the fields, compare, post the invoice or flag it. That is level 3, and it covers most of the volume. An invoice that matches no order means someone searches emails, contracts and earlier credit notes before deciding. An agent can do that search and present the findings, and the accountant makes the call. A chatbot over the accounting manual helps new staff with "which cost centre does this go to?", and it changes nothing in the books.

Verdict: a workflow for matching, an agent for the exceptions, with the accountant deciding.

An HR onboarding flow

A new hire signs a contract. Then come accounts, equipment, payroll data, a training plan and a welcome email. The steps are known and the order rarely changes, so a workflow triggered by the signed contract does it. A level 2 assistant answers the new hire's questions from the employee handbook: holidays, expenses, who to ask about the laptop. An agent adds little here, and the personal data involved argues for the narrowest access possible.

Verdict: a workflow and an assistant. No agent.

Part 06

A decision checklist

Run each use case through these questions before you ask for a quote.

1. Write the task down as steps.If the steps fit on one page and rarely change, start with a workflow.
2. Mark what it touches.If the output is only information, a chatbot or an assistant over your documents may be enough. If it has to write into a system, you need a workflow or an agent.
3. Put a price on one error.A wrong answer, a wrong email, a wrong booking. Above a level you set, a person approves before anything goes out.
4. Check your documents.An assistant is as good as what it reads. If the policies are out of date, fix them first.
5. Count the cases.An agent needs enough volume to pay for its testing. Low volume points to a workflow or to staying manual.
6. Collect real cases, including the awkward ones.They become the test set, whichever level you build.
7. Ask the provider why a simpler level will not do.If nobody can explain why a workflow is not enough, you are paying for an agent you do not need.

At Alpgency this checklist opens the free Audit, and we recommend the lowest level that does the job. The first use case then gets a Demo on your real data and a build, under one fixed price agreed after the Audit. After go-live you choose: a full handover, a fixed monthly support fee, or us as tech partner on a fixed schedule.

FAQ

Frequently asked questions