Chatbot vs AI agent: which one your business needs (and when neither)

Chatbot or AI agent: which one do you need?

Start from the job, not the technology. To answer, a chatbot; to complete a task in your systems whose steps change case by case, an agent; if the steps never change, a workflow; if the data arrives structured, perhaps no model at all.

What is a chatbot, what is an AI agent and what sits in between?

A chatbot is a conversational interface: it takes a message and returns an answer, today usually generated by a language model from your documents. OpenAI's guide to building agents (April 2025) is explicit: simple chatbots and other applications that don't let the model control how a workflow runs are not agents.

For OpenAI, an agent independently accomplishes tasks on your behalf: the model runs the workflow (it knows when the task is done, corrects itself, hands back control when it fails) and picks tools to gather context and act, within defined guardrails.

Anthropic (December 2024) names the middle ground: in a workflow, the model and tools follow code paths you define; in an agent, the model directs its own process. Its advice is to find the simplest solution and add complexity only when needed, which may mean no agent at all.

How do chatbots, workflows and agents compare?

The core difference is who decides the next step: nobody beyond the answer (chatbot), your code (workflow) or the model (agent). Integration, risk, cost and time to production all follow from that.

Cost deserves a number: in June 2025, Anthropic measured that agents typically use about 4 times more tokens than chat interactions, and multi-agent systems about 15 times more, which pays off only for valuable tasks. We break down the bill in how much an AI agent costs.

How do you spot agent washing?

Ask what the model actually decides. Gartner (June 2025) calls agent washing the rebranding of AI assistants, robotic process automation (RPA) and chatbots as agents without substantial agentic capabilities, and estimates that only about 130 of the thousands of agentic AI vendors are real.

The same release predicts that over 40% of agentic AI projects will be canceled by the end of 2027 (escalating costs, unclear business value or inadequate risk controls) and notes that many use cases sold as agentic don't need an agentic implementation. Five questions for any vendor:

• Which decisions does the model make at run time, and which are scripted?
• Which tools can it call, with which credentials, and what can each one change?
• When does it hand over to a person?
• What are its pass rate and cost per resolved task on your cases, in your language?
• Are its actions logged where the agent cannot edit them?

A scripted flow behind a chat window is often the right product. The problem is paying agent prices, and taking agent risks, for it.

When is an agent worth it? Four concrete cases

When the path depends on what the system finds and each task is worth the extra tokens and controls. OpenAI's guide gives three signals: decisions full of exceptions, rules too costly to maintain and heavy reliance on unstructured data; otherwise, a deterministic solution may suffice.

• FAQ support. Policies, hours, products: a chatbot with search over your documents. Tools would add risk, not value.
• Order status. One lookup, one answer, the same steps every time: a chatbot with a read-only query, or a fixed workflow.
• Refunds. Policy, history, evidence and exceptions are weighed case by case: here an agent earns its place. OpenAI cites refund approval as complex decision-making and lists large refunds among actions that need human oversight. Typical design: the agent gathers facts and proposes; a person approves above a set amount.
• Back-office reconciliation. Matching bank lines to ERP invoices follows rules, so a workflow does most of it; for the leftovers (partial payments, free-text references), an agent can propose matches for a person to confirm.

Anthropic sees customer support as a natural fit for agents: a chat interface plus tools to read customer records and orders, actions such as refunds and success measured in resolutions. Without actions to take, an agent is an expensive chatbot.

When is a form or a plain workflow better than both?

When the input is structured, the rules are stable and errors are expensive. Changing a billing address, booking an appointment or approving an expense under a fixed limit needs validation, not language understanding: a form, an API call or a classic automation is faster, cheaper and easier to audit.

• Structured inputs: if data arrives in fields, a model only adds cost and variance.
• Hard rules: rates, limits and eligibility belong in code, where they can be tested.
• Low volume: a task done a few times a month rarely repays design, evaluation and monitoring.

Models still help at the edges: routing incoming emails to the right form, or extracting fields from a PDF before a rule-based process. More in what to automate first with AI.

What risks does each option carry, and how do you control them?

A chatbot mostly risks bad answers; an agent, bad actions. So controls sit on the text for a chatbot and on tools and permissions for an agent.

Simon Willison's “lethal trifecta” sums up the agent case: private data, untrusted content and a way to send data out, combined in one agent, let an attacker steal data through what it reads. Remove one leg before launch; the full rules are in AI agents in production.

Europe: since August 2, 2026, Article 50(1) of the AI Act requires providers to make sure people know they are talking to an AI system, unless it is obvious. Colombia: Law 1581 of 2012 governs the personal data a chatbot or agent processes, including data sent to models hosted abroad. And Gartner expects rules guaranteeing access to a human to raise assisted service volume by 30% by 2028: design the handoff from day one.

How do you start without overbuilding?

1. Pick one task with an owner and a metric: resolution rate, handling time or error rate.
2. Build the simplest version that could work: one model call with search over your documents, then a workflow.
3. Write 20 to 50 test cases from real conversations, in your customers' languages; Anthropic suggests that size for a first evaluation set.
4. Connect tools read-only first, then add write actions one at a time, with confirmation and logging.
5. Define when a person takes over: after how many failures, and for which actions.
6. Measure cost per resolved task on real traffic before widening the scope.

An assistant inside your own product is a sound first step. In the app we built for Réseau Mazette, a network of French-speaking professionals and entrepreneurs in Spain, Mia, the AI copilot, has its own tab next to the directory and events, where members open the chat to ask their questions (see the case). For assistants on your website or WhatsApp, see our AI chatbots for business.