Guide

Chatbot vs AI Agent vs Workflow: which does your business need

Published 6 October 2026

Most pitches treat these three as tiers of the same product, with the agent at the top and the chatbot as the starter version. They are not tiers. The question is not which is most capable, it is: when this thing finishes running, what was it responsible for delivering?

Three kinds of accountability

A chatbot is accountable for an answer, a workflow is accountable for a process completing, and an agent is accountable for an outcome it had to work out how to reach. Picking the wrong one of the three is usually a scoping mistake rather than a technology mistake.

A chatbot is accountable for an answer. Someone asks, it responds, and the person reading the answer is the check on the answer.

A workflow is accountable for a process completing. You define the path in advance, the system follows the path you defined, and success is a state you can inspect afterwards: the invoice was posted, the ticket was closed. Anthropic's engineering team puts the architecture plainly: "Workflows are systems where LLMs and tools are orchestrated through predefined code paths."

An agent is accountable for an outcome it had to work out how to reach. In the same post: "Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks." You supply the goal and the tools. It picks the route at run time, and the route may be different tomorrow.

Anthropic keeps both under one umbrella term, agentic systems. The workflow versus agent line is about who fixes the path, not about which one counts as real AI.

This is not only a vendor's framing. The UK Information Commissioner's Office, the country's data protection regulator, puts the same axis this way in its Tech Futures report on agentic AI: "while traditional software typically follows a fixed way to solve problems, agentic AI might generate different ways to approach a problem or achieve a goal." It describes the change as developers "designing modern AI agents that can create and execute context-specific plans in more variable environments, with less human direction".

The comparison in one table

Chatbot Workflow Agent
Accountable for An answer A process completing An outcome, route unspecified
Who decides the path The person asking, one turn at a time You, up front, in the design The model, at run time, inside limits you set
What "done" means A useful reply A completion condition you can check A goal condition met, by some route
Typical failure A wrong answer the reader can see A stage stalls, or a bad record moves downstream It takes a wrong action and keeps going
Who notices The user, immediately Whoever owns the stage Nobody, unless you built the alert

What IBM actually says, and what it does not

IBM Think's page defines agentic workflows and contrasts them with a chatbot running on rule-based decision trees. It does not claim that a conversational LLM chatbot is obsolete.

IBM Think's definition is short and specific: "Agentic workflows are AI-driven processes where autonomous AI agents make decisions, take actions and coordinate tasks with minimal human intervention." The same page sets a hard floor on the word: "In artificial intelligence (AI), a workflow is not agentic if it does not consist of an AI agent."

The comparison IBM is drawing there is narrower than it is usually reported. Its worked example is an IT support chatbot running on a rule-based automation system, one where, in IBM's wording, "the chatbot runs through static decision trees and provides predefined responses". IBM's verdict on that setup is that it "is efficient for basic, well-defined issues but struggles with complex, multistep troubleshooting that requires adaptability."

That is agentic work contrasted with rule-based decision trees, not a claim that a conversational LLM chatbot is obsolete, and the page never makes one. If a slide cites IBM for the line that chatbots are finished, the citation does not carry that weight.

How common are agents, really

Google Cloud's AI agent trends 2026 report offers one usable adoption figure: 52% of executives in gen AI-using organizations have AI agents in production. That is executives at organizations already using generative AI, not all enterprises, and the number is footnoted back to a prior Google Cloud survey, The ROI of AI 2025, of 3,466 enterprise decision makers. Any version quoted as a share of all businesses has drifted from the source.

The report describes the shift this way: "This is AI that moves beyond answering questions to understanding a goal, making a plan, and taking actions across applications to achieve it with extensive human guidance and oversight." Note the last clause. Even the vendor making the case for agents is describing supervised ones.

Four questions that usually settle it

  1. Does the path change depending on what it finds? If the sequence is the same every time, you want a workflow. Branching on outcome is not by itself a reason to reach for an agent, because a workflow can hold a decision tree you wrote.
  2. Can you write down the steps in advance? Anthropic's framing is that agents suit "open-ended problems where it's difficult or impossible to predict the required number of steps, and where you can't hardcode a fixed path." If you can hardcode it, hardcode it.
  3. Does anyone need to approve individual actions? If yes, you are describing a workflow with a human gate, whatever you call it.
  4. Is one exchange the whole job? If so, you want a good chatbot or a good prompt, not a system. Anthropic recommends finding the simplest solution possible and increasing complexity only when needed.

Anthropic states the tradeoff directly: "Workflows offer predictability and consistency for well-defined tasks, whereas agents are the better option when flexibility and model-driven decision-making are needed at scale."

The buying question: who finds out when it gets something wrong

All three will be wrong sometimes. The difference is who is standing there when it happens. With a chatbot it is the user, with a workflow the failure lands on a known stage, and with an agent the failure compounds.

With a chatbot, the user is the error handler by default. That is cheap, and it works right up until the answer gets acted on without anyone re-reading it. With a workflow, failures land on a known stage, which is worth a great deal, but only if the workflow has an explicit stop condition and an error path rather than a silent skip.

With an agent, the failure compounds. Anthropic is blunt: "The autonomous nature of agents means higher costs, and the potential for compounding errors. We recommend extensive testing in sandboxed environments, along with the appropriate guardrails."

So ask the vendor or the build team two questions in this order.

  1. What does it do when a step fails, specifically, not "it handles errors"?
  2. Who is told, by what route, within how long?

If the answer to the second is that someone will spot it in a dashboard eventually, you have not bought an agent. You have bought an unattended process with a name that sounds accountable.

Where this fits if you use Bespoke Prompting

The same split runs through Bespoke Prompting's build types. A Workflow spec has to state a measurable completion predicate, something like complete when published_url is set and brand_check_passed is true, rather than describing itself as a workflow assistant. An Agent spec has to carry machine-checkable stop conditions, and every tool in its registry needs a named fallback, retry once then do X, fall back to Y, never the phrase handle the error. The generator is also deliberately biased toward supervision: when the autonomy level is a close call it picks the more supervised one, on the reasoning that a needless approval gate is annoying while a missing one can take an irreversible action.

Sources

Suggested internal links