A chatbot is software that holds a conversation. An AI agent is software that pursues a goal, using tools and judgment along the way. A chatbot can sit inside an agent as its interface, and an agent can sit behind a chatbot as its engine, which is why the two are confused.
This article is part of our technology guides. For what AI software costs to build, the sibling guide on custom software cost gives the published ranges.
What is a chatbot?
A chatbot is a program that holds a conversation with a person, in text or voice; IBM's definition is "a software application that communicates with people through text or voice" (updated 16 July 2026). The first generation followed scripts: a decision tree of fixed questions and answers, good for "what are your hours" and poor for anything unexpected. The current generation uses a large language model, which lets it understand a question phrased any way and answer in natural language, often grounded in your own documents so it answers about your products rather than the world in general.
IBM puts the relationship plainly: every AI chatbot is a chatbot, but not every chatbot is an AI chatbot. What both generations share is the shape of the job: a person asks, the bot answers, and the exchange ends. A chatbot does not go and do anything. If the answer is "your order will ship Tuesday", a chatbot looked that up or was told it; it did not reschedule the shipment.
What is an AI agent?
An AI agent is software that is given a goal and works toward it: it decides which steps to take, uses tools (a database, an email system, a calendar, a browser, an internal API), checks the result of each step and continues until the task is done or it needs a person. Anthropic's guide Building effective agents (19 December 2024) describes agents as "typically just LLMs using tools based on environmental feedback in a loop".
The same guide draws the line that matters. Workflows are "systems where LLMs and tools are orchestrated through predefined code paths"; agents are systems where "LLMs dynamically direct their own processes and tool usage". A chatbot, scripted or AI, is on the workflow side: the path is fixed, the model fills in the words. An agent chooses the path. IBM describes the same difference from the other side: AI agents take a more active role, and "conversation is only one part of their function"; they can plan, make task-specific decisions and carry out multi-step workflows with limited human involvement.
AI agents vs chatbots: the difference in one table
The table below compares the two on what they do, how they decide, what they connect to, the risk they carry and what they cost to run. The short version: a chatbot replies, an agent acts.
The three tiers most businesses choose between
In practice the choice is rarely binary. Three tiers cover most needs, and each is a step up in capability, cost and care:
- A scripted chatbot. Fixed questions, fixed answers, buttons and menus. Cheap, predictable, limited. Right for hours, locations, order status lookups and routing a visitor to the right form.
- An AI assistant grounded in your content. A language model that answers open questions from your documents, policies and product data, with citations and a hand-off to a person. Right for support, pre-sales questions and internal knowledge, where the answer exists and the job is finding and phrasing it.
- An AI agent with tools. A model that can take actions: update a record, book a slot, draft and send a reply for approval, reconcile a list, chase a document. Right for repetitive multi-step tasks with clear rules and a measurable outcome, where a person reviews exceptions.
Which jobs suit a chatbot?
Jobs where the output is an answer and the stakes of a wrong answer are low or recoverable: frequently asked questions, order and appointment status, first-line support that hands difficult cases to a person, lead qualification that collects a few details, and internal help with policies and procedures. A well-grounded assistant can handle a wide range of these; a scripted bot handles the narrowest well and the rest badly. Our chatbots service page sets out how we build the grounded kind.
Which jobs suit an AI agent?
Jobs where the output is a completed task: processing an inbound request from email to system entry, keeping two systems in sync, preparing a report from several sources, triaging and routing tickets with the first steps done, chasing missing information, or running a checklist across many records. The common features are repetition, several steps, access to tools, clear rules for the routine cases and a person for the exceptions. Our AI agents service and AI automation pages describe how we scope that work.
Cost, risk and control
A chatbot's cost is mostly set-up and content: the script, or the documents and the grounding. Its running cost is low and its risk is a wrong or awkward answer, which is embarrassing but recoverable. An agent's cost includes integrations with every system it touches, the rules and limits it works within, logging so you can see what it did, and a review step for exceptions. Its running cost is higher because it does more, and its risk is a wrong action, which is why permissions and limits matter more than the model.
Both need the same two things from you: a clear definition of the job, and someone who owns the outcome and reviews what the software does in its first weeks.
How to decide: start simple, add complexity only when needed
Anthropic's advice to developers applies to buyers too: "finding the simplest solution possible, and only increasing complexity when needed", with workflows for well-defined tasks and agents where "flexibility and model-driven decision-making are needed at scale". A practical sequence:
- Write down the job as a person does it today: inputs, steps, decisions, outputs, exceptions.
- If the job is answering, build the assistant and ground it in your content. Measure the share of questions it resolves and the hand-off rate.
- If the job is doing, start with the steps that follow fixed rules and automate those as a workflow, with a person approving the output.
- Only where the steps genuinely need judgment across many tools should an agent choose its own path, inside limits you set, with every action logged.
- Review weekly for the first two months: what it got right, what it got wrong, what people stopped doing because of it.
Common mistakes when choosing between an agent and a chatbot
The mistakes we see most:
- Buying an agent to answer questions. An assistant grounded in your content does that job at a fraction of the cost and risk.
- Buying a chatbot to do tasks. A bot that can only talk about the order cannot fix it, and users learn that fast.
- No grounding. An assistant that answers from general knowledge rather than your documents will be confidently wrong about your business.
- No limits. An agent with broad permissions and no logging is an incident waiting to happen.
- No owner. Both need a person who reviews the first weeks and keeps the content or rules current.
Conclusion
A chatbot converses; an AI agent acts. Anthropic's line between predefined code paths and a model that directs its own process is the useful test: if the path is fixed and the output is an answer, you need a chatbot or a grounded assistant; if the steps need judgment and tools and the output is a finished task, you need an agent with guardrails. Start with the simplest tier that does the job.
PromotEdge International is an international strategy, technology, marketing and design company, and our delivery team is in India. We build chatbots, AI agents service and AI automation service within our technology practice, and we scope the job before recommending the tier. We name no client example here because none is approved for publication yet.
Facts we rely on: see our company facts for what we state about ourselves and where each fact comes from.