AI agent vs chatbot: which one does your business actually need?
AI agent vs chatbot: which one does your business actually need?
Chatbots answer; agents act. Choose an agent when the job has steps, tools and decisions, and a chatbot when it is pure Q&A. The test is simple: if the person who does this work today opens other systems and makes choices, you need an agent. If they only look things up and reply, a chatbot is enough.
The AI agent vs chatbot question comes up in almost every first call we take, and the answer is less about technology than about the job. A chatbot answers questions. An agent takes a goal, works out the steps, uses tools, and returns a result. If the work you want to automate is answering, buy or build a chatbot. If it involves opening systems, checking conditions and doing something, you need an agent, and pretending a chatbot can do it is the most common way to waste a quarter.
This article gives you a test you can apply to any workflow in ten minutes, a comparison of what each actually costs to build and run, and the cases where the honest answer is "neither yet".
What the chatbot vs agent difference really is
Both talk to people in natural language, which is why they are confused. The difference is what happens after the message arrives.
A chatbot maps a message to an answer. The answer may come from a script, an FAQ, or a language model reading retrieved documents. The bot does not change anything in your systems. When it cannot answer, it hands off.
An agent maps a message to a goal, then plans. It calls tools: your CRM, your order system, a payment gateway, a calendar. It reads the results, decides what to do next, and may loop several times before it is done. It can act, which is why it needs permissions, limits and an audit trail that a chatbot never needed. We describe the mechanics in What is an AI agent?.
The label "conversational AI vs agentic AI" describes the same split. Conversational systems optimise the exchange. Agentic systems optimise the outcome, and the conversation is just one interface to it.
Chatbot vs agent: a side-by-side
| Dimension | Chatbot | AI agent |
|---|---|---|
| Core job | Answer questions | Complete a task with a goal |
| Access to systems | Read-only, or none | Read and write through scoped tools |
| Reasoning | Single turn: message in, reply out | Multi-step: plan, act, observe, repeat |
| Failure mode | Wrong or unhelpful answer | Wrong action, which is why gates exist |
| Safety work | Content filtering, hand-off | Permissions, spend limits, approvals, audit log |
| Evaluation | Answer accuracy on a question set | Task completion rate on scenario suites |
| Typical build | Two to four weeks | Six to twelve weeks including shadow mode |
| Running cost | Low; one model call per message | Higher; several calls and tool round-trips per task |
The ten-minute test: when to use AI agents
Pick the workflow. Sit with the person who does it today. Ask them to do it once while you watch, and count three things.
- Systems opened. If they open more than one application to finish the job, the work has tools in it. A chatbot cannot reach them.
- Decisions made. If they pause to check a condition (is this customer eligible, is the invoice over the threshold, is the slot free), the work has judgement in it.
- Changes made. If anything in a system is different at the end, whether a refund, a booking or a status, the work has actions in it.
Zero on all three: chatbot. Answering from a knowledge base, order status lookups, policy questions, and "how do I" guides are chatbot jobs and a well-built one with retrieval does them well. One or more: agent. Refund handling, appointment rescheduling, lead qualification with CRM updates, claims triage and invoice exception handling are agent jobs, and a chatbot in front of them will deflect rather than resolve.
There is a middle case. If the work has decisions but the person is not allowed to make them alone, you want an agent that drafts and a human that approves. That is not a chatbot with extra steps; it is an agent in shadow mode, and it is how we launch nearly everything.
When a chatbot is the right answer
Chatbots get an unfair reputation because so many were deployed as FAQ walls. A chatbot is correct when the value is in the answer and the answer already exists somewhere. Internal IT help, product documentation, HR policy, store hours and returns policy all fit. Build it on retrieval over your real documents, not on hand-written intents, and measure answer accuracy on a golden set. The retrieval discipline is covered in Why basic RAG fails in production.
A chatbot is also the right first step when you are not yet sure which intents matter. Three months of chatbot logs tell you exactly which questions should become agent tasks.
When you need an agent, and what changes
The moment the job includes an action, four things enter the project that were absent before.
Tools with scoped permissions
Each system the agent touches is exposed through a tool with a narrow contract: look up this order, issue a refund up to this amount, book a slot in this calendar. The agent never gets a database connection. Increasingly this is done through the Model Context Protocol.
Policy gates
Actions above a threshold, or of a sensitive type, require a human approval. The agent prepares everything; a person clicks. Over time, as evidence accumulates, the threshold moves.
Evaluation suites
You cannot judge an agent by chatting with it. You judge it by running two hundred scenarios with known correct outcomes and counting completions, escalations and errors. We build this before the agent, not after.
Shadow mode
The agent runs alongside the human for weeks, proposing actions that people approve or reject. Only when the acceptance rate is high and the errors understood does it act alone, one intent at a time. The rollout approach is detailed in Shadow mode: the right way to launch AI agents.
Cost: what you are paying for
A chatbot with retrieval over your documents is a small project. An agent with three or four tools, policy gates and an eval suite is a medium one. Multi-step, multi-system work, where a planner coordinates several specialised workers, is a multi-agent system and starts at $24,500 or ₹16 lakh. A support-focused agent with account access and gated actions is a customer service agent from $12,500 or ₹8 lakh. Running costs differ too: an agent makes several model calls and tool round-trips per task, so cost per completed task, not cost per message, is the number to track. All starting prices are on the pricing page.
A worked example
A field-service SaaS company came to us with a chatbot that answered "how do I" questions about the product well and was ignored for everything else. Their support logs showed the questions users actually asked were tasks: reassign this job, find the technician nearest to this site, close these work orders. Those are three systems and several decisions each.
We kept the chatbot for documentation questions and built an in-app copilot that could act, with each action exposed as a scoped tool and reassignments over a certain scale routed for approval. It ran in shadow mode for several weeks while dispatchers accepted or corrected its proposals. The finished system is described in the in-app copilot case study. The lesson was not that the chatbot was bad; it was that it had been asked to do a job it could not reach.
Team and timeline
A retrieval chatbot takes an AI engineer and a content owner on your side, two to four weeks. An agent takes an AI engineer, a backend engineer for tool contracts, and a product owner who can sign off policy thresholds, typically six to twelve weeks including two to four weeks of shadow mode. If the workflow is not yet mapped, a ten-day Sprint Zero at $3,250 or ₹2,00,000, credited to the build, produces the decision, the intent list and the eval plan. A three-week ProofRun proves the hardest intent before you commit to the full build. After launch, a monthly Care Plan covers evals on every model change.
Before you start: a checklist
- Run the ten-minute test on the workflow with the person who does it today
- List the intents and mark each as answer-only, act-with-approval or act-alone
- Identify every system an agent would touch and whether it has an API
- Decide who signs off policy thresholds for each action
- Collect one to three months of real conversations for the eval set
- Agree the success metric: answer accuracy for a chatbot, completion rate for an agent
- Budget for shadow mode; it is not optional
- Name the owner of the weekly escalation review
Questions clients ask
- Can we upgrade a chatbot into an agent later? Yes, if it was built on retrieval and a real model rather than scripted intents. The knowledge layer carries over; the tools, gates and evals are new work.
- Is an agent riskier? It can change things, so yes, without gates. With scoped tools, limits and approvals the risk is bounded and auditable, which is more than most human processes can say.
- Do we need both? Often. A chatbot handles the long tail of questions; an agent handles the five intents that carry the volume and the value.
- What about voice? The same split applies. A voice bot answers; a voice agent books, reschedules and confirms.
Related reading
What is an AI agent? explains the mechanics for non-engineers, How much does an AI agent cost? breaks down the build and running budget, and 10 business workflows that are ready for AI agents today lists the intents we see automated most. Anthropic's guide to building effective agents is a useful primary source on when simple patterns beat agents.
Answering is a chatbot's job and acting is an agent's; decide by watching the work, not by reading the brochure.
Frequently asked questions
What is the main difference between an AI agent and a chatbot?
▾
A chatbot answers a message with a reply and changes nothing. An agent takes a goal, plans steps, uses tools to read and write in your systems, and returns a completed task. Agents therefore need permissions, limits and audit trails that chatbots do not.
When should a business use an AI agent instead of a chatbot?
▾
When the job involves opening other systems, checking conditions or making changes: refunds, bookings, lead qualification, claims triage. If the work is purely answering from existing knowledge, a retrieval chatbot is cheaper and sufficient.
How much more does an agent cost than a chatbot?
▾
A retrieval chatbot is a two-to-four-week build. Agents start at $12,500 for customer service agents and $24,500 for multi-agent systems, take six to twelve weeks, and cost more per task to run because of multiple model and tool calls.