AI agent vs chatbot: what is the difference and which do you need?
A chatbot answers questions in a conversation. An AI agent uses a language model to plan and carry out a task: it calls tools, reads the results, decides the next step and repeats until the job is done. Choose a chatbot when the goal is to answer, and an agent when the goal is to complete work inside other systems, such as issuing a refund or updating a CRM record.

What is a chatbot
A chatbot is a program that holds a conversation. The older kind follows a script: it matches keywords and walks the user down a decision tree. The current kind is built on a large language model, so it understands free text and writes its own replies.
A modern chatbot can be very capable. Connected to a knowledge base, it answers product questions, explains policies and summarizes documents. What makes it a chatbot is the shape of the work: one message in, one reply out. It informs. It does not go and do something in another system.
What is an AI agent
An AI agent is software that uses a language model to plan and carry out a task. It is given a goal, a set of tools and rules about what it may do. Then it runs a loop:
- Decide the next step.
- Call a tool (an API, a database query, a search, a calculation).
- Read the result.
- Decide whether the task is done. If not, go back to step 1.
The loop is the difference. In a chatbot, the developer decides the sequence of steps in advance. In an agent, the model decides the sequence at run time, based on what each step returned. That makes agents able to handle tasks whose path is not known ahead of time, and it also makes them harder to test and secure.
There is a middle category worth naming: the workflow. A workflow uses a model inside fixed steps written by a developer, for example “classify the email, then draft a reply, then send it for approval”. Workflows are more predictable than agents and cheaper to run. Many projects that are sold as agents are better built as workflows.
AI agent vs chatbot: side by side
| Chatbot | AI agent | |
|---|---|---|
| Main job | Answer questions | Complete tasks |
| Output | A reply | A result in another system, plus a reply |
| Who decides the steps | The developer, in advance | The model, at run time |
| Tool use | None, or a single lookup | Multiple tools, chained |
| Model calls per request | Usually one | Several, sometimes dozens |
| Can change data | No | Yes, if granted write access |
| Failure mode | A wrong answer | A wrong action |
| Testing needed | Answer quality | Answer quality plus action safety |
| Security work | Protect the knowledge base | Permissions, approvals, audit logs, injection defenses |
| Relative build effort | Lower | Higher |
| Running cost per request | Lower | Higher |
The row that matters most is the failure mode. A chatbot that is wrong gives a bad answer, and a person can ignore it. An agent that is wrong can send the email, change the record or move the money. Everything that makes agents cost more follows from that.
AI agent examples
These are common patterns, described generally.
- Support agent. Reads the ticket, looks up the customer and the order, checks the policy, issues a refund under a set limit and writes the reply. Escalates anything above the limit.
- Ordering agent. Takes an order in a messaging app, checks the catalog and stock, builds the cart and confirms. Shai built and runs a production WhatsApp ordering bot on Claude with a model fallback chain, timeouts, retries and deterministic fallback replies for when the model fails.
- Research agent. Takes a question, searches several sources, reads the results, follows up on gaps and writes a cited summary.
- Coding agent. Reads a repository, plans a change, edits files, runs the tests and opens a pull request. Tools such as Claude Code work this way. See our Claude Code best practices.
- Internal data agent. Answers business questions from the ERP, CRM and SQL databases. Shai’s on-prem MCP gateway does this read-only, with per-user permissions and approved queries only.
Chatbot examples
- A website assistant that answers shipping, pricing and returns questions from the help center.
- An HR assistant that explains the leave policy.
- A documentation assistant inside a software product.
- A lead qualification bot that asks five questions and books a call.
Each of these informs or collects. None needs to change a record in a system of record, so none needs to be an agent.
AI agent use cases that justify the cost
An agent earns its extra cost when three things are true: the task has several steps, the steps depend on what earlier steps found, and a person currently does the work by moving between systems.
Use cases that fit:
- Resolving support requests that need a lookup and an action, such as refunds, address changes and cancellations.
- Order intake and changes over chat or messaging.
- Preparing quotes or reports that pull from several internal systems.
- Triage of inbound email into the right queue with the right fields filled in.
- Reconciliation tasks where records in two systems must be compared and exceptions flagged.
Use cases that do not fit: anything where a wrong action is costly and cannot be reviewed or reversed, and anything a fixed workflow already handles reliably.
How to choose
Work through these four checks in order.
1. Whether the work ends with an answer or with a change. If the user only needs information, build a chatbot. If something must be created, updated or sent, keep going.
2. Whether the sequence of steps is always the same. If it is, build a workflow with fixed steps and a model inside them. It will be cheaper and more predictable than an agent.
3. Whether the systems involved can be reached through APIs or database access. If not, the integration work comes first, and it may be the larger project. Our AI integration service covers this, including MCP servers for systems several AI tools need to reach.
4. What happens when the agent is wrong. If a wrong action is cheap to reverse, an agent can act on its own. If it is expensive, require human approval for that action. If it is unacceptable, keep the agent read-only.
A short version of the same guide:
| Situation | Build |
|---|---|
| Answer questions from documents | Chatbot |
| Answer questions that need a live lookup | Chatbot with one tool |
| Same multi-step process every time | Workflow |
| Multi-step task, path varies, actions reversible | Agent |
| Multi-step task, actions costly to reverse | Agent with human approval |
What changes in the build
Moving from a chatbot to an agent adds four pieces of engineering.
- Tool connections. Each system needs authentication, error handling and limits.
- Permissions enforced outside the model. The agent acts with the rights of the user it serves, never more. Instructions in a prompt are not a security control.
- Durability. Multi-step tasks fail in the middle. The agent needs retries, timeouts and the ability to resume without repeating an action.
- Evaluation and logging. Every tool call is recorded, and a test set of real cases runs on every change.
These are the same items that drive the budget. The detailed breakdown is in how much it costs to build an AI agent.
Start as a chatbot, grow into an agent
The lowest-risk route is staged. Launch a read-only assistant on one channel. Read the real conversations and find the requests where users wanted something done. Add the first action with a tight limit and an approval rule. Measure, then add the next.
This order gives you evaluation data before any action carries risk, and each stage produces something useful on its own.
What to do next
List the ten most common requests your team handles and mark each one “answer” or “action”. If most are answers, look at AI chatbot development. If the valuable ones are actions, look at AI agent development. Shai builds both, Build projects start at $25,000 with a fixed price per phase, and the first one-hour consultation is free.