AI Agent vs Chatbot vs Automation: Which One You Need

AI agent vs chatbot vs automation: what each one does, where it breaks, what it costs to run, and a 5-question test to pick the right one for each job.

· Guide · · 9 min read

Automation runs steps you defined, the same way every time. A chatbot answers questions in a conversation and stops there. An AI agent works toward a goal over several steps and chooses its own tools, inside limits you write down. Fixed steps need automation, questions need a chatbot, and work that changes case by case needs an agent.

Many "which AI do we need" calls I take start with the wrong noun: the company has seen a chatbot demo, been quoted for an agent, and already pays for an automation tool nobody maintains. This guide puts the three side by side so you can tell which one a specific job needs. If you already know you want an agent, Agentic AI for SMBs covers how to build and secure one.

AI agent vs chatbot vs automation at a glance

Automation (rules / RPA) Chatbot AI agent
What it does Runs the same steps every time: moves data, routes approvals Answers questions in a conversation, from a model or your documents Works toward a goal over several steps: reads, decides, calls tools, drafts
Who decides the steps You, in advance Nobody: it replies one message at a time The agent, inside limits you wrote
Input Structured data in known formats A person's question A goal plus messy input: email, documents, records
Output A finished action: record created, file moved Text on a screen Drafts and updates in your systems, logged
Where it breaks An input it doesn't expect, a changed screen or field Questions outside its sources, confident wrong answers A plausible but wrong step, and errors that add up across steps
Cost to run Lowest, flat per run Low, per conversation Highest, and it varies with each case
Testing Fixed test cases Real questions with expected answers A golden set of 20 to 30 real cases, scored on every change
Risk Low: it only does what you scripted What it tells people What it can touch in your systems
Typical example Supplier invoice into the accounting system Internal policy Q&A, website FAQ Inbox triage with draft replies
Choose it when The steps never change People need answers and nothing needs doing The work changes case by case and a good draft saves real time

What is a chatbot?

A chatbot is a conversational interface: a person asks, it answers. Modern chatbots run on a language model and often read from your documents, but each turn ends with text on a screen, not with work done in your systems.

That makes them good at getting people to answers faster: an internal policy assistant, a product helpdesk, a website FAQ. The risk sits in what it says. A confident wrong answer to a customer is still your company's answer.

Buyers have noticed the limits. US searches for "AI chatbot for business" fell 39% year on year (Bullock, June 2026).

What is automation (and RPA)?

Automation runs a sequence of steps you defined in advance: when an invoice arrives, read these fields, create the record, send it for approval. RPA (robotic process automation) is the variant that drives screens the way a person would, clicking through an old system that has no API.

It wins on cost and predictability. Each run costs close to nothing and does the same thing every time, which is exactly what you want for an accounting entry. It breaks on the unexpected: a new invoice layout, a moved button, a field renamed in an update. Then it stops, or writes to the wrong field.

There is a middle option people tend to skip: a fixed workflow with one AI step inside it. The steps stay yours, and the model does only the part rules can't, such as reading a PDF invoice into fields. Anthropic recommends this order: find the simplest solution and add complexity only when it is needed (Anthropic, December 2024). Most of our proven AI solutions look like this.

What is an AI agent?

An AI agent is software that pursues a goal over several steps and decides which tools to use and in which order, inside limits you write down. Anthropic draws the line clearly: in a workflow, models and tools follow predefined code paths, while an agent directs its own process and tool use (Anthropic, December 2024).

In practice, an agent reads an email, looks up the customer, checks the contract, drafts a reply and flags the exception, choosing each next step from what it found. It is the right tool when the input varies and the judgement behind it can be written down. It is the wrong tool when the steps never change, because you pay for decisions nobody needed.

Outside large companies, agents are still early. Among organisations under $1 billion in revenue, the share scaling AI agents stayed flat at 22% (McKinsey, August 2026). That is a good reason to start with one job, not a platform.

What's the difference between an AI agent and a chatbot?

A chatbot produces an answer and waits for the next question. An agent takes a goal and carries out steps in your systems until the job is done or it needs a person.

Three differences matter in practice:

  • Who starts the work. A chatbot waits for a message. An agent can start from an event: a new email, a task assigned to it, a deadline.
  • What it touches. A chatbot reads. An agent reads and writes: records, drafts, files. That is why it needs an agent identity: its own credentials and permissions, separate from any person. OWASP lists identity and privilege abuse among the top risks for agentic applications (OWASP, December 2025).
  • How you test it. A chatbot is judged on its answers, an agent on the result of a whole sequence, tested end to end on real cases.

AI agents vs automation: when does each win?

Automation wins when the steps are known and stable. An agent wins when each case needs a different path. Most companies have far more of the first kind of work than they think.

  • Automation for data sync, invoice intake with a fixed layout, approvals, recurring reports: anything you could draw as a flowchart that fits on one page.
  • An agent for inbox triage, a first read of contracts, research before a sales call, chasing stalled tasks. Work where today a person "has a look and decides".

RPA vs AI agents is the same question with an older tool. RPA is cheaper and more predictable on stable screens. An agent copes with variation but costs more per task and can take a wrong step that looks right. Anthropic notes that agentic systems trade latency and cost for better task performance, and that you should check the trade is worth it (Anthropic, December 2024). If your RPA bots break every time a supplier changes a layout, the fix is usually an AI step inside the workflow, not an agent.

Choosing an agent where automation would do is expensive. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, mainly because of rising costs and unclear business value (Gartner, June 2025).

Can they work together?

Yes, and in production they usually do. Automation handles the trigger and the plumbing, an agent handles the judgement, and a conversational interface is where people talk to it.

Take a morning inbox digest. A scheduled job pulls the night's email (automation). An agent sorts it, reads the threads that matter and drafts replies. The founder approves or edits each draft in chat. Nothing leaves without that approval.

Agents also call automations as tools. An agent that decides a supplier invoice is in order triggers the same posting workflow your team already uses, instead of writing to the ledger itself. It needs fewer permissions and you keep the same audit trail.

How do you choose? A 5-question test

Answer these about one specific job, not about "AI" in general:

  1. Are the steps the same every time? Yes: automation. Stop here.
  2. Does the job end with an answer to a person, with nothing to do in a system? Yes: a chatbot, or better search over your documents.
  3. What varies? If one step varies (free-text email, different document layouts), use a workflow with one AI step. If the path itself changes, consider an agent.
  4. Can you write down how a good person decides? If not, no AI is ready for the job yet. Write the judgement down first.
  5. What happens if it is wrong once? If the answer is money out, a contract signed or a customer misinformed, a person approves every output, whichever tool you pick.

Then measure before you build: hours, cycle time and errors on the job as it runs today. Without that baseline you can't tell whether the agent paid off.

What does it cost to run each?

Automation is the cheapest to run, a chatbot sits in the middle, and an agent costs the most per task and varies the most.

  • Automation: close to nothing per run, plus the platform. The hidden cost is maintenance when screens and formats change.
  • Chatbot: model usage per conversation, plus keeping its sources current. Stale sources are the usual failure.
  • AI agent: several model and tool calls per task, logs, and a person's review time. Add the test set that runs on every change. That is engineering time, and it is the part people skip.

None of this needs per-user AI subscriptions. We deploy in your environment or as private AI on your own servers, so the running cost is your own model and cloud bill.

Agent washing: how to spot a chatbot sold as an agent

"Agent washing" is Gartner's name for vendors relabelling chatbots and AI assistants as agentic AI. Gartner estimates that only about 130 of the thousands of vendors claiming agentic AI are the real thing (Gartner, June 2025).

Six questions to ask in the demo:

  • What does it do without a message from a person? If nothing, it is a chatbot.
  • Who decided the order of the steps? If a developer fixed it, it is a workflow. That can be the right answer, but it shouldn't be priced as an agent.
  • Which agent identity does it act under? "The admin account" is the wrong answer. OWASP's guiding principle is least agency: no more autonomy than the job justifies (OWASP, December 2025).
  • What does it see for each task? Context engineering decides what an agent sees: the facts, records and instructions the task needs, and nothing else. An agent pointed at the whole drive is harder to trust and costs more to run.
  • Where is the human in the loop? Ask which actions need a person's approval, then ask to see the approval step itself, not a slide about it.
  • How is it tested? Ask for the test set and the score on the last release. No test set means nobody knows whether last week's change broke something.

Buyers already ask for control over autonomy. In G2's analysis of enterprise reviews of AI agents, only 7% of reviewers would give an agent total autonomy (G2, October 2026).

Examples by job: finance, sales, support, admin

Same department, different tools:

Job Usually the right fit Why
Finance: bank reconciliation, posting supplier invoices Automation with one AI step The rules are known; only reading the documents varies
Finance: preparing the investor update Agent, with a person approving It pulls from several sources and the story changes every month
Sales: logging calls in the CRM Automation Same fields every time
Sales: researching a prospect before a call Agent Each company needs a different search path
Support: answering product questions Chatbot over your documentation The output is an answer
Support: refunds and exceptions Agent that drafts, person who approves Judgement plus money out
Admin: deadlines and recurring paperwork Automation A calendar and a template
Admin: first read of a contract against your standard positions Agent Every contract is different

For the automation side, see AI business automation. For agents, see agentic AI.

What does this look like in a real company?

An operations-heavy engineering company asked us for AI assistants for its team. We started with an Ops & AI audit, and the brief set the conditions: production-grade from day one, data under the company's control, a person approving everything outbound, and no new per-user AI subscriptions.

The audit sorted the work into the three buckets above:

  • Chatbots spread across personal accounts. The team had a set of custom assistants, each living in one person's account. We consolidated them onto one agentic framework, with a shared knowledge layer and memory, and each capability is tested against real cases.
  • The mixed pattern for email. The morning digest described above, in production and tested against a golden set of real cases. Every reply stays a draft until a person on the team sends it.
  • Agents, one job at a time. Each agent gets its own agent identity and a human in the loop, and hand-offs between agents are piloted on one case first.

The details are in the case study.


If you have a job in mind, bring it to a call. We'll tell you whether it needs automation, a chatbot or an agent, and what it takes to run it in your own environment. Build with us

Sources

  • Anthropic, "Building effective agents", 19 December 2024: anthropic.com
  • Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", 25 June 2025: gartner.com
  • OWASP GenAI Security Project, "OWASP Top 10 for Agentic Applications for 2026", 9 December 2025: genai.owasp.org
  • McKinsey, "The state of AI in 2026: On the road to ROI", 25 August 2026: mckinsey.com
  • G2, "Stop Selling Autonomy: What 3,222 Enterprise Reviews Say AI Agent Buyers Want", 30 October 2026: learn.g2.com
  • Lilach Bullock, "The AI for Marketing Era Is Quietly Ending. AI Agents Are Taking Over", 14 June 2026 (US Google search data via DataForSEO): lilachbullock.com

Last checked: 2 November 2026.

Frequently asked questions

A chatbot answers a question and waits for the next one. An AI agent takes a goal and carries out several steps in your systems to reach it: it reads, decides the next step, calls tools and prepares drafts or updates. A chatbot's risk is what it says. An agent's risk is what it can touch, so it needs its own agent identity, limited access and a human in the loop approving what leaves the company.

If the steps are the same every time, automation is enough, and it is cheaper and more predictable. If only one step varies, such as reading invoices in different layouts, use a workflow with one AI step inside it. Consider an agent only when the path itself changes from case to case, as with inbox triage or a first read of a contract.

Not for stable, repetitive screen work. RPA is cheaper and more predictable when the screens and formats don't change. Agents help where RPA keeps breaking because the input varies, but they cost more per task and can take a plausible wrong step. In most companies the two end up working together: the agent decides, and the existing automation does the posting.

Only if you give it tools, a goal and permission to act, and at that point it needs to be built and tested as an agent. Attaching one or two actions to a chatbot does not make it an agent. Gartner calls the relabelling of chatbots and assistants as agents agent washing.

Yes. Automation costs close to nothing per run once built. An agent makes several model and tool calls per task, so its cost varies with each case, and it needs a test set that runs on every change plus a person's time to review drafts. It pays off only on work where that judgement saves real hours.

Ask what it does without a message from a person, who decided the order of its steps, which agent identity it acts under, what context it sees for each task, where the human in the loop sits, and how it is tested. If it only replies to messages, it is a chatbot. If a developer fixed the order of steps, it is a workflow, which can be fine but should not be priced as an agent.

Start from the job that costs you the most time, measure it, then pick the simplest tool that fixes it. For most small and mid-sized companies that is an automation, often with one AI step, on finance or admin work. Agents come next, one job at a time, once there is a baseline to measure them against.

Tell us what you're building.

Book a strategy call. We'll tell you where we can help, and say so quickly if we can't.

Let's talk