Fundamentals

AI agent, automation or chatbot: which one do you actually need?

Most companies are sold an agent when a workflow would do, or a chatbot when they needed neither. Here is the difference in plain terms, with costs.

Published 21 Oct 2026 · 6 min read · Sentriqa

Three different tools

A chatbot answers questions. It is a conversation surface over a knowledge base or a model. A workflow automation follows fixed rules across systems: when an invoice arrives, extract these fields, create this record, notify this person. An AI agent is given a goal and tools, decides the next step itself, and keeps going across systems until the goal is met or it must escalate.

ChatbotWorkflow automationAI agent
Decides the next step?NoNo — rules doYes, within limits
Touches your systems?RarelyYes, fixed pathsYes, as tools
Best forFAQs, support deflectionStable, repeatable processesMulti-step work needing judgement
Typical cost (AU)A$5k–A$20kA$6k–A$25k per workflowA$12k pilot, A$30k–A$120k production
Main riskWrong answersSilent failuresWrong actions

The five-question test

  • Is the process the same every time? → Workflow.
  • Does the next step depend on reading several things and judging? → Agent.
  • Do people mostly need an answer, not an action? → Chatbot (governed, with citations).
  • Does it move money, commit the company or affect a person? → whichever you choose, add a human approval.
  • Can you name the number it has to move? If not, start with an audit, not a build.

Where companies go wrong

Buying an agent for a stable process (expensive, over-engineered), buying a chatbot and expecting it to do work (it cannot), or automating a process nobody has written down (it breaks on the first exception). The fix in all three cases is a short, paid discovery that produces a scope and a scorecard before any build.

Our default advice

Automate the six stable workflows first — document-to-data, quote-to-cash, onboarding, reporting, CRM hygiene, approvals. Then pilot one agent where judgement is the bottleneck. Chatbots last, and only governed ones.

Questions people ask

Can an automation include a model?

Yes — reading a document or classifying an email is a model step inside a rule-based workflow. That is still a workflow, because the path is fixed.

Is an "AI employee" an agent?

Yes. It is an agent given a role, a scope and a manager. The term is useful because it makes the governance question obvious.

Related: AI agents · AI workflows · What an agent costs

Not sure which you need?

One week of mapping gives you the answer and a fixed quote for it.