By KYN AI Advisory Team — AI implementation specialists, Singapore
A chatbot answers questions in a conversation. A workflow automation runs fixed steps when a trigger fires. An AI agent reads unstructured input, such as an email or a receipt, and chooses the next step from rules you set, within limits. Most SMEs need a mix: workflow automation for steps that never vary, an agent for steps that need reading, and a chatbot only where customers ask the same questions in the same channel. Choose by the shape of the task, not by the label.
What is the difference in one table?
| | Chatbot | Workflow automation | AI agent | |---|---|---|---| | Input | A message in a chat | A structured trigger: form, row, webhook | Unstructured or structured: email, PDF, reply, record | | Logic | Scripted flow or model-generated answers | Fixed if-this-then-that | Rules plus a model that reads and decides among allowed steps | | Output | A reply | A completed step in another tool | A drafted message, a labelled record, a proposed posting | | Handles variation | Little | None outside the branches you wrote | Within the rules you set | | Typical risk | Wrong or invented answer | Breaks when the input changes shape | Acts on a misread input | | Human role | Handles what it cannot | Fixes broken runs | Approves actions that matter |
The three overlap. An agent often contains workflow steps. A chatbot can be the front end of an agent. The useful question is what each part does, not which one it is called.
What is a chatbot good at?
A chatbot suits a narrow set of repeated questions where an answer already exists: opening hours, delivery status, how to book, what a plan includes. The better ones retrieve the answer from your documents and cite the page.
Where it fits:
- High volume of the same questions in one channel
- Answers that are already written down
- A clear route to a person when it does not know
Where it fails: questions that need account data it cannot see, requests that need a decision, and anything where an invented answer costs money. Set a rule that the chatbot says "I don't know" and hands over, instead of guessing.
What is workflow automation good at?
Workflow automation connects tools by fixed rules. When a form is submitted, create a row, send a confirmation, notify a channel. It is cheap, predictable and easy to audit because every run follows the same path.
Where it fits:
- Structured inputs with consistent fields
- Steps that never need interpretation
- Moving data between two systems
Where it fails: the input changes shape. A supplier sends a PDF in a new layout, a customer replies with a question instead of a form, a field arrives empty. A fixed workflow either stops or passes the error along. The fix is more branches, which become hard to maintain.
What is an AI agent good at?
An agent handles the step that needs reading. It opens an email and decides what the sender wants. It reads a receipt and proposes a category. It looks at a registry page and writes one fact about the company. Then it passes the result to a workflow step or to a person.
Where it fits:
- Unstructured inputs: email, documents, replies
- Tasks a new hire could do from a one-page instruction
- Output a person can check quickly
Where it needs limits: anything outbound, anything about money, anything irreversible. The agent drafts; a named person approves. Without that, a misread input becomes an action. For a risk view, read is it safe to let an AI agent send emails to customers, and for hostile input see prompt injection in business AI agents.
How do you decide which one a task needs?
Ask four questions in order.
- Is the input structured and consistent? If yes, use workflow automation.
- Is the answer already written, and the question repeated? If yes, consider a chatbot with retrieval.
- Does the step need reading or judgement within rules? If yes, use an agent.
- Does the output move money, send a message or delete something? If yes, add an approval by a named person whichever tool you use.
Examples:
| Task | Fit | Why | |---|---|---| | Form submission to CRM row | Workflow | Structured, fixed | | "What are your opening hours?" on WhatsApp | Chatbot | Repeated, answer exists | | Sorting replies to outreach | Agent | Needs reading, output is a label | | Matching bank lines to invoices | Workflow plus agent | Exact matches by rule, near matches by agent | | Reading supplier quotes in mixed layouts | Agent | Unstructured input | | Sending a payment | Workflow with approval | Money waits for a person |
What does each cost to run?
Costs differ by orders of magnitude between a fixed step and a model call, and a model call is still small per item. In KYN's live systems, a reply sorted costs about S$0.01, an outreach email with one checkable first line about S$0.02, and a morning report about S$0.05. These depend on the model and the text length. The larger costs are the build, the integration work and the review time. Compare total cost per item over a month, not the price of a licence.
Maintenance differs too. A workflow breaks when a system changes. A chatbot drifts when documents change. An agent drifts when the world changes, such as new reply types. Each needs monitoring and a named owner.
Can they be combined?
Yes, and they usually are. A common pattern:
- A workflow triggers when an email arrives.
- An agent reads it and labels it.
- A workflow routes it by label.
- A chatbot or a person replies, depending on the label.
- A workflow logs the outcome with actor and time.
The agent sits in the middle, only where reading is needed. Everything around it stays fixed and auditable.
What are the failure modes to plan for?
- Chatbot: an invented answer. Control: answers only from your documents, with a hand-over.
- Workflow: a silent break after a format change. Control: an alert on any run that fails or produces an empty field.
- Agent: a misread input. Control: an approval step for anything outbound, and a sample check of its labels each week.
Name an owner for each, and test the failure before launch, not after.
Where KYN fits
KYN builds systems that combine all three: cloud infrastructure in the client's own account, a front end designed around each team's workflow, and agents for sales, marketing, operations and finance with rules-based logic. Agents have their own restricted accounts, and the database uses row-level security. Outbound messages, money, filings and deletions wait for a named person.
The first 14 days are a review of your stack and interviews with your team. A full build ships in under 90 days. See integrations and operations systems.
Frequently asked questions
Is an AI agent just a smarter chatbot?
No. A chatbot's output is a reply in a conversation. An agent's output is an action or a proposed action in your systems, such as a label, a draft or a record, usually with no conversation involved.
Do I need an agent if I already have Zapier-style workflows?
Only for steps that need reading. If your inputs are structured and consistent, workflow automation is cheaper and easier to audit. Add an agent where emails, documents or replies break your branches.
Which one should I buy first?
Start with the task, not the tool. Score your candidate tasks using our scoring sheet for a first automation, then pick the lightest tool that does the job.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
