AI Agent vs workflow vs chatbot
Quick answer: A workflow is the typed component graph, an agent is a component using instructions and tools, and a chatbot is the conversational input/output interface. These concepts do not establish deployment outcomes.
By Ai Flowi · Updated 9 October 2026.
Technical review: Janice Lee, Technical Support, Ai Flowi.
Published · Updated
In AI Flowi Workflow, separate the component graph, the agent component and the conversational input/output path. A workflow connects typed components; the documented agent component accepts instructions and tools; a chatbot describes the conversational input/output interface. These describe different parts of a design, not three proven deployment outcomes. Repository documentation dated 29 September 2026 supports the learning path. The saved 30 September Ask AI test generated 3 components and 2 connections and imported a new flow, but omitted requested validation.
This guide explains the documented product concepts. Documentation does not establish a deployed chatbot or a working tool-using AI Agent.
Compare the roles before choosing a design
| Term | Documented role in AI Flowi Workflow | Starting check | Evidence boundary |
|---|---|---|---|
| Workflow | Components move data from input or trigger through processing, logic or AI, action and output. | Write the input, required processing and expected output. | Connected ports do not prove the flow can run with your credentials. |
| AI Agent | The agent catalog entry describes instructions, a task and tools. | Inspect the documented component interface and the exact tools your task would require. | Tool use and external actions in your environment are not yet verified. |
| Chatbot | A conversational input/output interface within a component graph. | Define the conversation requirements and deployment target. | A conversational design is not evidence of deployment to a customer-facing channel; that is not yet verified. |
A conversational input does not by itself establish tool use. A model step can sit inside a workflow without proving a deployed agent. Keep the graph, the conversational interface and the external action separate in your acceptance criteria.
When is a fixed workflow enough?
Start with explicit processing or conditions when the incoming data already contains the fields the destination needs. The documented automation guide permits processing without an AI step in that case. If-Else is documented for a fixed condition; model or Structured Output components cover tasks involving text interpretation or extraction.
Treat this as a design choice to test. If a rule completely describes the decision, first check the rule-based path. Add a model only where the task needs interpretation. AI involvement does not remove the need for validation and error handling.
Read workflow basics for component interfaces and typed connections. This guide owns the architecture choice; the existing basics page owns detailed connection rules.
Related: AI automation guides.
AI Agent vs chatbot: check the required action
Separate a conversational interface from the actions it must perform. If the goal includes a tool or an external change, inspect the agent component and the selected tool’s accepted types and configuration. Catalog presence does not prove successful execution. For how model requests relate to tool execution, read the function calling guide before choosing an action path.
For subscriber updates and AI or human handling, follow the chat channel fields and workflow handoff guide.
Limitations: AI Agent and chatbot deployment
Not yet verified: a working AI Agent selecting and using the required tools, performing external actions and respecting your task’s controls. Check the repository agent interfaces and collect a scoped test result before describing that behavior as working.
Not yet verified: customer-facing chatbot deployment, channel availability and the operation of the proposed conversation path outside the Playground. Check the official docs repository and require deployment-specific evidence. No customer-channel capability is claimed here.
For a hands-on exercise, see the first-workflow tutorial.
Related: Chatbot vs live chat: choose response responsibility and the handoff boundary.
What did the live Ask AI test establish?
On 30 September 2026, Ask AI generated a webhook-to-spreadsheet starting flow and imported it as a new flow. The requested validation was absent. The user still needed to configure the account, spreadsheet identifier and row mapping.
Download JSON and Replace Current Flow were displayed but were not exercised in that saved test. The repository documents those paths, including replacement confirmation and locked-flow protection; their current live behavior is not yet verified. This guide recommends the observed new-flow import path only as a reviewed starting design, not proof of a complete automation.
On 30 September 2026, spreadsheet creation succeeded but direct mapping of raw webhook JSON to Append Rows failed: the full request object is not a row. On 1 October 2026, an editor test with a sample webhook body wrote one row to an app-created Google Sheet through Universal Webhook → JSON Operations (JQ) → Append Rows, with no AI model. Production-endpoint delivery, sheets not created by the app, and field validation were not exercised. A type-compatible graph alone does not establish the final row write. See the verified no-model tutorial.
Select tools for an AI Agent
In AI Flowi Workflow, the AI Agent component’s Tools input accepts the Tool type. Connect a component output that provides this type, then open its tool-selection control to inspect the available names and descriptions. Select the tools you want enabled; closing the selector saves each tool’s enabled or disabled status.
Configure and test the selected tool before relying on it for a task. Selecting a tool records the configuration; it does not establish that an external action succeeded. Use the component guide to inspect the component, then check the result of a small test. Use AI agent evaluation to define checks for the selected tool and its expected action.
For document retrieval, follow the knowledge bases and document retrieval guide; for uploading and selecting component files, see workflow file management. For the retrieval concept behind document-based answers, read retrieval augmented generation before adding a model response stage.
Check the result before extending the design
Write down the input and expected output for the next stage. Inspect the source output type and the destination’s accepted types, then configure required fields and credentials. Run a small test and stop at the first failed stage.
Use getting started to choose the learning sequence and Ask AI documentation for its full procedure. Keep generated designs, documented behavior and observed runtime results distinct as you learn.
Sources
- What is AI Flowi Workflow? — Workflow stages; documentation dated 29 September 2026. Accessed 1 October 2026 (official documentation copy).
- AI Flowi Workflow basics — Typed components and configuration; documentation dated 29 September 2026. Accessed 1 October 2026 (official documentation copy).
- AI Agent component interfaces — Instructions, tasks and tools; interface documentation is not tool-execution proof. Accessed 1 October 2026.
- Workflow automation design — Optional AI and fixed-rule processing; documentation dated 29 September 2026. Accessed 1 October 2026 (official documentation copy).
- Ask AI documentation — Download, replacement and confirmation procedures; documentation dated 29 September 2026. Accessed 1 October 2026 (official documentation copy).
- First-workflow tutorial — Learning route only; no new runtime result asserted. Accessed 1 October 2026 (official documentation copy).
- Official AI Flowi Workflow documentation repository — Check deployment evidence separately. Accessed 1 October 2026.
- Ai Flowi tests, 30 September and 1 October 2026. On 30 September, Ask AI generated 3 components and 2 connections and imported a new flow. Requested validation was missing, and connection and row configuration remained. Download JSON and Replace Current Flow were visible but not exercised. Spreadsheet creation succeeded; direct mapping of raw webhook JSON to Append Rows failed. On 1 October 2026, an editor test with a sample webhook body wrote one row to an app-created Google Sheet through Universal Webhook → JSON Operations (JQ) → Append Rows, with no AI model. Production-endpoint delivery, sheets not created by the app, and field validation were not exercised. See the verified no-model tutorial.