An AI agent is a system that uses a model to choose actions and tools, observe results and work toward a goal. A workflow defines the path in code. A chat interface can front either system: the architecture and permissions, not the chat box, determine what it can do.
A conversational interface. A basic version answers in text; a tool-connected version can also take actions. Inspect the system behind the interface.
Follows a path you defined. When X happens, do Y, then Z. Reliable and predictable, and it breaks on anything you did not anticipate.
Has a goal and a set of tools. It decides which to use, observes the result, and continues until the goal is met or it gives up. The path is not fixed in advance.
Take handling an inbound enquiry.
| Approach | What happens |
|---|---|
| Chatbot | A basic text-only chatbot answers the question; tool-connected chat systems can do more. |
| Automation | Sends a fixed acknowledgement, creates a CRM record, notifies you. Always the same steps. |
| Agent | Reads the enquiry, looks the company up, checks whether they are an existing customer, decides whether it is a quote request or a support issue, drafts an appropriate response, schedules a follow-up — and escalates to you when it hits something it cannot judge. |
This matters more than the capability list, because agent failures are less obvious than automation failures:
The workflow/agent distinction follows Building effective agents, published by Anthropic on 19 December 2024. The enquiry scenario is illustrative, not a measured client result. The implementation advice is Kanaky Tech’s editorial guidance; capabilities and costs depend on the chosen system.
Editorial review: 16 September 2026 · Kanaky Tech. Report a correction with the passage and reference: contact.
An automation follows a path you defined in advance, which makes it predictable and brittle. An agent decides its own path using tools available to it, which makes it flexible and less predictable. If you can write down the steps, build an automation — it will be cheaper, faster and easier to debug.
For well-scoped tasks with a human reviewing consequential actions, yes. For unsupervised work with real consequences, be careful — they fail in ways that look like success, and errors compound inside loops. The common pattern that works is an agent that prepares and a person who approves.
It depends on how many steps it takes, which varies per task and is not fully predictable. That is a genuine budgeting difference from fixed automations. Running a model locally can avoid a model provider’s per-token fee, but hardware, electricity, maintenance and any external tools still cost money — see running AI locally.
Most businesses need automation and are being sold agents. If you can write the steps down, you do not need something that decides its own. Agents earn their complexity when the next step genuinely depends on what the previous one found.
Kanaky Tech is an AI automation agency working across New Zealand and the Pacific. Start with a free AI opportunity audit: we map how your business actually runs, rank what is worth automating, and give you a clear scope before anything is built — no obligation.