An AI agent is software in which a language model decides which steps to take and which tools to use to finish a task, checks the result, and stops or hands back to a person when it is done or stuck. A chatbot only replies to messages. A workflow runs steps someone fixed in advance.

Picture an operations manager at a wholesale distributor who sits through three vendor demos in a week. The first shows a website chat window that answers questions about delivery times. The second shows an automation that copies each new web order into the accounting system. The third shows a system that reads a supplier's email about a late shipment, looks up the affected orders, drafts a message to each customer and waits for approval before sending. All three vendors call their product "an AI agent". Only the third one is.

What is an AI agent?

An AI agent is a system where a large language model (LLM) directs its own process: it chooses the next step and the tool to use, based on what it has seen so far, until the task is finished. The defining trait is that the model, not a fixed script, controls the path.

The two companies behind the most widely used models define it in similar terms. In Building effective agents (December 2024), Anthropic separates two kinds of system: "Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks."

OpenAI's A practical guide to building agents puts it in one line: "Agents are systems that independently accomplish tasks on your behalf." It names three core components:

• Model: the LLM that does the reasoning and decision-making.

• Tools: external functions or APIs the agent can use to take action, such as a CRM lookup or an email send.

• Instructions: explicit guidelines and guardrails that define how the agent behaves.

A tool, in this sense, is any function the model is allowed to call to read data or change something outside itself.

How is an AI agent different from a chatbot?

A chatbot answers; an agent acts. A chatbot takes a message and returns a reply, while an agent can take several steps across other systems to get a task done, with or without a conversation.

OpenAI's guide is explicit on this point: applications that use an LLM but do not let it control how a workflow runs, "think simple chatbots, single-turn LLMs, or sentiment classifiers", are not agents.

The line is not about the interface. A chat window can sit in front of an agent, and an agent can run with no chat at all, triggered by an incoming email or a new row in a database. What matters is what happens behind the reply:

• If the system only drafts text from what it already knows or retrieves, it is a chatbot.

• If it decides to look up an order, change a booking or open a ticket, and chooses the order of those steps itself, it is an agent.

How is an AI agent different from a workflow or automation?

A workflow follows a path a person designed in advance; an agent decides its own path at run time. Both can use an LLM, and both can call the same tools.

An automation is a workflow that runs without a person pressing a button, usually built in a tool such as n8n, Zapier or Make. Many useful "AI automations" are workflows with an LLM step inside: classify this email, extract these fields, summarise this call. Anthropic's article calls these patterns prompt chaining, routing, parallelization, orchestrator-workers and evaluator-optimizer. In each one a developer still decides the sequence.

Agents, chatbots and workflows side by side

• Chatbot. Who decides the steps: no steps; it replies to each message. Uses tools: rarely, or for retrieval only. Best for: answering questions from known content. Main risk: wrong or made-up answers.

• Workflow with an LLM step. Who decides the steps: the person who designed it. Uses tools: yes, in a fixed order. Best for: repeatable processes with clear rules and messy inputs. Main risk: breaks on cases nobody designed for.

• AI agent. Who decides the steps: the model, within limits you set. Uses tools: yes, chosen at run time. Best for: open-ended tasks where the number and order of steps can't be known in advance. Main risk: wrong actions taken with real permissions.

In practice most production systems mix the three. A workflow can hand one hard step to an agent, and an agent can call a workflow as one of its tools.

When should a business use an agent instead of a workflow?

Use an agent only when the steps can't be written down in advance; if they can, a workflow is cheaper, faster and easier to test. Anthropic's advice is to find "the simplest solution possible", and to add complexity only when it is needed.

OpenAI's guide lists three signs that a task is a good fit for an agent:

1. Complex decision-making: judgement, exceptions or context-sensitive decisions, such as approving a refund.

2. Difficult-to-maintain rules: rule sets so large that every update is costly or error-prone, such as vendor security reviews.

3. Heavy reliance on unstructured data: reading documents, interpreting natural language or talking with users, such as processing a home insurance claim.

It adds: "Before committing to building an agent, validate that your use case can meet these criteria clearly. Otherwise, a deterministic solution may suffice." Anthropic makes the same point from the other side: agents suit "open-ended problems where it's difficult or impossible to predict the required number of steps."

For the distributor above, copying web orders into accounting is a workflow. Handling a late-shipment email, where every supplier writes differently and the affected orders vary, is where an agent earns its cost. Our comparison of n8n, Zapier and Make covers the workflow side of that choice.

What does an AI agent need before it can do real work?

An agent needs clear instructions, the smallest set of tools that does the job, a checkpoint before any action that is hard to undo, and a record of everything it did. Without those, an agent is a demo, not a system.

OpenAI's guide describes an agent that "recognizes when a workflow is complete and can proactively correct its actions if needed", and that can "halt execution and transfer control back to the user" when it fails. Anthropic's article warns that the autonomy of agents means "higher costs, and the potential for compounding errors", and recommends testing in sandboxed environments with guardrails. Government guidance points the same way: the Five Eyes agencies' Careful adoption of agentic AI services advises organisations to "deploy agentic AI incrementally, beginning with clearly defined low-risk tasks".

Checklist before an agent goes live

1. The task, its inputs and its finished state are written down.

2. The agent has its own identity and only the permissions this task needs.

3. Reads run on their own; writes, deletions and outbound messages wait for a person.

4. Every tool call is logged with its inputs, the result and the identity behind it.

5. There is a test set of real past cases, and the agent's results on it have been reviewed.

6. Someone owns the agent after launch and can switch it off in one step.

If the agent connects to business systems through the Model Context Protocol (MCP), our guide to securing the MCP server first covers the permission side. For the screens people use to approve and undo an agent's actions, see designing for AI agents.

How we approach it at Vizio AI

We start by writing the task out as steps and asking which of them a person could specify in advance. Those become a workflow. Only the steps that need judgement over messy input go to an agent, and often that turns out to be one or two steps, not the whole process.

When an agent is the right answer, we give it a narrow job, its own credentials, read access first and an approval step on every write. We run it against past cases before it sees live ones, log every tool call into the client's existing monitoring, and widen its permissions only when the logs show it behaves as expected. The client's team owns the instructions and the switch-off after handover.

This is the work behind our DX Studio.

Sources

• Anthropic, Building effective agents (19 December 2024): https://www.anthropic.com/engineering/building-effective-agents

• OpenAI, A practical guide to building agents (PDF): https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf

• ASD's ACSC, CISA, NSA, Canadian Centre for Cyber Security, NCSC-NZ and NCSC-UK, Careful adoption of agentic AI services (1 May 2026): https://www.cyber.gov.au/business-government/secure-design/artificial-intelligence/careful-adoption-of-agentic-ai-services