AI Assistant, Automation, or Agent: Which Does Your Workflow Need?

Vendors use these terms loosely. The useful distinction isn’t how advanced the technology sounds. It’s who decides what happens next: a person, a fixed set of steps written in advance, or the AI model itself.

The three approaches in plain terms

An assistant: a person asks, reviews and decides

An employee-directed assistant is ChatGPT or Claude set up with your instructions and approved material. A person asks it for something, reads what comes back and decides what to do with it. Nothing changes in your other systems unless that person makes the change.

• A property manager asks for a first draft of a reply to a tenant’s maintenance question, based on the approved maintenance policy, then edits and sends it herself.
• A bookkeeper pastes in a confusing vendor statement and asks for a plain summary of what is owed, then checks it against the ledger.
• A new hire asks what the expense policy says about mileage and where the onboarding checklist lives.

Assistants fit work that varies case by case, where judgment stays with the person. The main risk is a wrong or incomplete draft, which the person reviewing it should catch.

An automation: fixed steps triggered by an event

A predefined workflow, built in a tool such as Make, Zapier or Monday.com’s automations, runs the same steps every time a specific event happens: a form is submitted, a booking is confirmed or a status changes. People decide the steps in advance and write them down.

• A website quote request creates a CRM record, is assigned by service type and region from an approved routing table, and notifies the assigned person.
• A confirmed onboarding booking creates a client item on the project board with the standard task list.
• A deal marked “won” produces a draft engagement letter from the approved template for a manager to review.

AI can sit inside one step to handle what fixed rules can’t: reading a free-text message to find the service requested, sorting an inquiry into a category, summarizing a long email or drafting a document. The surrounding steps stay fixed: where the result goes, who is notified and what needs approval. If the AI step returns something unexpected, or a required field is missing, the item goes to a person instead of being guessed.

An agent: the model chooses its own actions

An agent is given a goal and a set of tools, such as search, email, a CRM or a calendar. It decides for itself which steps to take, in what order and when it’s finished. Anthropic’s engineering guide Building effective agents draws the line clearly. Workflows are “systems where LLMs and tools are orchestrated through predefined code paths” (LLMs, or large language models, are the AI models behind tools like ChatGPT and Claude). Agents “dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.”

• Researching a prospective commercial client across their website and public information, then compiling a briefing, where the steps differ for every company.
• Working through a queue of support tickets, where each one needs a different combination of order lookups, policy checks and a drafted reply.

The same guide describes agents as suited to “open-ended problems where it’s difficult or impossible to predict the required number of steps, and where you can’t hardcode a fixed path.” Much administrative work isn’t like that. The steps are known, and the hard part is following them consistently.

Why the simplest reliable approach usually fits

More autonomy isn’t automatically better. In Building effective agents, Anthropic recommends “finding the simplest solution possible, and only increasing complexity when needed,” and adds, “This might mean not building agentic systems at all.” The guide says workflows “offer predictability and consistency for well-defined tasks,” and that the autonomy of agents means “higher costs, and the potential for compounding errors.” The guide dates from December 2024 and now notes that the tools it describes have changed, but its distinction between workflows and agents is still a practical way to decide.

Make, an automation platform, gives similar advice in the documentation for its own AI agent feature. It recommends a standard scenario “for tasks with predefined logic that always produce the same output for a given input,” and tells users to “Avoid tasks involving sensitive data, high-stakes financial or strategic decisions, or strict legal requirements” (Introduction to Make AI Agent).

In practice, work through the options in order:

1. Can a person do it with an assistant and review the result? Start there.
2. Are the steps the same every time, starting from a clear event? Build an automation, and add an AI step only where rules can’t read or write the text.
3. Are the steps genuinely unpredictable, and is the task worth the extra oversight? Only then consider an agent, with narrow tools, a test period and approval before anything consequential.

Authority: what it can see, what it can do, who approves

Whatever the label, three questions define the real scope. Write the answers down before anything is built.

• What can it see? Which documents, folders, inboxes, boards and records it can access. Connected tools usually follow the signed-in person’s permissions. Anthropic says “Claude inherits each person’s permissions from the connected service.” But a connection set up with one shared credential reaches whatever that credential can access.
• What can it do? Read only, create drafts, update records, send messages or move money. Each step up that list needs its own justification.
• Who approves, and when? Name the person and the point of approval: before an email goes to a customer, before a record changes, before a document is shared.

These questions apply to assistants too. In Anthropic’s words, connectors “let Claude access your apps and services, retrieve your data, and take actions within connected services.” An assistant with write access to your email or CRM is no longer only drafting. The platforms include controls for this. Claude Team and Enterprise owners can set each connected service’s actions to “Always allow, Needs approval, or Blocked” across the organization, and individual users can’t override that setting (Use connectors to extend Claude’s capabilities). ChatGPT’s workspace-wide permission options include “Always ask,” “Allow read actions” and “Allow low-risk actions” (OpenAI admin controls for plugins and apps).

Quick calls on common requests

• “Answer staff questions about our policies.” An assistant.
• “Every web inquiry should reach the right person with a record created.” An automation, perhaps with an AI step to categorize free-text requests.
• “Draft the renewal letter when a lease is 90 days from expiry.” An automation with an AI drafting step and a person’s approval before sending.
• “Figure out why sales are down and fix it.” None of the three yet. A person needs to define the question first.

If the answer is an assistant, see AI workspace setup. If it’s a predictable process, see workflow automation and three administrative workflows worth evaluating. If you’re still unsure, the automation assessment starts with a short description of the work.