The short answer
Automation follows steps you decide in advance: when this happens, do that. It’s cheap, fast and gives the same result every time.
An AI agent decides its own next step. It reads a request, chooses which of your tools to use, acts, and adapts when the situation isn’t what it expected.
Most businesses don’t need to pick one. The question to ask about each task is: where, exactly, does a decision have to be made? If nowhere, use automation. If at one point, add one AI step to a normal workflow. If the decisions change from case to case, that’s when an agent earns its cost.
Three levels, not two
“Agent or automation” sounds like a yes-or-no choice. In practice there are three levels, and most of the value sits in the middle one.
| Level | What decides the next step | Good for | Running cost |
|---|---|---|---|
| 1. Rules-only automation | You, in advance | Structured data with a known path | Lowest |
| 2. Workflow with an AI step | You, except at one point where AI reads or sorts something | Messy input, such as emails or documents, feeding a fixed process | Low |
| 3. AI agent | The AI, within limits you set | Requests that vary every time and need several tools | Higher |
Anthropic, the company behind Claude, draws the same line: workflows run AI and tools along predefined code paths, while agents direct their own process. Its advice is to start with the simplest option that works and move up only when you need to.
Here’s what each level looks like in systems we’ve built for clients.
Level 1: rules-only automation
Client build: a real estate brokerage with seven agents. Their website form asks enquirers which agent they want to work with. Someone in the office used to read each submission and assign it by hand.
We replaced that with a GoHighLevel workflow. The form submission triggers one condition that reads the answer and sends the lead down one of seven branches. Each branch tags the contact and assigns it to that agent. A fallback branch flags anything that doesn’t match.
There’s no AI in it, and there shouldn’t be. The input is a form field, the rule is clear, and the same answer should always give the same result. Adding AI here would only add cost and the chance of a wrong guess.
Use rules-only automation when the input is structured (form fields, order data, records) and you could write the steps on a checklist.
Level 2: a normal workflow with one AI step
Client build: a business whose Gmail inbox mixed client emails with promotions and sales pitches. They wanted help replying, but no AI sending emails on its own.
We built an n8n workflow that runs on every new email. One AI step, a text classifier, reads the email and decides which category it belongs to. Everything after that is fixed: each category has its own reply instructions, and the result is saved as a Gmail draft for a person to check and send.
The AI does the one thing rules can’t: read messy text and judge what it is. The workflow around it stays predictable, and a person still approves every email that goes out.
Use a workflow with an AI step when the input is messy (emails, PDFs, call transcripts, free-text form answers), but what happens after you understand it is always the same.
Level 3: an AI agent with tools
Client build: a product business that tracks stock in Google Sheets. Staff kept interrupting whoever knew the sheet best to ask stock questions or record changes.
We built an AI agent in n8n with a chat window, an OpenAI model, short-term memory and two tools: one that searches the inventory sheet and one that updates or adds rows. Someone types “how many of the large ones are left?” or “we just received 40 more”, and the agent decides whether to search, update, or ask a follow-up question.
No fixed path could cover every way people ask about stock. That’s what makes it an agent’s job: the next step depends on what the person says.
Use an AI agent when requests vary every time, the work needs more than one tool, and the right next step can’t be mapped in advance.
Side by side
| Automation | Workflow with an AI step | AI agent | |
|---|---|---|---|
| Input | Structured | Messy, then structured | Messy and varied |
| Who decides | Your rules | Your rules, plus AI at one point | AI, within limits |
| Same result every time? | Yes | Mostly | No, so it needs testing and guardrails |
| What goes wrong | Skips cases it wasn’t built for | AI mis-sorts an item | AI takes a wrong action |
| Safety net | Fallback branch | Human review of the AI’s output | Limited tools, logging, human approval |
| Build effort | Lowest | Moderate | Highest |
Why so many AI agent projects get cancelled
Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, because of rising costs, unclear business value or weak risk controls. It also warns about “agent washing”: vendors relabelling chatbots and basic automation as agents. By its estimate, only about 130 of the thousands of vendors claiming agentic AI actually offer it.
In our experience the cause is usually simpler: an agent was built for a job that a level 1 or level 2 workflow would have done more cheaply and more reliably. The habits that avoid it:
- Find the decision first. Write the task as steps and circle the ones that need judgment. If nothing is circled, you don’t need AI.
- Give the AI one job. In the email workflow, the AI only classifies and drafts. Sending stays with a person.
- Limit the agent’s tools. The inventory agent can search and update one sheet. It can’t delete rows, email customers or touch anything else.
- Plan for “doesn’t fit”. Every system needs somewhere for unexpected cases to go, like the brokerage workflow’s fallback branch.
- Measure it. Pick one number before you start, such as minutes to assign a lead or replies drafted per day, and check it after launch.
A five-question test for any task
- Can you write the steps as a checklist? Yes means level 1.
- Is the input structured or messy? Messy means you need AI at least once: level 2 or 3.
- After the AI reads the input, is the rest always the same? Yes means level 2.
- Does it need a back-and-forth conversation, or several tools in a different order each time? Yes means level 3.
- What does a mistake cost? The higher the cost, the more you want human approval at the end, whatever the level.
How we decide on client projects
We map the process first and pick the lowest level that does the job: automation for the fixed steps, an AI step where something has to be read or sorted, and an AI agent only where the work really varies. Most of our agent work runs on n8n, with logging, limited permissions and a human handoff built in. For costs, see our AI automation pricing guide.