Guide · Automation vs AI Workflow vs AI Agent
Three terms that get mixed up constantly. Here's what each one actually is, what it costs, and which one your business needs, in plain English.
The distinction that matters is who controls the path. In automation and AI workflows, you do. With an agent, the AI does. That single difference drives everything else: reliability, cost, auditability, and what each is safe to use for.
| Automation | AI Workflow | AI Agent | |
|---|---|---|---|
| What it is | Fixed rules: when X happens, do Y | A designed sequence of steps where AI handles the judgement steps | A goal plus tools; the AI picks its own steps |
| Handles messy input? | No. Inputs must be structured and predictable | Yes. Reads emails, PDFs, free text, photos of invoices | Yes, plus it can go looking for more input on its own |
| Predictability | Total. Same input, same output | High. The path is fixed; only the judgement inside each step varies | Lower. The path itself can change run to run |
| Auditability | Easy | Easy when built properly: every prompt, output and decision logged | Harder. You must reconstruct why it chose the steps it did |
| Typical NZ SME example | New booking lands, send a confirmation text | Inbound email read, classified, customer history pulled from the CRM, reply drafted for one-click approval | "Research this new lead and update the CRM", where the agent searches, reads and writes on its own |
| Running cost | Cents. No AI calls | Low. A few AI calls per item, usually well under $0.10 | Higher. Many AI calls per task, and slower |
| Best for | High-volume, identical, structured tasks | High-volume tasks that need reading and judgement | Open-ended tasks with low stakes or human review at the end |
| Watch out for | Breaks the moment input varies | Needs guardrails and human checkpoints designed in | Unpredictable failures; cost creep; hard to debug |
Work up the ladder, not down it. Use plain automation wherever the input is structured; add an AI workflow where a person currently reads and judges; reserve agents for narrow jobs where an unexpected result is annoying rather than expensive.
A useful test for each rung:
Out of the box, no. ChatGPT is a model you converse with; it acts only when you prompt it. It becomes agent-like when it's given tools (web browsing, file access, code execution) and a goal to pursue across multiple steps. The same underlying models power automations, workflows and agents; the difference is the harness around them.
Classic Zapier/Make flows are plain automation: triggers and structured actions. Both now let you insert AI steps, which turns a flow into a simple AI workflow. Where they run out of road is deep integration (legacy systems, complex business rules, audit requirements), which is when a custom build makes sense. I cover this in my AI workflow automation page.
Yes, for the right jobs with the right guardrails: scoped tools, spending and step limits, full logging, and a human checkpoint before anything customer-facing or financial happens. What's risky is giving an agent broad access to live systems and letting it act unsupervised on high-stakes work.
Plain automation is cheapest, then AI workflows, then agents. For NZ SMEs, a custom AI workflow typically lands in the NZ$500 to $15,000 fixed-price range and pays for itself within 3–6 months. Agent projects cost more to build and more to run, and need a stronger business case. My 8-step implementation guide covers how to scope this properly.
Agents keep improving, and the line will keep moving. But businesses will always want predictable, auditable paths for core processes, for the same reason they use checklists and standard procedures with human staff. Expect agents to take over more of the open-ended edges while workflows keep running the middle.
Book a free 30-minute call and describe the task. I'll tell you whether it's an if-statement, a workflow or an agent, and what it would honestly cost. Discovery is always free.
Book a free strategy call