AI automation means two different things. Label them.

AI Automation: AI automation is the use of machine learning models — in practice, large language models — to perform steps in a work process that previously required a human to read, judge, or write.

Where the term came from, and the date that matters

Merriam-Webster dates the first known use of "automation" to 1948, out of the US auto industry, where it meant machines moving parts between machines. For seventy years after that, automation meant deterministic. A fixed path, written by a person, executed identically every time. Zapier, Workato, RPA scripts, cron jobs. Same input, same output, no judgement anywhere in the chain.

The compound "AI automation" only becomes a coherent category on 13 June 2023, when OpenAI shipped function calling and a model could reliably emit a structured call into your code rather than just prose. Before that, putting a model in a pipeline meant parsing free text and hoping. After it, the model could be a step.

No standards body defines "AI automation." It is a market term, not a technical one.

The two things people mean by it

They are not the same system, they do not fail the same way, and vendors sell both under one word. Anthropic's engineering team draws the same line, calling them workflows and agents: workflows run predefined code paths, agents direct their own process.

	Type 1 — Model in a fixed path	Type 2 — Model-directed loop
Who chooses the next step	A human, at build time	The model, at run time
Shape	Trigger → step → model call → step	Goal → loop of tool calls → stop condition
Number of steps	Known before it runs	Unknown until it stops
Typical failure	Bad classification, one record wrong	Unbounded loop, wrong tool, cost spike
What you test	Output quality on a fixed set	Whether it terminates, and what it touched

Type 1 is most of what ships. Type 2 is most of what gets written about. If someone says "AI automation" and you do not know which one they mean, you do not know what you are buying. I wrote about that split under its other name in agentic future: one word, two claims.

What it is not

It is not RPA. Robotic process automation clicks a recorded path through a user interface. It carries no judgement and no tolerance: change a selector, rename a field, move a button, and the script fails loudly. An RPA bot that reads an invoice is doing optical character recognition against a fixed template. An AI automation that reads an invoice is producing a probabilistic guess about which number is the total. The second one handles formats it has never seen and is sometimes confidently wrong. Those are different risk profiles, not different speeds.

It is not an agent, necessarily. Every agent is AI automation. Most AI automation is not an agent. The distinguishing property is who decides the next step at run time. If you can draw the full path on a whiteboard before it executes, it is a workflow with a model in it — and that is the right build. Computer-use agents are a last resort for the same reason: more autonomy buys you capability you did not need and costs you reliability you did.

It is not "AI-powered." Autocomplete in a text box is a feature. Automation means the step runs without a human in the chair.

One worked example, costed

Inbound form fills arriving at a B2B site: 1,400 a month. Each one needs reading, tagging as support / sales / recruiter spam, and routing. A person doing it at two minutes each spends roughly 47 hours a month.

The Type 1 build: webhook fires, an enrichment API call attaches company data, one model call returns a JSON label and a confidence, the workflow writes to the CRM and Slack. One model call, no branching. That call is around 1,200 input tokens and 120 output tokens. Over 1,400 records that is 1.68M input and 168K output tokens. At gpt-4o-mini's listed rate of $0.15 per million input and $0.60 per million output, the model spend is about $0.35 a month.

Thirty-five cents. The model is never the expensive part at this volume — the per-task billing on the orchestrator, the engineering hours, and the cost of a misrouted enterprise lead all dominate it. That gap is why I argue for building a margin meter before you price an AI feature: people size the token bill and miss the retry, review, and rework bill underneath it.

Now make it Type 2 — "handle inbound and book meetings where appropriate" — and the step count per record goes from one to unknown. Same business goal. Entirely different thing to operate.

How to tell whether something qualifies

Four tests, in order:

Does a model make a decision that a person used to make? If no, it is plain automation with an API call in it. Classification, extraction, drafting, triage, routing all count. String concatenation does not.
Does it run without a human initiating it? A chat window where someone pastes a prompt is assistance. A webhook, a cron, or a queue consumer is automation.
Does it write somewhere? Automation that only produces text for a human to approve is a drafting tool. Useful. Not the same category, and not the same risk.
Is there a cap enforced outside the model? For Type 2 this is the test that separates production from demo. Unbounded loops are the most common production failure in model-directed systems. The fix is a hard step and token ceiling in your code — not a sentence in the prompt asking the model to be brief.

On reliability, the published benchmarks are the honest reference point. WebArena reported a GPT-4 agent completing 14.41% of realistic web tasks against 78.24% for humans. OSWorld reported the best model at 12.24% on real desktop tasks against 72.36% for humans. Those numbers are from 2023–2024 and current models do better, but the shape holds: narrow and bounded works, open-ended computer operation does not yet. Tool plumbing got easier — the Model Context Protocol, published November 2024, standardised how models connect to tools and data — and plumbing was never the bottleneck.

I took Metadata.io from $0 to $15M ARR and I am now building a company where agents do the work. Start with Type 1. Make one step boring. Reach for the loop only when the fixed path genuinely cannot express the task, and build one bounded agent before you build a fleet.