What an AI automation business actually is
Two different companies use the same name — one sells automations to clients, one runs itself on them. Here is the line between them, and the test that separates both from RPA.
- AI Automation Business
- An AI automation business is one whose recurring work is done by software that decides as well as executes — either sold as a service to clients, or run as the company's own operating model.
Where the term came from, and the date that matters
OpenAI shipped function calling on June 13, 2023. Before that date a model produced prose; after it, a model returned enrich_company(domain="acme.com") — a structured call into your code. Every product sold as "AI automation" since depends on that mechanism or a copy of it.
The automation half is older and duller. Zapier has been selling trigger-action automation since 2011, and RPA vendors like UiPath go back further, scripting the user interface of software that never shipped an API — software robots following rules a person wrote. Gartner stacked those pieces into "hyperautomation" and put it on a trend list.
None of that is this term.
Three things then made it buildable by small teams. Anthropic published "Building effective agents" in December 2024, which drew the line between a workflow (fixed path, model inside one step) and an agent (model chooses the path). The Model Context Protocol landed November 25, 2024, standardising how a model reaches tools and data. OpenAI's Agents SDK followed in March 2025.
Eighteen months. That is the whole history of the thing people now put on a landing page.
Two different businesses wear this name
They share almost nothing except the phrase.
Sense A is an agency. A good one is very profitable and the build skills are real, but the margin comes from hours, and the automation is an artifact you hand over.
Sense B is what I am building publicly as a zero-human company — a business where agents do the work and I pay the inference bill. I spent eight years on the other side of this: Metadata.io, $0 to $15M ARR, $50M raised, six patents in AI-driven marketing, with a human team doing the operations. The interesting question now is which of those operations needed the human.
The label is applied to both senses in common usage, which is why the two are routinely conflated. They diverge immediately on revenue model, on how headcount tracks volume, and on who carries the operational risk.
What it is not: RPA, and a workflow with a prompt in it
RPA executes a path somebody drew. If the invoice PDF changes layout, the robot fails, and it fails loudly — which is its best feature. Determinism is a property, not a limitation.
An agent picks the next step at runtime. Same goal, different route depending on what it observes. That buys you tolerance for messy input and costs you predictability.
That dashed box is the part people skip. Unbounded loops are the most common production failure in agent systems: the model re-reads the same page, re-calls the same tool, burns tokens until something times out. The fix is a step cap and a spend cap enforced outside the model. LangGraph ships a recursion limit for exactly this reason, and Anthropic's own guidance is to define stopping conditions and human checkpoints rather than trust the prompt. A prompt that says "do not loop" is not a cap.
More on the three caps I would start with: one agent, three hard caps.
A worked example, with the arithmetic shown
Inbound lead qualification. A CSV of 2,000 signups lands weekly. For each row: resolve the email domain, call an enrichment API, read the company's own site copy, score fit against your ICP, write the score and a one-line reason back to the row, route anything above 70 to a human.
Token arithmetic, using OpenAI's published list price for gpt-4o-mini at $0.15 per million input tokens and $0.60 per million output:
| Item | Volume | Cost |
|---|---|---|
| Input | 2,000 × 3,000 tokens = 6.0M | $0.90 |
| Output | 2,000 × 300 tokens = 0.6M | $0.36 |
| Model total | — | $1.26 |
The model is the cheap part. The enrichment API, the proxy for page fetches, and the retry budget are where the real bill sits — which is why "what does the model cost" is the wrong first question. The build is written up step by step in qualifying inbound leads from a CSV, and the rent-versus-self-host decision for the surrounding infrastructure is here.
Note what the agent does not do in that example. It does not email the lead and it does not touch spend. I would not hand an agent spend authority on day one, and I would not let it send external mail before I had a week of logs I had actually read. That is a judgement, not a rule of physics — but it is the one I hold.
How to tell whether something qualifies
Four questions.
| Question | Qualifies | Does not |
|---|---|---|
| Who chooses the next step? | The model, at runtime | A branch a developer drew |
| What happens on an input nobody anticipated? | Attempts, logs, escalates | Errors out |
| Where are the limits enforced? | Step and spend caps in code | In the prompt |
| Where does the margin come from? | Software runtime | Billable hours per client |
If a system fails the first question, it is workflow automation with a prompt in one node. That is often the right build — cheaper, auditable, easier to debug — and calling it an agent does not improve it.
If it fails the last one, it is Sense A: an agency with an LLM in the stack. Fine business. It is not the same business as the one that runs on the software it built, and the two should stop sharing a name. What stays human in either case is a separate decision, and I have written about where I draw that line.
Sources
- Function calling and other API updates — OpenAI — June 13, 2023 launch of function calling: the point a model could return a structured call to your code
- Building effective agents — Anthropic Engineering — December 2024 post drawing the workflow/agent distinction and recommending stopping conditions and human checkpoints
- Introducing the Model Context Protocol — Anthropic — MCP announced November 25, 2024; standardised tool/data access for models
- New tools for building agents — OpenAI — March 2025 release of the Responses API and Agents SDK
- OpenAI API pricing — Per-million-token list prices used in the worked example
- Robotic Process Automation — UiPath — Vendor definition of RPA as software robots following predefined rules/UI steps
- About Zapier — Zapier's founding and trigger-action automation model
- n8n pricing — Self-hosted versus cloud execution-metered pricing for workflow automation
- LangGraph documentation — Graph-based agent orchestration with explicit loop and recursion limits
- Hyperautomation — Gartner Glossary — The analyst term people confuse with AI automation