Automation, internal tool, or AI agent?

Most requests we get start the same way: a process is eating hours, and something should be done about it. The interesting question is which of three quite different things the problem actually needs — because they differ by roughly an order of magnitude in cost, and picking the expensive one for a cheap problem is the most common way this money gets wasted.

The three options

An automation moves data between systems on a trigger. No interface, no judgment, no state beyond what the tools already hold. A form submission creates a CRM record and posts to Slack. Built in n8n, Make, or a small amount of custom code.

An internal tool gives people a place to do work. It has a database, an interface, permissions, and a notion of who did what. A dashboard, an admin panel, an approvals queue. Built properly, as software.

An AI system handles the cases where the rule cannot be written down — reading unstructured documents, answering open-ended questions, classifying things that resist a clean taxonomy. It also brings evaluation, monitoring, and failure behavior along with it, which is most of its real cost.

The questions that separate them

Four questions, asked in this order, resolve most cases before anybody writes a line of code.

1. 01Can you write the rule down completely? If yes, you want an automation. Not 'mostly' — completely, including the exceptions. If you can specify it, you should not be paying a model to guess it, and you should not be paying us to build an interface around it.
2. 02Do people need to see or decide something? If humans have to review, approve, prioritize, or simply look at the state of things, you need an interface. An automation with no interface becomes a black box that someone has to trust blindly, and they won't.
3. 03Is the input unstructured? Free text, documents, images, email bodies, anything a person currently reads before acting. This is the honest trigger for AI, and it is much narrower than the current enthusiasm suggests.
4. 04What does being wrong cost? Low cost and easily reversible means you can tolerate a model's error rate. High cost or hard to reverse means you need a human review step regardless, and that step is an internal tool whether or not there is a model behind it.

Where people get it wrong

The dominant error at the moment is reaching for an AI system when the rule is perfectly writable. It is more expensive, less reliable, harder to debug, and it introduces a monitoring burden that would not otherwise exist. If the logic is 'when an invoice arrives from one of these twelve suppliers, file it under that supplier', that is an automation. A model will do it correctly most of the time, which is worse than a rule that does it correctly all of the time.

The second error runs the other way: building an elaborate internal tool whose real purpose is to let a human do a job a model could do, because 'AI' felt risky. The tell is that the tool's main screen is a queue of items where a person reads something and picks one of four options. That is a classification problem with a review interface on top, and the review interface is the part worth keeping.

The expensive mistake is not choosing the wrong tool. It is choosing before anyone has written down what the process actually does.

Most real answers are combinations

In practice the good answer is usually two of the three. Document extraction is an AI system that classifies and extracts, plus an internal tool where anything below the confidence threshold gets reviewed by a person, plus an automation that files the approved result. Each part does what it is good at.

That combination is also why the sequencing matters. Build the automation first, because it is cheap and it reveals how the process really works — including the exceptions everybody forgot to mention. Add the interface when people need to see or intervene. Add the model last, once you know exactly which decision it is being asked to make and what happens when it gets that decision wrong.

A shortcut

If you want one heuristic instead of four questions: write down the process as if you were training a new hire. If the instructions fit on a page and contain no sentence beginning 'it depends', it is an automation. If they contain judgment calls, you need either an interface or a model — and which one depends entirely on what it costs to be wrong.