There is a lot of pressure at the moment to put AI into things. Some of it comes from boards, some from suppliers, and almost none of it from the process itself.
Ordinary automation — rules, integrations, scheduled jobs, form logic — is cheaper to build, cheaper to run, easier to test and predictable in a way that a model is not. It should be the default. AI earns its place only where the ordinary version cannot do the job.
The test
For any given step, ask: is the input structured, and is there a right answer?
Structured input, right answer. Use ordinary automation. “When payment clears, create the booking” needs no intelligence at all. Putting a model in the path of a rule this clear adds cost, latency and a failure mode you did not previously have.
Unstructured input, right answer. This is AI’s best case. An invoice arrives as a PDF in a layout nobody has seen before, but there is exactly one correct total. The model is doing perception, not judgement, and you can check its work — which is why confidence thresholds work so well here.
Structured input, judgement required. Usually a person, sometimes a model as an assistant. If the data is already clean and the difficulty is deciding what to do about it, the difficulty is the part you should not hand over.
Unstructured input, judgement required. Be careful. This is where AI is most impressive in a demo and most dangerous in production. Build it so the model prepares and evidences, and a person decides.
Signs you are using AI where a rule would do
- You can write the logic down in a paragraph without using the word “usually”.
- The same input must always produce the same output, and someone would notice if it did not.
- The step is being tested by comparing the output against a rule you already wrote.
- You are paying per call for something a lookup table does for free.
Signs a rule is not going to be enough
- The rules already run to dozens of cases and someone adds another every month.
- The input arrives as prose, or as documents in layouts you do not control.
- Every attempt to write the logic down ends in “it depends on the context”.
- The current process depends on one experienced person recognising something.
The combination that usually wins
Most real workflows are neither, they are both. The model handles the messy edge — reading what arrived, working out what it is, pulling out what matters — and hands structured data to ordinary automation, which does the deterministic part reliably and cheaply.
That split has a useful property beyond cost: when something goes wrong, you can tell immediately which half was responsible.