Most real projects are mostly not AI
A typical automation we build has one or two AI steps and a dozen deterministic ones. The AI reads the email; rules decide what happens next.
That mix is deliberate. Rules are cheaper to run, instant, perfectly explainable and impossible to talk into doing something strange.
Where AI earns its place
- Extraction from documents, emails and forms — unstructured in, structured out
- Classification — what is this, where should it go
- Answering from your own material, with citations
- First drafts a human reviews and owns
All four share one property: a person can verify the output in seconds. That is what makes them safe to deploy.
Where it stays out
Matching, tolerances, routing and posting are rules. You need to be able to read them, change them without a retrain, and explain them to an auditor. A probability is not an explanation.
Anything with legal or financial consequence where the decision must be defensible also stays deterministic, with AI at most surfacing information for a person.
The confidence boundary
Every AI step returns a confidence figure, and the rules use it: high confidence flows through, low confidence routes to review, and the thresholds are set per field by consequence.
A supplier name being wrong is recoverable. A bank account number being wrong is not. Different fields deserve different bars.
What this means for cost
- Route by difficulty — a cheap model for classification, a stronger one only where needed
- Send less context; most retrieval sends more than accuracy requires
- Cache the stable parts of prompts
- Cap spend per day and alert before the cap
Done this way, the AI portion of a business automation typically costs pennies per unit of work rather than being the dominant line.