Decision trees or AI: what to automate with each
Use decision trees for anything with a single correct, predictable answer: opening hours, prices, availability, the steps of a process. Use AI for what arrives phrased a thousand different ways or requires interpreting intent. Putting AI where a tree works better makes the operation more expensive and less reliable.
The question is not which is better, but what each one solves
A decision tree is deterministic: the same input always returns the same answer, it can be audited step by step, and its cost does not depend on how much text passes through it. A language model interprets, generalises, and writes — which is precisely why it does not guarantee the same answer twice, is harder to audit, and costs something on every interaction.
None of those properties is good or bad in the abstract. They are different, and they suit different problems.
What should you automate with decision trees?
Anything with a correct answer where being wrong has consequences: hours, prices, availability, membership terms, order status, the steps of a process.
In the sports club operation we centralised, court bookings, hours, and membership terms were handled by tree. Not because AI could not answer them, but because those answers have to be exact, identical every time, and verifiable by whoever runs the operation. A customer told a court is free when it is not is not a tone problem, it is a business problem.
What should you resolve with AI?
Whatever does not fit a tree without building a hundred branches: open questions, messages that mix several topics at once, phrasings nobody anticipated, and reading intent in order to route a conversation to the right place.
It also works as a comprehension layer in front of the tree: AI works out what the person is asking, and the tree delivers the exact answer. That combination usually beats either one on its own.
The expensive mistake: AI for everything
Resolving with AI what a tree already handled carries three costs. The obvious one is economic: every interaction consumes credits, and service operations are high-volume by definition.
The other two matter more. First, reliability: where accuracy counts, an answer that sounds right but is wrong is worse than no answer. Second, governance: a tree is reviewed branch by branch and you can prove what it replies; a generated answer has to be evaluated case by case. In regulated sectors, that difference decides it.
People are a layer, not a system failure
The goal is not that no conversation ever reaches a human. It is that when one does, it arrives with the full history and context, and to the right person.
A well-divided operation leaves the team the exceptional, the delicate, and the revenue-generating — which is exactly where a person contributes something no automation does.
How to divide it in practice
Before deciding what handles each query, run it through these four filters.
- Is there a single correct answer? If yes, tree.
- Does a wrong answer carry operational or legal consequences? If yes, tree.
- Does the same question arrive phrased unpredictably? If yes, AI in front.
- Can you explain why it answered what it answered? If not, it is not ready for production.