AI is attractive when a process contains language, judgement, or a large number of possible inputs. Those characteristics make a model relevant. They do not automatically make a model the right intervention.

Many automation projects begin with a description such as: “Someone reads the request and decides where it goes.” That can sound like a classification problem. A closer look may reveal that most requests follow a handful of explicit rules, while the difficult cases depend on missing information or unresolved ownership.

Adding a model does not resolve either problem.

Map the decision, not just the task

Before choosing an implementation, identify:

  • What information is available when the decision is made
  • Which outcomes are possible
  • Which rules are already explicit
  • Where judgement genuinely occurs
  • How exceptions are handled
  • What makes a decision correct
  • What happens when confidence is low

This often separates the workflow into different kinds of work.

Some parts are deterministic and should remain rules. Some need better data validation. Some need a systems integration. Some need a redesigned form. A smaller part may benefit from language understanding or probabilistic ranking.

Prefer the simplest dependable mechanism

Rules are easier to test, explain, and change when the policy itself is deterministic. They are usually cheaper to operate and do not introduce variable behaviour.

A model becomes valuable when the input cannot reasonably be reduced to stable rules: interpreting varied natural language, extracting meaning from unstructured material, ranking uncertain evidence, or supporting a human decision across a large possibility space.

The most dependable systems often combine both. A model interprets an unstructured input; rules enforce permissions, thresholds, routing constraints, and consequential actions.

Keep policy outside the prompt

If a business rule must always hold, hiding it inside a prompt makes the control difficult to inspect and unreliable to enforce. Prompts can guide behaviour, but they should not be the only place where a hard policy exists.

Explicit rules create a boundary around the model:

  • The model proposes; a rule validates
  • The model classifies; a threshold determines automation or escalation
  • The model extracts; deterministic checks verify required fields
  • The model drafts; an authorised person approves a consequential action

Audit before automating

The goal is not to minimise the use of AI. It is to apply it where uncertainty is real and useful, while keeping stable decisions explicit.

Mapping the process first produces a smaller technical problem, clearer acceptance criteria, and a system that is easier to operate when conditions change.