You Don’t Automate Broken Workflows: Prepare the Operation for AI.

Before the agent

Start with the operation, not the model.

When a business asks where AI can help, the first useful question is not which model or agent framework to use. It is how the work currently moves from input to decision to action.

If information arrives through several channels, ownership is unclear, exceptions live in people’s heads and nobody can describe what should happen under different conditions, AI inherits that ambiguity. The result may look impressive in a demo while remaining difficult to trust in production.

The goal is not to make every workflow rigid. It is to distinguish what is deterministic, what requires interpretation and what still needs accountable human judgment.

Five layers of an AI-ready workflow.

What should AI do — and what should ordinary automation do?

A useful architecture does not use an LLM for every step. Traditional automation is often better for deterministic actions: moving data, validating required fields, calling an API, calculating values, changing a status or sending a known notification.

AI becomes useful when the workflow contains unstructured information or interpretation: extracting meaning from documents, classifying requests, summarizing context, drafting responses, matching imperfect descriptions or assisting with decisions that cannot be expressed as a simple rule.

Do not confuse autonomy with value.

The most useful system is not necessarily the one with the least human involvement. A workflow can create substantial leverage even when AI prepares information, recommends an action or handles routine cases while a person approves exceptions.

Human-in-the-loop design is particularly valuable when decisions are difficult to reverse, inputs vary significantly or accountability matters. Autonomy can increase later when actual operating data shows where it is safe and useful.

A practical sequence for introducing AI.

Map one workflow end to end. Identify repetitive information work. Separate known rules from interpretive work. Connect the necessary systems. Add the smallest AI capability that removes meaningful friction. Then observe exceptions before expanding its authority.

This sequence also makes technology choices easier. Sometimes the answer is an agent. Sometimes it is a conventional workflow with one LLM step. Sometimes the business needs an internal tool before it needs either.