Start with the workflow, not the AI

When someone asks me where AI can help, my first question is about the work. What gets copied between systems? Which decision waits for someone to read a document? Where does a team repeatedly lose time checking the same information?

Those questions reveal a workflow. A workflow gives an AI project something concrete to improve.

Define the job before choosing the model

Write down the input, the expected output and the person responsible for the result. “Help with operations” is too broad. “Extract the fields from an incoming request and prepare a draft for review” gives you something you can test.

Sometimes the answer is a SQL query, an integration or a better form. An AI model earns its place when the task involves language, ambiguity or information that is difficult to handle with fixed rules.

Plan for uncertainty

Models make mistakes. The application around the model should make those mistakes visible and manageable.

• Validate the format and required fields before accepting an output.
• Keep the source material available so a person can check the result.
• Route uncertain or consequential cases to human review.
• Limit which systems the model can read or change.
• Provide a useful fallback when a provider is unavailable.

The right boundary depends on the workflow. Summarizing a document and authorizing a payment require different levels of control.

Measure useful work

Agree on a baseline before building. Track time to complete the task, correction rate, queue age or another outcome the team already understands. Compare the complete workflow, including human review and exceptions.

A polished demo shows that an idea can work. A maintained workflow shows whether it actually helps.

Make ownership part of delivery

Someone needs to own the prompts, permissions, monitoring and changes to the surrounding systems. Document those responsibilities alongside the architecture.

My preferred starting point is one bounded workflow with a clear owner. Get that working, learn from real use and expand where the evidence supports it.

Explore AI and automation, or get in touch about a workflow you want to improve.