The customer asks you to automate operations but cannot name a specific workflow. How would you begin?

Instruction: Use a clearly hypothetical workflow if you need an example. Explain how you select an engineering problem before selecting an automation technology.

Context: Turns a vague automation request into one bounded user task with an observable result and an explicit failure path.

Updated

Example Answer

I'd ask the customer to show me where work currently waits, gets repeated, or produces a costly mistake. I'd talk to the people doing that work and inspect an approved example of the input and finished result. Automate operations is too broad to build against, so my first goal would be a specific task and its owner.

In a hypothetical order-dispatch workflow, I might learn that operators repeatedly compare order details with warehouse availability before proposing a new dispatch date. I'd clarify which information is authoritative, when it becomes stale, and who may approve a change. I'd separate gathering the facts from deciding an exception so I do not accidentally automate authority nobody agreed to delegate.

I'd propose a narrow first slice that prepares the comparison and leaves the decision with the operator. We'd test whether it reduces the actual manual steps without hiding missing information or creating extra corrections. Only then would I consider automating more of the workflow, using rules or AI according to the variation we observe.

Make it your own

Replace dispatch with a workflow you understand. Keep the example hypothetical unless you can support it with your actual experience.

Why this works

It creates a buildable boundary, identifies business authority and tests workflow improvement. The employer reference establishes discovery as part of FDE work, not a requirement to use AI for every task.

Interviewer follow-up

What if the sponsor insists the first version must use an AI agent?

I'd ask what capability they expect from the agent that the simpler implementation cannot provide. I could test AI on an uncertain step, such as interpreting an operator's note, while keeping approved rules and final writes outside that experiment. I'd compare the extra correction work and failure cases with the baseline, then recommend the design that meets the workflow need.

Assessment criteria

These are practice criteria for this scenario, not an employer's scoring rubric.

  • Strong: Turns the broad request into a named task, authoritative inputs, bounded implementation and an outcome check.
  • Adequate: Clarifies a workflow and proposes a small prototype with visible assumptions.
  • Weak: Chooses an agent before establishing what work should improve.

A tempting weak answer

"I'd connect their data to an agent and let them explore it."

Why it fails: The proposal leaves the task, authority and correctness criteria undefined.

References

Related Questions