Good questions.
Useful experiments.
A little curiosity.

A sensible starting point is the decision and its commercial consequence. What changes if it gets better? How often does it happen? Is your company’s own knowledge essential to getting it right? Can you measure progress without mistaking activity for value?

If existing software solves the problem well, use it. If the volume is small or the cost of mistakes is high, a person may be the right answer. If data is missing, fix that first. Custom engineering earns its place when company-specific logic creates an advantage large enough to justify building and maintaining it.

That still leaves room for ambition. Some of the best opportunities are things a team could not previously afford to do consistently: checking thousands of signals against its own expectations, for example. Start with a narrow, valuable loop that can prove the idea. Its first job is to produce evidence, not justify an empire.

A marketing team knows that promotions often dip halfway through. A sales team recognizes customers who look promising but rarely convert. Finance knows which apparently healthy sales disappear into returns and low margins. Some of this is in data, some in operating rules, some in people’s heads.

Collecting all of it is not the goal. Ask what would change the decision. A campaign alert that ignores your usual response curve can be technically correct and operationally useless. A system that knows the curve, and its limits, can direct attention to a deviation that deserves investigation.

That knowledge needs testing too. A familiar rule may be outdated; a historical relationship may not survive a new market or offer. Useful systems make those assumptions visible and revisit them as outcomes come back. Internal expertise makes the system specific. Evaluation keeps it honest.

Start with the business outcome you want to affect. Incremental profit is a strong target when you can establish a credible comparison. Customer lifetime value and qualified leads can better fit longer buying cycles. A rise in output or a fall in time spent is useful information, but it does not by itself prove a business gain.

Sometimes the first measurable improvement is a better decision, fewer errors or more tests your team can run. Name it clearly. Establish a baseline and decide how you will evaluate quality. Do not quietly relabel a proxy as profit, or an observed change as a causal effect.

A production system needs ongoing evaluation as well as a launch test. Track whether its inputs still mean the same thing, whether people act on its recommendations and whether the economics still make sense. The right answer can be to adjust the system, narrow its remit or stop it.