Framework for picking rules, code, ML, generative AI or agents for a product problem

As a product manager, how do you decide whether a problem calls for a simple rule-based approach, conventional programming, machine learning, generative AI, or an agent-based system? ```hint Climb only when forced Picture the cheapest option that could possibly work for a given problem, then ask what specific property of that problem would force you to move to something more capable. ``` ### Clarifying Questions - Does "AI" here mean predictive machine learning such as classifiers and rankers, as distinct from generative models? - Should I give a general framework, or apply it to a specific LinkedIn problem you have in mind? - Are unit cost and latency at LinkedIn's scale part of the decision, or only quality and fit? ### What a Strong Answer Covers - Criteria that separate the options, such as how fully the logic can be specified, the shape of inputs and outputs, error tolerance and the need to explain or audit decisions - A default toward the simplest approach that meets the bar, and the evidence that justifies moving to machine learning, generative models or agents - Data needs, unit cost, latency and long-term maintenance of each option - Hybrid designs in which rules or deterministic code constrain or act on model output - At least one concrete product example worked through the framework ### Follow-up Questions - Your team wants to replace a rule-based system that screens job postings with a large language model. What would you need to see before agreeing? - When, if ever, would you let an agent take an action on a member's behalf without asking for confirmation? - A generative feature tested well, but its cost per use is too high at full scale. What are your options?