In 2026, every PM's instinct is to throw an LLM at it. Mine was too, for about a year. I watched us wrap regexes in chatbots, turn dropdowns into "chat experiences," and replace working SQL queries with RAG pipelines. Most of it was worse than what it replaced. Slower, more expensive, less predictable. The senior move is not adding AI everywhere. It is knowing when to take AI out of a flow.
The decision tree I run now
Before wrapping any surface in an LLM, I walk five questions in order and stop at the first yes.
Can a rule do it? If the input space is small or the output is well-defined, a rule, regex, or state machine outperforms a model on cost, latency, and reliability. I do not use a model to check if an email is valid.
Can a query do it? If the answer lives in the database, query for it. RAG is a query with a creative writing flourish on top. "What is my MRR last month" is a SQL query rendered as a number, not a 200K-context model call.
Can a form do it? This one breaks most PMs. Many chat experiences shipping today are worse than the form they replaced. Most agentic SaaS products are forms in chatbot clothing. Take the chatbot off and ship the form.
Can a heuristic and a lookup do it? For ranking, prioritization, or scoring, a weighted heuristic with a lookup table beats an LLM on cost, latency, and explainability. You can debug a heuristic. You cannot debug "the model decided."
Now use a model. By the time you get here you have earned it: open-ended generation, synthesis across many sources, judgment under ambiguity, conversation where the value is the dialogue. A smaller slice of your product than you think.
The hybrid pattern most teams miss
The most powerful pattern is not "use AI" or "do not." It is use a rule for 80% and escalate to a model for the 20%. Support routing: keywords for obvious cases, model only for ambiguous ones, around 80% cost reduction with no quality drop. Document extraction: regex for structured fields, model for the rest, around 90% cost reduction. Search: lexical first, vector as fallback. The PM job is identifying the 80%. Most PMs assume the whole thing is the 20%, ship model-only, and discover they are losing money on every transaction.
The cost case
A cent per call sounds like nothing until you multiply. 100 actions per user per day, 50,000 users, 30 days is $1.5M a month for something a regex could do for free. The reverse is also true: summarizing a 50-page legal document into a one-paragraph brief is worth the cost ten times over. The PM job is knowing which is which, and you know by running both and looking at eval scores against cost.
Pick one AI feature on your roadmap and apply the tree. If you cannot get past step 5 with a one-sentence reason a rule would not work, the feature is an AI feature for marketing reasons, not product reasons. Kill it or ship the simpler version first.