When Not to Use AI

The instinct everyone has right now

In 2026, every PM's instinct is: throw an LLM at it. Mine was too for about a year. I watched us wrap regexes in chatbots, turn dropdowns into "chat experiences," replace working SQL queries with RAG pipelines, and generally add a model call to every workflow we touched.

Most of it was worse than what it replaced. Slower. More expensive. Less predictable. Users had to type out what a dropdown would have collected in one click. The model had to infer what the form would have constrained. Both sides did more work for a worse outcome.

The senior move in 2026 isn't adding AI everywhere. It's knowing when to take AI out of a flow. That's the skill that separates product builders from hype builders.

The decision tree I run now

Before wrapping any surface in an LLM, I walk through these questions in order. I stop at the first "yes."

1. Can a rule do it?

If the input space is small or the output is well-defined (validation, parsing, format conversion, classification with under ten categories), a rule or regex or state machine will outperform a model on cost, latency, and reliability. I don't use a model to check if an email is valid. I don't use a model to extract a phone number. I don't use a model to route between three workflows when an if/else does it cleanly.

2. Can a query do it?

If the answer lives in the database, I query for it. RAG is a query with a creative writing flourish on top. When a user asks "what's my MRR last month," that's a SQL query rendered as a number, not a 200K-context-window model call. Give them a chart.

3. Can a form do it?

This one breaks most PMs' brains. Many "chat experiences" shipping in 2026 are worse than the form they replaced. Chat is the right interface when the user is exploring, the inputs are open-ended, and the value comes from synthesis. Form is the right interface when the user knows what they want, the inputs are enumerable, and the cost of error is non-zero.

Most agentic SaaS products are forms in chatbot clothing. Take the chatbot off. Ship the form. Ship faster.

4. Can a heuristic and a lookup do it?

For ranking, prioritization, or scoring problems, a weighted heuristic with a lookup table beats an LLM on cost, latency, predictability, and explainability. You can debug a heuristic. You can A/B test a weight. You cannot debug "the model decided." Save the model for the hard cases.

5. Now use a model.

By the time you get here, you've earned it. The remaining cases are: open-ended generation, synthesis across many sources, judgment calls under ambiguity, conversational interfaces where the value is in the dialogue. These are real LLM use cases. They're a smaller slice of your product than you think.

The hybrid pattern (the one most teams miss)

The most powerful pattern in 2026 isn't "use AI" or "don't use AI." It's "use a rule for 80%, escalate to a model for 20%."

A few examples from products I've worked on:

• Support ticket routing. Rule-based on keywords for the obvious cases. Model only for ambiguous tickets. About 80% cost reduction with no quality drop.
• Document extraction. Regex for structured fields (dates, amounts, IDs). Model for the unstructured remainder. About 90% cost reduction.
• Search. Lexical first. Vector as fallback when lexical returns under five results. Cuts retrieval cost in half.
• Content moderation. Rule-based for slurs and known bad patterns. Model only for ambiguous edge cases. Faster, cheaper, more legally defensible.

The PM job is identifying the 80%. Most PMs assume the whole thing is the 20%, ship a model-only solution, and discover they're losing money on every transaction.

The "form versus chat" gut check

For any chat interface in your product, ask three questions:

1. Could a dropdown collect the same input?
2. Could a 3-5 step wizard replace the conversation?
3. Are users typing roughly the same phrases repeatedly?

If yes to any of these, the chat is wrong. Every PM eventually learns this by shipping a chatbot, watching the analytics, and discovering 70% of the user inputs are the same five phrases. That's a form.

The cost case

A 1,500-token prompt with a 500-token completion on a flagship model costs about a cent. Sounds like nothing. Now multiply: 100 actions per user per day, 50,000 active users, 30 days. That's $1.5M a month for a feature a regex could have done for free.

Your CFO will eventually find this. Better that you find it first.

The reverse is also true. A feature that genuinely needs a model (like summarizing a 50-page legal document into a one-paragraph brief) is worth the cost ten times over and shouldn't be cheaped out by a rule that produces garbage. The PM job is knowing which is which. The way you know is by running both and looking at the eval scores against the cost.

The cultural fight

A peer (sometimes a founder, sometimes the CEO) will want the product to be "AI everywhere" because that's what the market is rewarding. They're confusing a marketing position with a product decision.

The market rewards companies that use AI thoughtfully. It doesn't reward companies that wrap a regex in a chatbot for the demo. The customer eventually figures out which is which. So does the bill payer. Credibility comes from the LLM working brilliantly where it matters and being absent where it doesn't. That's the AI-native posture.

Pick one thing this week

Pick one AI feature in your current roadmap. Apply the decision tree to it.

1. Could a rule do it? If yes, what's the rule?
2. Could a query do it? If yes, what's the SQL?
3. Could a form do it? If yes, sketch the form.
4. Could a heuristic do it? If yes, what's the formula?
5. If it genuinely needs a model, write one sentence explaining why a rule wouldn't work.

If you can't make it past step 5 with confidence, the feature is an AI feature for marketing reasons, not product reasons. Either kill it or ship the simpler version first and see if users notice.

AI is a tool. Putting it down when something cheaper does the job is the move that actually makes you look senior.