I Turned Down 4 AI Feature Requests in One Quarter. Here's Why That Was the Right Call.
At Sonic Linker, I watched our founding team ship an AI SaaS platform in 3 months. We had the tech, we had the momentum, and suddenly everyone had ideas about what else AI could do in our product.
Sales wanted AI-generated email subject lines. A potential client asked for AI-powered analytics summaries. Our CEO wondered if we should add a chatbot. An advisor suggested AI content recommendations.
I said no to all four. Not because they were bad ideas, but because they failed a test I started using after nearly derailing our roadmap twice.
The Test: Does AI Actually Solve the Job, or Does It Just Sound Cool?
Here's what I learned the hard way. When someone asks for an AI feature, they're usually asking for one of three things:
- A solution to a problem they already have (good)
- A feature their competitor just launched (neutral, depends)
- AI for the sake of having AI (almost always bad)
The sales team's request for AI email subject lines? That was number three. When I dug in, the real problem wasn't writing subject lines. It was that our users didn't have enough context about their recipients to personalize emails at all. AI would've just generated better-sounding garbage.
We fixed it by surfacing recipient data higher in the UI and adding simple merge tags. No model required. Took 2 days instead of 3 weeks, and it actually worked.
The chatbot idea? Same thing. Our users weren't confused about how to use the product. They were confused about whether the product could handle their specific use case. A chatbot would've just been an expensive FAQ that gave vague answers. What they needed was better onboarding and clearer capability documentation.
When I Actually Said Yes: The One AI Feature That Made Sense
We did add one AI feature that quarter. It auto-categorized user-generated content based on intent, something users were doing manually in spreadsheets for hours.
Why did this one pass the test?
First, the manual alternative was genuinely painful. I watched a user spend 90 minutes categorizing 200 items. She had a system, but it was tedious and error-prone.
Second, the AI made the task 10x faster, not 10% better. We weren't adding a slight improvement. We were removing a time sink entirely.
Third, failure was obvious and fixable. If the model miscategorized something, users could see it immediately and correct it. The AI didn't make irreversible decisions or hide its logic.
That feature became one of our top retention drivers. Not because it was AI, but because it solved a real problem that users cared about deeply.
The Red Flags I Now Look For
I've developed a quick gut check before saying yes to any AI feature request:
Is the person asking able to describe the current workaround? If they can't tell me how users solve this problem today (even badly), it's probably not a real problem. At Finvestfx, I managed 20+ enterprise clients. The features that stuck were always the ones where clients could walk me through their current janky process. If they couldn't, it meant they weren't actually doing the thing we were trying to automate.
Would a simple rules-based system get us 80% of the way there? AI is expensive to build and maintain. If you can hardcode logic that covers most cases, do that first. You can always add the model later when you hit the edge cases that matter.
Can we measure whether the AI output is actually better than the alternative? If you can't define what "better" means in concrete terms, you can't know if your AI feature is working. I learned this at Sonic Linker when we almost shipped a feature with no clear success metric beyond "users like it." That's not a metric. That's hope.
What I Tell Founders Now
AI is not a product strategy. It's a tool that sometimes happens to be the right solution.
The best AI features I've seen (including the ones we shipped) all had one thing in common: you could describe the value without mentioning AI at all. "It categorizes your content automatically" is compelling. "It uses AI to categorize your content" is just noise.
When someone pitches you an AI feature, ask them to describe it without using the words "AI" or "machine learning." If they can't make it sound valuable in plain language, it's probably not valuable at all.
I'm not anti-AI. I spent months building an AI product. But I am anti-waste, and I've seen too many teams burn weeks on AI features that users tolerate but don't actually need. The best product decisions I made weren't about saying yes to clever ideas. They were about saying no to everything that didn't pass the "does this actually matter" test.
AI just happens to fail that test more often than founders want to admit.