I Almost Added AI to Our Dashboard Just Because Everyone Else Was. Here's Why I Didn't.

The pitch sounded perfect

Last year at Sonic Linker, we had users asking if our platform could predict campaign performance using AI. Investors were asking when we'd add "AI-powered insights" to the roadmap. Our competitor just announced an AI feature. The pressure was real.

I spent two weeks working with our AI team to prototype predictive analytics for link performance. The model was decent, accuracy was around 75%. We could have shipped it.

But I killed it.

Not because the tech was bad. Because I realized we were solving the wrong problem. Our users weren't struggling to predict performance. They were struggling to understand why their current campaigns were underperforming. They needed better diagnostics, not predictions.

That's when I learned that the real question isn't "can we add AI here?" It's "what problem are we actually solving, and is AI the simplest way to solve it?"

Three questions I now ask before adding AI

1. Can users articulate the problem without mentioning AI?

When I went back to customer calls, nobody said "I wish I had AI predictions." They said things like "I don't know why this link isn't converting" or "I can't figure out which audience segment to target next."

If users are asking for AI specifically, it's usually because they think that's the only way to solve their problem. Your job is to understand the actual problem. Sometimes the answer is better filters. Sometimes it's a clearer data visualization. Sometimes, yes, it's AI.

At Finvestfx, we had enterprise clients asking for "AI-powered forex rate predictions." What they actually needed was faster alerts when rates hit their thresholds. We built a better notification system instead. Took two weeks, solved the real problem.

2. Does the user need to trust the output completely?

AI is probabilistic. It's going to be wrong sometimes. The question is: can your user workflow handle that?

For our link performance predictor, the workflow would have been: AI predicts campaign will underperform → user decides whether to cancel or adjust it. But if the prediction was wrong (and 25% of the time it would be), we'd either waste a good campaign or let a bad one run.

The trust threshold was too high. Users needed certainty for that decision.

Compare that to our AI-powered link categorization feature, which we did ship. If it miscategorizes a link, the user just recategorizes it manually. Low stakes, high tolerance for error. That's when AI works.

3. Is there a simpler version that gets 80% of the value?

This is the one that kills most AI features for me.

At Sonic Linker, instead of building predictive analytics, we built a "similar campaigns" comparison view. It showed users how their current campaign was performing versus similar past campaigns (same audience size, same content type, same time of day).

No AI. Just smart filtering and basic stats. Users got the insight they actually needed ("is this normal or should I be worried?") without the complexity or uncertainty of predictions.

It took three days to build instead of three weeks. And honestly, users loved it more because they understood exactly how it worked.

When AI actually is the right call

I'm not anti-AI. We use it heavily at Sonic Linker, just not everywhere.

We added AI for content suggestions (generating link descriptions from page content). That made sense because: - Users explicitly wanted to save time writing descriptions - They could edit the output, so trust threshold was low - The alternative (writing it manually every time) was genuinely painful - There was no simpler solution that got close to the value

That feature ships value every single day. Because we added it for the right reasons.

The real test

Here's what I do now: before speccing any AI feature, I write down the user workflow without AI. Then I write it with AI.

If the AI version is only marginally better, or if it introduces new decision points the user has to think about, I don't build it. If it removes painful steps or enables something genuinely impossible before, I do.

At Sonic Linker, this framework killed about 60% of the AI features we initially planned. The ones we shipped performed way better because they solved real problems instead of checking a "we have AI" box.

The hardest part of product management isn't figuring out what to build. It's figuring out what not to build. Especially when everyone's telling you to build it.