Not Every Bottleneck Should Be Solved With AI

Not Every Bottleneck Should Be Solved With AI

• AI & Machine Learning
• August 24, 2026

Not every bottleneck should be solved with AI. Sometimes AI just scales the bottleneck itself.

A recent Harvard Business Review piece, “Research: The Innovation Problems AI Can’t Solve,” makes an important distinction: before adding AI to a process, you need to understand what is actually slowing that process down. AI is very good at scaling things. But if a process already contains a bad assumption, a biased decision, or the wrong signal, AI can scale that too.

Take idea screening. Choosing which ideas to fund has always involved human judgment. Novel ideas tend to feel riskier and harder to evaluate, so people naturally lean toward safer, more familiar options.

According to the research discussed in the HBR article, AI can amplify that bias in two ways. AI-generated pitches are often more polished and well-structured, and that fluency can be mistaken for quality. A better-presented idea may score higher without actually being a better idea.

There’s a second, less obvious problem. If an AI screening system is trained on a company’s historical accept/reject decisions, it can learn the same biases embedded in those decisions. Instead of removing the bottleneck, you’ve automated it and made it scalable.

The article also describes a field experiment in which evaluators received an LLM recommendation together with a written rationale. They were more likely to accept its rejections without becoming better at distinguishing correct AI judgments from incorrect ones. The AI explanation started replacing human judgment rather than supporting it, especially in borderline cases.

And screening isn’t the only example. The same question applies to bottlenecks across the entire process: what is actually causing the constraint in the first place?

Some bottlenecks are exactly where AI works well. When the problem is too much information or repetitive processing, AI can be extremely effective. Post-launch feedback is a good example: models can process and organize reviews, support tickets, and social media conversations at a scale no human team could handle.

So before asking “Where can we add AI?”, there’s a better question: “What is actually causing the bottleneck?”

If it’s volume, repetition, or too much information to process, AI may be exactly the right tool. If it’s human judgment, incentives, or flawed assumptions, automating the same process may simply scale the problem you already have.

Good AI strategy isn’t about adding AI everywhere. It’s about knowing which bottlenecks AI can actually solve, and which ones need the process itself to change.

Source: Harvard Business Review, “Research: The Innovation Problems AI Can’t Solve,” August 2026.

Image credit: Kersten Winegeart / Unsplash