6 blind spots in every AI rollout

Here’s something we see a lot. Leadership announces an AI initiative. They ask every team to submit their best use cases. The ideas start flooding in, and there’s a ton of excitement.

The initial wins are exciting. Setting up Claude and GSuite connectors. All of a sudden AI can help you draft emails, organize your day with morning briefings, and summarize Slack threads. But that’s about it.

Everyone saves a few minutes here and there using AI, but no workflows in production have been fundamentally transformed. Maybe it’s because the models are good at coding, but not good at “what we do”. Or at least that’s what we tell ourselves.

But the truth is simpler than that. “Send us your best AI ideas” was a broken ask in the first place, and it creates blind spots that make it hard to see the best opportunities for applying AI.

This article is about the blind spots we see, and how we reframe the questions we ask to help companies find the opportunities hiding in plain sight.

The six blind spots

1. Pain is a bad map. People are quickest to nominate the work they hate doing. But painful work often requires expert judgement. On the flip side, boring and repetitive work that might rank low on the pain index often makes the best candidate for AI automation.

An example we ran into is a KPI deck that gets generated every month. On the surface it sounds like a great use case. 90% of the deck could be written blindfolded. But when we dug deeper we realized that all of that happens after updating a crucial financial model that is owned by an expert team member. And the assumptions going into that model are based on the company’s goals this quarter, the market signals, and decades of their own experience in this industry.

2. Invisible work. As touched on above, the best use cases are the ones that experts stopped noticing years ago. The work feels easy when it’s baked into muscle memory, but that doesn’t mean it’s fast. And these hours add up quickly.

A good example is a routine data entry task like submitting timesheets. It takes an hour a week, but you stop noticing it because it’s easy. So instead of asking “what’s the most painful part of your job?”, we simply ask folks to walk us through a typical week.

3. The idea contest trap. We frequently see AI positioned as a time-boxed experiment. But hackathons reward flashy solutions, which means you will never discover the boring workflows that don’t demo well. Instead you get vibe-coded mega apps that will never hit production.

On the flip side, we’ve found that the best way to surface the best ideas is to let people experiment. That means investing in AI training and education. Hosting consistent workshops where folks are encouraged to bring their “dumb questions” about AI and get unblocked. At the end of the day, it’s the same principle that made agile development work. Push decision making and agency to those who sit closest to the problems.

4. Stuck in time. In many conversations, we’ve found that people’s perception of what AI is capable of is grounded in the last time they tried it. And that could be anywhere from last week to 2024. There’s a narrative that AI hasn’t changed much since the first few years, but nothing could be further from the truth. Sure, we’ve saturated a lot of benchmarks, but model capability and the cost of intelligence continue to progress at a rapid rate.

A good example is security. Many people I talk to still believe the prompt injections we uncovered in 2023 will work, and I always have to ask “when’s the last time you tried this?” Even in normal software, documentation is constantly out of date. In the AI industry this is even more true. The only way to really know is to try.

5. The capacity blind spot. It’s easy to nominate work that already exists, but what people often miss is the work that doesn’t. Things that if you “had more time” you would do “the proper way”. Reframing what AI can do for you around this can be very illuminating.

A general example is “checking your work”. Many industries have SOPs that are much closer to an ideal of how work should be done than to how it actually gets done. And that’s simply because no one has the time. With AI, you can have something constantly checking work orders and purchase orders to make sure your project is still within budget.

6. The handoff tax. I love this one because it’s a real easy one to miss. I know because I’ve missed it a lot. It’s easy to get excited about automating an end-to-end workflow and forget to place yourself in the shoes of someone receiving the final work output. But if it takes you just as long to reverse engineer the report and verify the claims, you might need to hit the drawing board again.

A classic example is using AI to generate an analysis or report. Sure, the AI is great at pulling together multiple sources, but the second it spits out a 10 page report it becomes your problem. A good litmus test is that real work done with AI should be something you can share with any teammate. And if you’re sharing it with someone, you better bet there will be questions. So you need to engineer the workflow to make sure that your understanding of the final work output is defensible.

Reframing the questions

The good news is that despite all this, we still see a ton of opportunities for AI hidden in plain sight. It just takes a bit of reframing to see them.

The mental model I like to use is a very smart and eager intern you just hired. They know little about the business, but with enough coaching and guidance could definitely figure it out.

• “What painful work can AI take away from me?” becomes “What kind of tasks would I give an intern if they started tomorrow?”
• “What can I have AI do that will free up more of my time?” becomes “What am I not doing today that if I had a team of interns I would definitely do?”
• “What can AI automate for me end to end?” becomes “What are things that I can easily delegate to an intern and verify quickly?”

Expanding on that analogy one step further, the people I’ve seen get the most out of AI often think about themselves as a manager: sharing context, unblocking teammates, managing timelines.

Conclusion

You can’t possibly discover the right use cases if you’re asking the wrong questions. Reframe where you start, and you’ll end up in a better place.