For the average non-technical person, AI has a steep learning curve. This is fortunate for people with technical experience. Many of the same instincts that come with software development carry into effectively leveraging an AI harness. Context-first, fix it at the source, automate tedium, abstractions, subroutines. If you can think in software, then you can think in AI. This is part of why I still feel pretty good about my career prospects. I'm confident in my ability to do all of these things substantially better than the average person. I don't think I'm some rare genius, but I think I'm competent. Anyone can build something with AI, but building something good, meaning robust, secure, scalable, intuitive, is still hard.
I hear from friends and family about their company trying to shove AI down their throat. Someone high up hears about how much it helps with productivity, or maybe they have their own personal "aha" moment where it does something amazing, saving them a couple hours of work. But this only works if:
- The employees actually want to use it
- Have enough patience to push through the initial learning curve and
- Enough technical wherewithal to notice AI-shaped problems
All of these are considerable bottlenecks. Just handing someone a Copilot license and saying "use it" doesn't work. So what does?
-
Injecting AI into existing workflows where it has been proven consistently effective.
What are the systems that people actually use, and how do they currently use them? Which parts of those systems are costing people the most time? Can those time-consuming steps be reasonably replaced with AI? Example: a document is created from a template and goes through a review workflow before being issued to a customer. The template requires some manual clean-up and rewrites per customer. If you can pattern match on what those updates typically entail, you can build a skill to automate the tedious ones, letting humans focus on the parts that require actual judgment calls.
-
Clearly demonstrate AI value in situations with known frustrations.
This means actually listening to the people doing the work, asking questions, and engaging with their struggles. How does it look when you're doing your job? What gets in the way? Which parts are unnecessarily repetitive?
-
Encourage thinking in AI and create self-perpetuating loops.
This looks a lot like thinking algorithmically. I know from personal experience that as I make more things work with AI, it's easier to see even more cases where it will work well. The loop is rewarding as tedious tasks become easier and easier. Future problems become simpler to solve as the tools layer on top of each other. I feel like this can be distilled into:
-
Create order from chaos.
LLMs are chaotic and noisy. They need to be reeled in. This is a bit abstract, but AI thrives in explicit structure and fails in ambiguity. If you can provide clear guidance in scenarios that were previously flush with inference and intuition, you will find more success when using AI. This is a slow process. What does it look like in practice?
-
When it does something wrong, take the time to explain why.
As you build up a knowledge base of the whys, you will be astounded at what it can infer. Ask yourself: "How did I know this was wrong? What information was the AI missing that caused it to be wrong?" If you fill in those gaps, then next time it won't be wrong. Keep doing this, and you get closer and closer to an effective coworker. This takes patience, but it's worth it.