Customer Expectations of AI are OUTRAGEOUS

One of my favorite memories as a kid was getting a new video game from the mall. Riding home in the back seat, I would rip open the shrinkwrap and dive into the instruction manual. I couldn’t wait to learn all about the missions, the bad guys, and the controls.

Fast forward a couple decades and the video game instruction manual is a bygone relic. My kids wouldn’t even understand the concept of a game manual, much less be willing to actually read something before playing. It’s a foregone conclusion that every video game today comes with a built-in tutorial for the first-time user. The best ones have progressive disclosure, teaching you the basic moves before exposing the more complex ones.

‍

Obviously, this is a better system. The options for video games today are approaching infinity and asking a kid today to invest any time in learning the mechanics of a game before deciding if the game is up their alley is akin to learning how to be an airplane pilot in order to decide if you want to fly United or Delta on your next vacation. 

But, also: kids are so programmed. The days of spending 5 hours on a Saturday watching G.I. Joe are long behind us. There are ballet lessons, robotics clubs, birding classes, robotic ballet bird building societies, and on and on—all of which have to be finished before noon because after noon there is a birthday party to go to and also the travel soccer team has a match seven and half hours away. 

When they finally sit down to play that’s ALL they want to do. No reading skibidi manuals. Just start cooking, bruh! (I think I did that right?)

The same is true with business software. I’ve been building software for a very long time and if I can’t figure out how to use your software and get value from it very quickly, I’m gonna bounce. 

Within minutes, a prospective customer / user needs to answer the following questions:

1. Is this product for me / someone like me?
2. Will this solve my problem?
3. And how?
4. And how QUICKLY!

Which is why the best modern software products invest so heavily in onboarding and the new user experience. We constantly want to minimize the “time to magic”. 

And yet, 25 years of building software and 40(!) years of playing video games did not prepare me for the expectations, assumptions, and impatience of customers testing out AI agents 🤯.

Customer expectations of AI Agents are NEXT LEVEL

As we started onboarding early customers to Teammates, most didn’t want to read any docs or get any training. They wanted to dive right in. Which we planned for! Just like many AI products, our teammates can chat and get assignments with natural language. 

But what happened next was surprising. 

The very first external customer we onboarded into the product created their first teammate (a virtual marketing manager) and typed into the chat: “Do your job.” As you might expect, this did not go well. At all. 

The next customer tried something similar: “Your job is to read the latest news about renewable energy and post the interesting stories on our X account. My username is @example and password is *******.” followed by “Now do today’s work.” 

Once again, this did not go well.

What I found most fascinating was that these expectations were both real and common. In a post-LLM world, customers have an outsized assumption that software is now pure magic and that anything will work.

In a post-LLM world, customers have an outsized assumption that software is now pure magic and that anything will work.

It’s us who need to adapt

Just like video game designers had to adapt to a world where no one reads the manual, building tutorials into the start of every game, we as software designers need to adapt as well.

Tutorials define HOW to play but they also teach you the overarching boundaries of the game. They teach you what happens when you try to make Link jump off a cliff, but they also teach you that Link can’t eat rocks. 

And what’s truly wild? Is that today, Link can’t (and shouldn’t) eat rocks, but tomorrow that may be different. Our AI agents evolve at such a rapid state that our changelog looks like a stock ticker. All of which leads to a much more complex and MUCH more exciting state than any I can remember in all my time building software.

Unless we want to throttle the development of our AI Agents (we don’t) or drastically lower users’ expectations (we don’t think we should have to), we need to teach people how to fly a plane that is building itself midair. It’s a new problem for so many of us that grew up bridging gaps and solving problems. But bridging that gap may be the coolest thing since the Contra Cheat Code. And it’s only getting cooler.