Generative AI’s New Clothes
How to think about AI adoption
No boots!
Every time he comes up, he's got no knife, he's got no jacket, he's got no pants, he's got no boots. All he's got is that stupid gun he carries around like John Wayne. That ain’t gonna help you.
That was Mike (played by Robert De Niro) talking to Stan (played by the late John Cazale) in one of my favorite films, ‘The Deer Hunter.’ Mike and his friends work at a steel plant in Pennsylvania.
I’ve been to a few steel plants over the last six years. The scale, the physicality, and the intensity (heat, light, sound) are something. It’s an engineering and technology marvel to take a bunch of scrap and turn it into a beautiful new sheet of steel with desired properties. Easy for an outsider to say though - the operator of a 3200 oF electric arc furnace might not have such romanticism associated with steel.
Every industry and every company within that industry has come to be through invention and discovery. Lot of which I took for granted. Partially because of living in the Silicon Valley bubble. Drowning in a sea of bits, I forgot about the atoms. Trips to scrap yards, factories, warehouses, and cubicle-lined offices cured that.
“All he's got is that stupid gun he carries around like John Wayne.” If you’re a company building Generative AI products, don’t be that kind of John Wayne. Don’t just carry your AI around with no knife, no jacket, no pants, no boots. That ain’t gonna help you.
Generative AI has seen unprecedented interest and usage by consumers. Novelty, controversy, and ease of use (natural language interface) drove a lot of that. I’m a sucker for Generative AI but keep reminding myself that changing user behavior is hard even though we might think we have invented the wheel. The irony is that you invent the wheel, and people expect it to fly. C’mon! What?
The AI adoption challenge
Apart from the complexity of building AI products, to begin with, companies creating AI software for enterprises face three core, interrelated challenges: a) data - fragmentation, quality, sufficiency and context, b) systems fragmentation, and c) workflow fragmentation.
All this fragmentation will be there even within the same industry and companies that are very similar – literally the closest competitors in that industry! Large Language Models right now may not face all these issues as those are trained on large, available datasets, and therefore work well enough out of the box, and are still being used for narrow applications that don’t require linkages across business systems and workflows. Foundation models are more general purpose compared to use-case specific models, and that removes a lot of the initial friction (although at an extremely high training cost). However, expanding to multi-modal foundation models, fine-tuning those on company-specific data, and integrating them into more complex workflows will expose more challenges.
To deal with the fragmentation, if the AI company customizes for each enterprise customer, that is too expensive for them, too demanding for the customer, and too non-standard for the end-user to benefit from a large external community of users. Also, very difficult to profitably scale a business model that relies on custom AI.
Think about the end-users of the AI product you’re building. These users could be graphic designers, mill operators, financial modelers, purchasing agents, or benefits administrators. They already have a lot of tools. They have invested in learning those tools and are comfortable using those. Their company has invested in integrating the tools. Neither the users nor the company has a strong incentive to try new tools. They are concerned about the heavy lifting to implement, integrate, learn, and maintain a new tool. “AI” sounds complex and overwhelming and black-boxy to them.
Traditional software is deterministic. AI is probabilistic. It requires users to embrace uncertainty and that doesn't come naturally. Because of the persona (or aura?) that AI has developed, people expect even lower uncertainty from AI. For example, self-driving cars likely drive better and cause fewer accidents than humans, but people expect that those cars should cause zero accidents.
So, you are going in from a position of weakness, even though you might think you are going in from a position of strength because you have the best AI, the best user experience, and the best value story in the world. You are the John Wayne with a gun.
What do do?
Think about the sponsor (the one championing and paying) and the end user (the one who must use this thing daily).
‘The Sponsor’ needs: a price within their approval limit, a burning platform, low human resource investment from their side, customer references, easy to get started, easy to see initial proof and fast time to value. Is that what heaven looks like?
‘The End User’ needs: user experience of course – learn from consumer apps because that is what they are used to and what they treat as the high bar. It should just “work” – can’t make excuses about data, or that AI is hard, or that it takes time to learn, or that the context/domain has changed, etc. Sometimes it should align with their expected outputs (to build trust) and other times it should surprise them with something they hadn’t thought about (to create value). The end-user doesn’t think in terms of dollars created or dollars saved. They think in terms of time. How can you give them a better result in lesser time? The “better result” could be implicit or explicit. When using an image editing tool, you should feel that the AI is accelerating, amplifying, or expanding your creativity. And maybe what you collaboratively create is getting you more engagement. But that higher engagement may not be due to your new shiny tool. When using an AI agent that completes a series of tasks for you to achieve a result, and the quality of that result is directly measurable and attributable to the AI agent, then we can think of giving achievement medals to the AI. Even though the medals are given to the AI (non-anthropomorphic), the human user benefits by gaining trust.
In situations where users can’t easily judge AI outputs or actions, it could be useful to have an optional layer of trust-building features. Users could opt-in to view how the AI performed on relevant, standard benchmarks before getting to them, how the AI has been doing since they have been using it, the AI’s confidence in the tasks at hand, and in some cases even the steps it took. Again, all this can create information overload and may not even be needed in all applications. Also, no matter how many metrics and proof points we build in, the users must then trust those as well! Another possibility is that these become non-issues as AI performance becomes highly accurate, reliable, and consistent.
I’m on the fence about trying to use economic value as an adoption lever for end-users. Honestly, I’ve been a strong proponent of that for many years. That’s because AI is built on data and directly or indirectly makes decisions, and takes actions. I felt that all the ingredients to calculate value were there, so why not do it and put it in the user’s face – here you go, now pay up and use this thing because it gives you dollars. Why would you say no to getting dollars? Over time, I’ve come closer to the center on this topic. Firstly, it is hard to figure out the real value. You need all sorts of data that you didn’t already collect to train the AI. You also need to figure out what would have happened without the AI! Secondly, attribution is tricky unless it is a narrow use of AI. There’s a lot of stuff going on in consumers’ lives, in businesses, and in the world. We don’t know how much of that helped (or didn’t) achieve a good outcome. Lastly, all this complexity of getting to the value number makes it even more difficult for the end-user to digest what you came up with. It’s not their math, and they know you could be biased.
Broadly, the AI product adoption should be built into the user experience i.e., the user should be able to start small and take baby steps nudged by the experience, as opposed to having to go off and take training or read documents. Baby steps even when it comes to results and value. Start with small wins that excite the users, wins that they can easily believe in without a lot of mumbo jumbo. You also need evangelists within the user group – believers who are willing to take a small chance, invest a bit more, get excited by new stuff, and change their own behavior. But you can’t disappoint them – they need to be wowed. You have one shot.
This article is mostly about user adoption focused on the AI product. The users’ relationship with the AI company’s brand and people also matters. Zooming out, their experience across all interactions with the company matters.
Paul Graham tweeted recently: “AI is probably as big a deal as microprocessors or the web. But for the first few years only nerdy enthusiasts cared about those, whereas everyone is using AI. I think the difference is that ordinary people don't have to wait for the nerdy enthusiasts to build stuff for them on top of AI. It's useful right out of the box. But the nerdy enthusiasts will build stuff for them on top of AI as well. Which will also presumably have higher adoption rates, since so many end users will already be believers.”
Paul is likely referring to the recent wave of Generative AI. Yes, it has been relatively easier to use and fun to use. Also, it’s more interactive than other forms of AI. All of that has led to faster initial awareness and adoption.
What is art?
I met an artist at a recent meetup organized by Adobe Firefly. He said that even steelmaking is art. Got me thinking – how can I use Midjourney in a steel mill?! We’ve come a long way from the Industrial Revolution and are in an AI Revolution. Remember that when you drive a car, you never think about how the steel was made. Similarly, when your customers drive an AI software, all they should be doing is happily gliding toward their goal.