Services-Systems Co-Design: The Lesson From Apple To Make AI Implementations Work

Failure comes from different workflows, that should be reimagined

Happy Sunday and welcome to Investing in AI. If you are new to the newsletter, be sure to also check out our AI Innovator’s Podcast, and we are always looking for new guests so reach out if you have something to chat about.

Today I want to talk about the challenge of AI implementations and how to fix them. This insight came about as my investment group was looking at a “services as software” play, where we buy either a services company and add AI technology, or buy a software company and merge it with a services company that the software supports. Someone pointed out how different these companies are to manage. Services companies and software companies run on different metrics.

As an engineer at heart, this seems to me a lot like the hardware vs software design paradigms. In most systems, they are designed independently. But Apple is famous for a model called “hardware software co-design”. In this model, both the hardware and the software are optimized for each other. This is a more challenging design process but it leads to a better optimized system overall.

In AI, the last few years have been focused on performance at all costs - pushing the limits of what AI can do. Now I believe we are starting to enter the era of TCO and ROI. We know what modern AI is capable of and we want to figure out how to run it efficiently and apply it effectively.

At the level of technology, I’m starting to see people think about AI systems - hardware/software/model/storage/network all together. This allows them to optimize for cost or performance or power consumption or some combination. But what I really want to focus on is making AI work at the workflow layer.

Services take advantage of human capabilities. Humans are flexible with generalized intelligence, and like interacting with other humans. They can adapt easily to unexpected surprises or changes in a workflow. Software is the opposite - optimized for efficiency and speed of one or more specific tasks.

Software for services companies, like accounting, legal, recruiting, consulting, etc, is usually built to accommodate their existing workflows. But as it starts to eat into part of the human work, this may no longer be the right way to do it. There now needs to be a recursive process of modifying the services workflow to take advantage of new AI software capabilities.

To take a practical example, lets use a firm that creates digital advertising for small to midsized companies. There are design workflows, approval workflows, and execution workflows. But now generative AI can do the design and execution workflows in minutes, and can do so at the scale of mass personalization. What does that mean for the approval workflows when there are thousands of ad variations created in a few minutes that are ready to go out immediately?

If you keep the existing approval process, it is so slow that you don’t really get the benefits of the AI powered workflows. And no human can validate thousands of digital ad options, so you’ve just created more work. The path out is to re-design the approval workflow using both software and services. Building a model that will catch the ads that are most likely to have problems, and having a human sample those for approval or disapproval can give you a sense of whether the AI content going out is worthwhile.

Thinking about services and software together as you design your workflows can lead to a much better overall optimized outcome. The insight here for investors is that, this will influence where companies buy software and where they build software (as software gets easier to build). The variability in the market across a specific workflow will be a key driver of that decision.

The right path is very situationally dependent. But I hope this gives you something to think about that can help with your AI implementations.

Thanks for reading.