Taking AI into production – learnings from 100+ successful AI projects

What are the paths and pitfalls in taking AI into production? How do you turn the first success into the next three? Taking an AI solution into production requires a certain kind of skill set. Here’s what we’ve learned from 100+ production-level AI projects.

Understand, verify, and deliver

The steps of successfully applying AI into a business problem are quite simple, and similar to many other uses of technology: First, we need to understand what is the vision we can and want to reach together with the customer. Then, we need to understand the potential benefits of the solution, and that they are high enough to help us succeed even with the usual technical and non-technical obstacles involved in introducing and deploying new technologies. As AI as a technology remains new and unfamiliar to many, successfully onboarding it depends on continuously building transparency and trust on reaching the business goals together with the customer.

With good collaboration and common understanding of the goal in place, it’s time to get stuff done. One obvious part of that is having great people and strong expertise to achieve the goal. Another approach would be to pay attention to whatever is the most present and significant risk for not reaching the goal. These risks can be plenty when dealing with AI, and this situation changes many times throughout the work.

In simple terms, the focus starts from really knowing what needs to be built, and then moves on to verifying whether the technology can actually reach expectations for business impact and with what effort. A big part of this is also looking beyond the project and understanding how the AI solution will be operated and how it will fit the existing operational processes.

Later, focus moves towards traditional project delivery and e.g. hitting schedules, but with AI the key part happens already prior to this. In many specific use cases there’s limited experience and history with applying AI for it, and so you need a solid plan for detecting and derisking the uncertainties of the project. While we’re ambitious about what we want to achieve, it’s never been about ignoring risks but rather “how little time and effort can we use to really know if our plan will work out and what impact can we really deliver?“

Our key principle for success that ties all the previous three steps (understand, verify and deliver) together, is to have a relentless focus on the strongest AI talent. There’s no general AI expertise, but rather many niches of techniques that fit specific use cases, and delivering the best cannot happen unless you know what best means for a specific project. In addition, AI technology keeps progressing very fast and what was state-of-the-art six months ago is often already hopelessly outdated. To keep up with these demands for quality we’ve really focused on finding strong AI PhDs (most of us hold one) and building several means to keep our people skilled with the latest in AI research. Basically, to be great you need to stay great.

https://vimeo.com/466507822/f576231743

Watch Niko's keynote at AI & Business Strategies 2020 on this topic.

Case: digitalizing the pharmacy industry

So how have we taken these principles into use in real projects? Treet by Apodigi is a next-gen AI-driven pharmacy solution, where we got to develop some of the key elements of the new user experience, and design an AI-based recommendation engine around the solution. Our starting point wasn't on what cool tech we could put into the system, but rather figuring what could really improve the care provided by the pharmacies. Once it was clear what really matters, designing the technology and our overall collaboration with the customer became more straightforward.

Next, once we knew what the technology should deliver, we moved onto making a quick validation of how far we can reach with the technology. As mentioned, this has been a key part of our approach and over time we’ve built our own development and deployment infrastructure Silo Operating Software (Silo OS), that really helps us run rapid PoC development and testing in a tight iterative loop.

Later, once we saw the technology working, it was time to get the plan done and deliver. Just like we saw with Treet, deep AI expertise isn't always just data and models. It's knowing how to make the technology operationally strong. Then again, it’s often also about models and stuff. Reaching high levels of AI performance and reliability has been the key to reducing manual work and also to improving the service experience, something we weren’t about to compromise on.

Read more about our work with Apodigi here.

Case: major equipment manufacturer

For a different kind of customer we ran through the same principles, but they worked out in a different way. In this case, we worked with a major Nordic equipment manufacturer. The story is a great example of a ramp up: We started with a small test project to prove our capabilities and ability to deliver. Again, Silo Operating Software was used as the basis for accelerating the development with the use of existing tooling and infrastructure and also some existing customizable components.

Soon after the initial PoC, we agreed to accelerate and broaden the work to new areas. In this case, we were working in mine and construction environments that required special technologies. We were able to bring in new AI techniques that are rarely seen yet in the real-world context, and make them work in order to match the particular requirements of this project.

Advisory isn't choosing this or that buzzword

As we work with our clients, we solve the problems together, each side providing their own expertise. Going big requires deep collaboration between domain knowledge, operational experience, and technology. Bringing new technology to the picture brings new open questions with it as well and derisking the project success needs to be seen as a joint effort, objectively considering what could derail us and how to avoid it.

It is often a surprise to many how fast things evolve in AI also beyond the headline-grabbing spectacles. AI evolution is creating numerous subfields, and each of them is maturing at a different speed in raw performance and operational soundness. We need to know what to demand from the technology and how to achieve it. Keeping ourselves constantly learning and staying close to the scientific AI world that is keeping up their rapid pace of new discoveries. As the scientific AI world keeps up their rapid pace of new discoveries in AI, keeping ourselves close to it by constantly learning is the best strategy to

Ready to get started? Get in touch with Niko to further discuss your how to take your AI plans to the next level, via email niko.vuokko@silo.ai or via LinkedIn.