We’re moving on from the initial excitement about generative AI. As a headline I recently read put it, “ChatGPT is not a strategy.” This next year will ask a lot of us as data leaders to define what is a viable AI strategy. Based on my experience working in machine learning and AI when it wasn’t in the headlines every day, here are some best practices for using AI that you might find useful.

  1. Find use cases that deliver business value. If you’re building AI algorithms to figure out what a customer may need next, it pays to start small and iterate. Remember, you’re unlikely to have enough accurate data to boil the ocean right away. To use a healthcare example, it’s a more realistic path to look at lung cancer oncologists in Spain and show what could be done with that dataset using AI. Then you can build it out so the application can scale and deliver value. I call this single, achievable instance a lighthouse. The use case might then expand to all lung cancer oncologists in the EU, then to North America, with more value delivered at every step as you add more lighthouses.   
     
  2. Gather the right talent for the job. When I was building a Next-best Action team at a previous job, I started at the top by hiring an experienced team leader and charged them to build out the organizational capability they felt was needed. The engine that their team built was superior to anything we could buy off the shelf from a vendor. As a data leader, you too should hire great people who know how to build things, then get out of their way and remove barriers to their success. 
     
  3. Take the shortest possible path. Especially in non-tech companies, talent can end up moving data around and claiming that when they get really good at data processing, they’ll arrive at AI. In my opinion you should learn where other companies are making money or saving money using AI, find areas that make sense for your organization, then do a little of what I call R&D: Rip Off and Duplicate. Remember, 100,000 good BI reports will not get you to AI.
     
  4. Bring doubters along on the journey. Part of any AI success is based on convincing the right people in the organization that we need this technology to tell us what our customers need. Take sales as an example. AI in sales is less about promoting a replacement of what people are doing and more about suggesting a complementary upgrade. Stress that your organization already has great customers and great customer experiences, but they’re mostly analog. Explain how AI will create a digital enabled customer experience for them. Offer to show them some examples. Bring in consultants who have solved similar problems before so they can demonstrate the art of possible. 
     
  5. Transition mature AI models to MLOps. You may have AI algorithms out there, but how do you know that no one touches the data that feeds them, or that your algorithms aren’t drifting; that is, that the underlying data distribution your algorithms were trained on no longer matches the real-world data they encounter in production? Once a model is up and running successfully, it’s critical to transition the work off of the data scientist’s desk and hand it over to an MLOps team. They are much better trained at overseeing the entire process, optimizing an entire system or platform of algorithms, and ensuring that you have lineage and traceability all the way back to source systems. They also can help to ensure that algorithm governance becomes part of overall data governance. 
     
  6. Bubble-wrap your data. This is how I talk about observability. When you're not at a tech-native company and you have AI in place, you may not be able to clean up every piece of data. This may require wrapping observability around the relevant datasets. This way if someone does touch the data, maybe indirectly through an ERP system that has an algorithm attached to it, you know right away and can take corrective action.
     
  7. Phone a friend. At large organizations where people have worked for a long time, it can be tough to find an infusion of new ideas. It can be more effective to reach out to your network and ask, Who solves X problem really well? Request meetings with two or three companies who have solved the problem or something very similar but are not in your competitive space. These people might be in large tech companies or in startups. Example: I learned about observability through a conversation with a VC who was asking me about a leading company in this space. He described what this company did, and I thought, “Hey, I need that!”

Stay current, not trendy
If you’re looking to get an overall sense of what’s happening in the AI market and who is at the forefront of AI best practices, I might start with the Quantum Black/McKinsey annual survey, which looks at companies leading in AI and what they’re doing differently. Then if your CEO comes to you with a requirement to use LLMs, you can point out that this may not be where your competitors are making money with AI. In my own case, a CEO asked me if we could use AI to improve working capital in supply chain; in other words, use it to do better inventory management, reduce excess stock, optimize logistics, streamline transportation, and so on. I said we absolutely could, and our team helped to generate $100 million of working capital savings in four months. 

Moral of the story: sometimes the AI you’re being asked to deploy isn’t the AI you need to win.