Build models with transparency and reliability. Machine learning (ML) can be a black box and unfamiliar to many people. Demystifying the process and including leadership and end users every step of the way is the best way to get buy-in from clients. This creates advocates for the product through deepened understanding — and people like things they can trust.
Make real impact. Having a data science (DS) team churn out 30 cutting-edge models is useless if the models are collecting dust in a shelf. A few simple models with large impact deployed in production creates tangible value, however large, for the client. Real impact > potential impact.
Just because you can, doesn’t mean you should. Not all problems are nails to be solved by the Artificial Intelligence (AI) hammer. In identifying optimization opportunities and process efficiencies, it’s tempting to insert AI wherever possible. Problems should be solved with the simplest route possible, starting with solutions like providing additional training to staff, updating steps in the process, automating a manual process through code or robotic process automation, and finally arriving at more complex solutions like AI. We want to simplify systems, not make them more complex.
ML client engagements offer a two-way street for learning. Consultants may bring the data science skills, and the client will bring domain or industry knowledge specific to their data, systems, and processes. Neither is an expert in all the inputs for a model, and success relies on remaining respectful and open to fill gaps in each other’s expertise.
Data is the whole pie. If model = algorithm + data, prioritizing clean and thought-through datasets is 90% of the outcome, and the most difficult to secure. Bad data in ⇒ bad data out.