How to define and scope your opportunity
You’ve decided to kick off a machine learning project. You’ve picked a problem or opportunity that ML is suited to address and lined up resources to work on the project. Everybody is excited to start.
Now what? Do you put a bunch of technologists in a room with the business case and say go?
Not quite. As with other projects, your success will largely depend on your organization’s ability to define and scope your problem and success criteria. This leads us to step 2 of 10 in solving a business problem with ML: Understand, scope and define the opportunity.
Fortunately, you don’t need to recreate the wheel to do this: existing scoping and planning frameworks provide a useful starting point. Here, we’ll draw from McKinsey’s problem definition framework (step 1 in their problem solving first principles):
Finally, here are some dos and don’ts that we’ve learned working with our clients.
Do include thoughts on a feedback loop in your initial implementation plan. Will feedback come from salespeople? If so, how will you design your product to collect feedback as data and use this data to improve your model? Many data science projects stop at the first version of a model; ongoing feedback is where the real power begins.
Do be clear on what success looks like for a proof of concept. It’s tempting to think that any improvement is good improvement, but you need specific success criteria for a PoC so it’s clear when to scale up or stop.
Don’t underestimate the importance of DevOps. Allocate appropriate budget. Big data and complex algorithms require the right infrastructure. A machine learning team without proper DevOps is like a team of analysts trying to use excel on their smartphone instead of a computer.
Do involve security, privacy, and compliance stakeholders from the beginning. Trust is central to any machine learning application, in particular those using consumer data. Understanding regulatory frameworks is only step one. Today, business leaders must think about the ethical implications of machine learning applications, not just privacy and security requirements.
Do be clear on how you will execute on the results of a machine learning model and how this will affect your employees and clients. When a model makes a prediction, who uses it and how? A machine learning algorithm is nothing but a series of numbers unless it is integrated effectively.
Don’t feel that you need 100% automation in your PoC. Remember to push your thinking on what qualifies as a minimum viable product to get an experiment in market quickly. Say you’re using machine learning to guide how center agents will engage with customers. You don’t need a perfect application from day one. You could start by delivering predictions daily via email to a call centre manager who distributes them. Call center agents could input feedback into a google doc spreadsheet in a pre-specified format fed back into the machine learning model daily. This is obviously cumbersome. You’ll want integrations and automations. But this will enable you to get an experiment in market quickly so that you can learn if there is value in scaling up the application.
Now that you have a clearly scoped opportunity and know the pitfalls to avoid, we can move onto steps 3–5: acquiring and preparing the data and doing initial data discovery. Stay tuned for the next post!