PM or get PM'd

Product- and Project Management are critical for data science initiatives (in this piece, I use “PM” to refer to both functions combined). This includes activities and tasks such as communicating, evangelizing, documenting, assessing risk, identifying blockers, quantifying the potential impact, and so on. IMHO, companies too often default into a dedicated PM role for these activities rather than enabling a data scientist to do it herself. In some circumstances this division of labor can hinder a data science initiative.?

Algorithms, and data science initiatives in general, are notoriously difficult to design up-front. Models, parameters, insights, performance, etc can’t be prescribed and are instead learned during development. You want to leverage this new information as it is revealed, yet separating the PM activities from development can impair fluidity and unduly increase coordination costs (See Beware the Data Science Pin Factory for more). In addition, having the data scientist do the PM activities herself has other benefits: she gets a broader context, she better understands the customer or end-user, she is able to try more approaches, she feels more accountability, and so on.

Of course, there are conditions that warrant a dedicated PM role and all that a seasoned PM can bring to the table. For example, when there is customer- or company knowledge that is inaccessible to the data scientist, or when data science is just one small part of a larger cross functional effort, or when the data science work is a component of an already existing product/project -- these are all great times to leverage a dedicated PM as well as benefit from the specialization their skillset affords. Done right, the dedicated PM role not only complements the work of the data scientist but also amplifies it. Unfortunately, many companies default into a dedicated PM role even when the conditions do not merit it. For example, many companies leverage ML algorithms that are autonomous to a single function in the company such as operations, forecasting, resource allocation, and so on. Such algorithms may not require cross-functional coordination or external knowledge and more benefits can be derived by having the data scientist do the PM activities herself.?

Yet, even under these conditions, the PM tasks are no less important. If the data scientist fails to communicate effectively or doesn’t properly manage the initiative, the organization can get frustrated. The tempting remedy is to assign a dedicated PM to manage the initiative. This “remedy” leads to the inefficiencies mentioned above. If institutionalized, it can lead to more deleterious consequences. The data scientist role can be relegated to mere order-taker, leading to less novel solutions because the data scientist's unique way of thinking is not leveraged. In turn, this causes the innovative data scientists to leave and the more subservient role attracts more compliant individuals. This is not what you want when you are trying to innovate. Prevention is worth many pounds of cure! A better route is to proactively equip your data scientists with effective PM skills to avoid getting into this situation. PM or get PM'd!

[Footnote: of course, this all assumes you have great people. Variation in talent levels can easily trump good org design and well-crafted roles.]