Mental Models 4 Life

What is the optimal way to introduce and embed analytics in an organisation? This is a question I have been wrestling with for years. There are three broad options and a lot of in-between: (1) a centralised analytics team that services different business units in the organisation; (2) a distributed structure where analytics professionals are spread across the organisation and located within individual business units; and (3) a hybrid of the two, where you have analytics professionals in business units focussing on day-to-day operational-analytics work and a central enterprise analytics team that focusses on building capabilities for the whole organisation and introducing consistency of practice across business units.

The right structure is of course dependent on the contextual realities of an organisation, including the nature of its business, the number of analytics professionals it has or can hire, the availability of technical managers, etc. In the absence of strong contextual reasons to go completely centralised, I have always had a soft spot for the hybrid structure and tend to recommend that when people sought my advice on such matters.

I’m starting to change my mind, however, and I now gravitate towards recommending a centralised analytics model as a default for most organisations.

The main reason is empirical rather than theoretical. My sample size is not large enough to make any statistically meaningful statement but I have simply seen one too many cases of the distributed and hybrid models failing, with attendant costs for otherwise brilliant analytics people and the broader organisation, to retain much confidence in those two models.

The two things that work against the distributed model are both related to the people dimension of managing an analytics team. The first one is put quite simply by Prof Andrew Ng in his State of Artificial Intelligence talk: Why would a top AI researcher work for, say, a gift-card business unit? Compounding the first problem of finding good analytics professionals is the arguably harder problem of finding a manager within a business unit who knows how to manage an analytics team well.

It saddens me to see, time and again, how non-technical managers are brought in to manage a team of data scientists only to run it to the ground because they have no understanding of what PhDs do and what motivates them, what’s the best way to communicate with them, and how to create an environment to get the most out of them. In most professions, we usually only put people who are technically proficient in a field to take charge of a team of technical people. It baffles me that we continue to see non-technical managers being asked to look after analytics teams in businesses with the  expectation that things will work out. It’s possible, but those would be the exceptions rather than the rules.

So the distributed model is unlikely to work in practice. How about the hybrid model? Unfortunately, they too tend to fail because the moment you have analytics people in business units, the natural tendency of the managers of those business units is to shut off communication and collaboration with the central analytics, believing they can exercise better control and achieve self-sufficiency by solving their own problems with their own people. And then they mismanage their analytics team, which get disbanded, and then they do it again with new analytics hires and so on and so forth ad infinitum.

Until our education system can produce data scientists at the quantity and quality we are producing, say, software engineers, I think the short-term solution for most organisations is to centralise their analytics function and then use the matrix management structure to support the analytics requirements in their business units. There are well-known challenges with the centralised model, but I feel they are more tractable problems than finding good analytics professionals and managers in a regular business unit that has to be managed using a completely different set of rules to that of managing an analytics team.

I will finish by describing a few necessary — but probably not sufficient  — attributes a centralised analytics team needs to be successful.

• Such a team needs to play a key role in the stewardship of an organisation’s data holdings, ideally serving as the path of least resistance for business units to access the data they need to solve analytics problems. One way to achieve this is to have DBAs report to the centralised analytics team, instead of corporate IT as is usual.
• The centralised analytics team must be staffed with a few people who are good at business development. Ideally, the team lead is both technical and business savvy.
• Finally, the team will benefit from the set up and control of a rapid prototyping analytics lab that allows operationalisation of analytics models to be done quickly.

I reserve the right to change my mind if the facts change, but the above is my current best advice on an effective analytics organisational structure for most companies and institutions.