Navigating the ‘office’ politics of analytics

In today’s world, the term “politics” generally stirs up emotional reactions, debates and strong opinions. Luckily, I won’t be delving into the realm of national political opinion. The politics I’m discussing here fall under the last of Merriam Webster’s five definitions: “competition between competing interest groups or individuals for power and leadership.” Large organizations have multiple functional areas and departments, often with competing interests and their own corporate culture – hence the term “office politics.”

There’s an old saying in analyst circles, usually attributed to legendary engineer and management consultant W. Edwards Deming: “In God we trust; all others must bring data.” As analysts, we like to believe that humans are inherently rational, and that producing the right data, the right formulas and the right strategy will automatically lead to our solutions being adopted and the problems they’re applied to solved.

Anyone who’s worked on a large project knows that isn’t quite true. The field of behavioral economics continues to demonstrate how “rational decions” are often influenced by social and cultural factors. Or as legendary management consultant Peter Drucker – equally popular in analytics circles – put it: “culture eats strategy for breakfast.” You can have the most accurate, relevant data available, pitch-perfect analysis and a foolproof strategy, and if you haven’t accounted for culture it will start chewing up your strategy the moment it’s launched and spit it out before noon.

That is the type of politics I’m writing about, and like it or not, it’s unavoidable.

How your collaborators’ experiences shape their decisions. Another quote beloved by analysts comes from philosopher George Santayana: “Those who cannot remember the past are condemned to repeat it.”

You’re probably familiar with this quote, too. It makes intuitive sense – history does indeed appear to repeat itself, and we learn from experience – but it hides a dirty little secret: We all think we’ve learned from history. And for better or worse, the truth is that many of our colleagues – including other analysts! – don’t view history through the same experiences we do. Which means that even if we believe we’ve learned from history, we often feel as if many of our collaborators haven’t.

In the world of advanced analytics, the shared history of the system being analyzed, predicted and/or optimized is captured in the data that we analyze and the models we build. But the data and the model don’t capture everything. The information that resides in the heads of everyone involved in and impacted by a project is often more powerful than the data in any database. It is based on beliefs, opinions, biases and experiences, which we can collectively refer to as mental models. Unfortunately, mental models are not easy to extract, and big data is no match for strong mental models.

My most vivid experience illustrating this observation happened to me about 20 years ago, while working on a project with colleagues I’ll call Steve and Craig. Steve was a senior-level supply chain leader I had been assigned to support, and Craig was the logistics leader who would ultimately implement our strategy.

When Steve and I first presented our solutions, Craig was not happy. Our model was wrong, he said. Our data was incorrect, even though his team had provided most of it. The word “no” was used a lot, and my list of things to investigate to convince Craig to believe in the model was growing longer by the minute. But one scenario flipped the script around: suddenly Steve the supply chain expert was the one wondering if the model was wrong while Craig believed it made perfect sense. The argument ended with Craig blurting out, “But Steve, the model says so!” After the initial shock wore off, we were all able to move on and laugh about it over drinks later that evening.

As analysts, we’re often so focused on the math behind our models that we don’t have the opportunity to understand our audience’s mental model(s). Craig and Steve’s reactions perfectly demonstrated that someone’s ability to understand and interpret our models are strongly shaped by their mental models versus the mathematics behind the model.

Craigs’s intuition was correct in many cases, and it helped us debug the model initially. But eventually we got to a point where Craig had to change – dare I say, improve – his mental model. It was a foundational lesson for me: Changing beliefs and mental models is often the project within every analytics project.

Strategies for conquering the “problem” of politics in analytics. Algorithms and software are a long way from being able to change minds. In fact, most recent “successes” in social network algorithms and software in the broader world have come from inflaming passions and beliefs that already existed instead of changing minds. I can, however, offer guidance on how to account for political factors when addressing workplace challenges, mitigating their influence and hopefully making your solutions more effective in the long run.

First and foremost, you must be willing to unlearn your own assumptions before you encourage others to learn yours. For instance, if people aren’t using your model the way you intended, get close to the process to learn why. If the reason comes down to the unfortunately all-too-common “they just don’t get it,” the unlearning before learning rule applies here, too.

Yes, to “get it,” users may have to unlearn rules and assumptions reinforced over many years before they can accept a new way of decision making; i.e., give up their old mental model in exchange for a new one aligned with the mathematical model.

But I’ve had to examine my own biases too – for example, by teaching myself that “useful” is better than “optimal.” We often strive to perfect our models, ignoring (or perhaps unaware of) the fact that the more sophisticated the model becomes, the harder it might become for people to understand it – and the less they understand it, the more likely they are to resist it. I’m a greater fan of producing something useful with a positive impact than something that is mathematically more powerful, but fails to make an impact because it’s too great a shift for the current culture. Business impact = model power × adoption, and trading off model power for higher adoption could be globally optimal for business impact.

I strongly believe that the road to simplicity is paved with understanding. You don’t have to over-simplify your model. Just keep trying to find the balance between model sophistication and adoption capacity. If a more complex math model is what’s needed, then that is the right model. Just don’t underestimate the effort needed in education and training to achieve your adoption rate. If you are the “math guru,” don’t fall into the trap of thinking that adoption is not your problem. If you create a model that is so complex that it becomes very difficult to explain, you just became part of the problem. 

The power of metrics. As analytics professionals we are very focused on metrics. Our optimization models come to the table with an objective function. I have had the misfortune – or should I say, “learning opportunity” – where we deployed a model optimized for a higher-level business objective, only to realize that the operations team were relying on process-level decision heuristics that were misaligned with our higher-level objective. For example, in a transportation system where the objective is to minimize empty miles, there are many good local heuristics, e.g., greedy, nearest-neighbor, etc. Using gamification, we were able to prove that the process-level heuristics were 10% to 25% worse than their optimal higher-level counterpart. However, once the model was implemented, the heuristics that were locally optimal were always being used to challenge the model’s globally optimal recommendations. In one case, in order to address the culture shift, we actually redefined the roles and responsibilities to create a break from the old metrics and the politics associated with it.

Develop trust in people – not just automation. Most of what I’ve written can be seen as common sense. Yet if it is common sense, why is trust between analytics professionals and business stakeholders notoriously low? Research firms claim that adoption rates of advanced analytics models are less than 50%. In high-trust environments, adoption of advanced analytics-based solutions should be much higher.

One solution is to create a high-trust environment by following Theodore Roosevelt’s maxim: “Nobody cares how much you know, until they know how much you care.”

As a younger engineer, I used to think caring meant using my skills and knowledge to fix complex business problems. However, I missed a step. In the years since, I have learned to demonstrate a different kind of caring, one that acknowledges a business problem and how it impacts people regardless of whether it falls under my technical jurisdiction or not. This creates the trust needed to build a foundation for adopting a solution that might be different, difficult to understand and even counterintuitive.

My experience illustrates another key to developing trust: recognizing the role humans play in the application of analytics, which, at their best, balance automation with augmentation. Used correctly, automation can lower the cost for model development and data integration, and serve as a useful tool for change management, leading to higher adoption rates. Used incorrectly, it’s a prime illustration of the phrase “garbage in, garbage out,” serving as little more than ammunition for your political opponents. Automation should only be employed for highly reliable processes and well-defined data models. For less reliable situations, use augmentation strategies that allow users to update the data and run scenarios. Allowing users to interact with the model in this way also brings the added benefit of supporting change management and increased adoption.

The “what” is important but the “why” is invaluable. One of FICO’s guiding principles is to make models understandable. We are investing in research areas such as explainable AI because it’s critical that we remember that our models are ultimately serving human decision-making, not the other way around. Decision augmentation models generally do very well in guiding users to their goal – the “what.” But in my experience we spend just as much time or more helping people understand “why.”

This is particularly pronounced in the optimization world, where complex dependencies within a model create greater opportunities to find counter-intuitive solutions. As an analyst, it’s exciting to uncover a counter-intuitive solution and a great opportunity for breakthrough results. But counter-intuitive is, by definition, counter-culture, and being able to explain “why” is an important tool in the tough task of making the counter-intuitive less counter-culture.

What if the last mile isn’t the last mile? Finally, it’s worth considering the “last mile” problem – a term that, in advanced analytics, refers to the challenges of operationalizing sophisticated models. If you think about your analytics processes as a supply chain, then your data is your raw materials, your models are manufacturing or creating better decisions, and operationalizing those models acts as the delivery of your final project to your customers, i.e., the last mile. The right technologies can help you become the Amazon Prime of operationalizing advanced analytics (and FICO would be happy to support that process with industry-leading technology).

But operationalizing is not the same as adopting. Value is only accrued when solutions are used. As advanced analytics professionals, we need to go beyond the last mile. We need to go the extra mile to make sure all the hard work that goes into operationalizing our models results in adoption and value creation. This is where leveraging the strategies listed above come into play.

Going the extra mile to create adoption should not be an afterthought. These strategies should be included in design decisions as early as possible. Designing with the culture of the organization in mind, and developing trust by demonstrating a level of caring that goes beyond technical skills, is a great start. By understanding existing metrics and locally optimal heuristics early, we position ourselves to create designs and project plans that go beyond the “what” and address the “why.” Understanding and using these strategies will arm you well to conquer the invisible forces created by the politics of analytics.