Table of Contents

    For nine seasons of (now classic) television, The Office’s Dunder Mifflin Paper Company somehow managed to skate by in an industry that has been experiencing a long, slow decline as the world becomes more digital. In Scranton, where the show is based, regional manager Michael Scott somehow manages (despite his best efforts) to build a branch that is, for all intents and purposes, successful. However, success is a relative term, as most of us know, and while Scranton might have been plugging along, the company itself went through near-bankruptcy, failed initiatives, an acquisition, and more troubles just to remain afloat. 

    What happens if we look at Dunder Mifflin’s sales numbers compared to real-world sales? We start seeing an interesting picture emerge, one that might explain why the company struggled so much, despite having branches like Scranton, which are surprisingly robust. One of the bigger issues (and we may be extrapolating here), is the company’s insistence on keeping things “old school.” Also, due to  some notably large failed tech initiatives (can you say “Dunder Mifflin Infinity”?), Scranton’s sales team remains strongly anti-technology:


    The company insists on face-to-face sales, which is fine, but it also refuses to optimize its sales processes. Michael Scott may believe that a human touch is best, but what happens when you need to spread out, or when your budget is limited, or when you simply need to determine better promotions? Going with your gut isn’t always a great idea (even if they sometimes pan out):



    So, with that in mind, how can we bring Dunder Mifflin into the 21st century? Let’s see how using data science could help us optimize the company’s sales pipeline and build a model to better score their leads to boost their sales. 

    The case for data science at Dunder Mifflin 

    Hard as it is to believe, Michael Scott actually built the winningest branch in all of Dunder Mifflin. As we said above, though, that may not be the brag it sounds like. The company itself struggles to stay afloat often, and it eventually gets absorbed by a printer company to avoid going under. To get a slightly better idea of how Dunder Mifflin ranks, let’s take a quick snapshot of the paper market. 

    Global paper sales have steadily expanded since 1992, but that is for the overall industry. If we focus exclusively on the graphic paper market (which includes Dunder Mifflin), the sector has actually seen compound annual growth of -1.5% between 2010 and 2018. According to Mckinsey, this is more of a permanent trend than a fixable problem. The rise of paperless operations and the turn toward eco-friendly business policies means that paper (specifically, for printing and graphic purposes) is less in demand. 

    Even so, Dunder Mifflin (by its own admission) is not positioned to be competitive. In fact, by Michael Scott’s own admission, the company’s prices place them near the bottom in the local and national markets. Let’s see if we can find a better price point and determine how we could boost sales. 

    Fixing Dunder Mifflin’s pricing strategy 

    We can (and will) focus on a few different areas where a data science platform might make an impact, let’s start at the place where Dunder Mifflin seems to have the most trouble — pricing. When it comes to flagging sales, finding a better price point can provide a quick win. It’s safe to assume (judging from what we can glean from the company’s relationship with technology) that Dunder Mifflin is technology-adverse at best.

    So, how do we start making heads and tails of the company’s pricing? First, we need to collect their historic price data and gather information not just from our beloved Scranton branch but also from the company’s nationwide network. We know the company’s inventory has been underpriced for some time, so relying exclusively on that data will only give us inaccurate results. So let’s look elsewhere to help bolster the model with a more robust data pipeline. We can start with some landscape data sources: 

    • Revenue numbers by industry that can point toward buying potential. This can help Dunder Mifflin price its products more dynamically based on industry and even clients. We can use industry codes to start matching potential customers to their industry and use that information to also see where the biggest sales opportunities could be. 
    • Procurement data that includes production levels, resource prices, operational costs, and supply figures. These numbers can help us understand the cost of acquisition and can lead to a more accurate model that produces better margins and avoids the underpricing issue we’re trying to avoid. 
    • Web search data and social media interactions. This is where it gets a little esoteric, but can actually help boost pricing predictions and even targeting models. Web searches reveal both the demand for your products and the average prices. Social media shows how people react to specific companies, their policies, and their business interactions. By including these data points, we can help Dunder Mifflin adapt, target businesses, and price their products competitively without being too aggressive. 

    Based on this data, we can start building a few different models that can lead us to a more accurate price point. 

    Let’s get to work building Dunder Mifflin’s ML models

    The first is a demand forecasting model that predicts which products will be in greater demand in the near future, letting Dunder Mifflin stock those products that will move quickly. This allows the company to price its goods competitively, instead of selling them off at a steep discount to move inventory that no one wants. We can make it dynamic so that every branch will understand their local market better and handle inventory ordering from the central warehouse on a more case-by-case basis. This will allow them to both lower their inventory costs, avoid unwanted products, and provide prices that match the real demand. 

    The next model we can set up is a targeting model. Understanding demand is one thing, but Dunder Mifflin Scranton, which already struggles to find customers, needs to target its leads better to ensure it can offer its products without underpricing them. Using their enriched data, we can find the company’s potential customers who are most likely to need paper, want a local firm, and pay at competitive, non-discounted prices. 

    Finally, and because we know Michael Scott likely can’t help himself, we can build the company a model for better promotion scheduling and planning. This model would help the company understand when it’s the best time to launch a paper sale, or another “golden ticket” situation. More importantly, it can show which products are likely to be the best suited for promotional prices and to drive the most interest in the company’s goods. 

    Could data science save Dunder Mifflin? 

    It’s tough to say that simply introducing data science could turn around Dunder Mifflin. The running joke is that the company isn’t great, and it has barely any name brand recognition. However, using machine learning models could help the company become more competitive and efficient. Instead of focusing solely on cold-calling to drum up sales, the company could, for instance, identify higher value and more interest leads. More importantly, it could finally stop losing money on all its sales. 

    The best part is that it wouldn’t even have to go full “Dunder Mifflin Infinity” and invest millions into a strategy doomed to fail from the beginning. The company could invest in a platform that could handle the data discovery, enrichment, and even modeling for them, and be on their way to avoid the dustbin of corporate history.