How Birchbox Stays Competitive using Machine Learning

Machine learning helps Birchbox stand out from the competition by sending the right products that customers will love and continue to order in an online and offline world.

What is Birchbox?

Birchbox is an omni-channel beauty retailer that offers an online subscription service, an e-commerce site, and brick and mortar retail experiences. Founded in 2010 by Katia Beauchamp and Hayley Barna, Birchbox started off with a simple online subscription business that would send beauty-related samples to their subscribers. Their early model had lots of competitors such as Ipsy and Glossybox. Because it was hard to differentiate on just providing samples, they also decided to sell full size products of their products online to generate increased repeat revenue [1].

How Birchbox Leverages Machine Learning

Machine learning is important to the success of Birchbox because it enables it to differentiate from the competition by sending the right products that customers will love and continue to order full sizes of [1]. If Birchbox can do this better than their other competitors, they will have an advantage in terms of customer retention and revenues. They do this by using data and machine learning to optimally allocate samples to customers and personalize the experience.

Birchbox needs to allocate a limited number of every changing samples to maximize customer “happiness” every month. They triangulate multiple metrics including star ratings, follow-up purchase of full-size products, referrals, cancellations, and upgrades. They add more data as they gather it and tweak the algorithm to try to improve the quality of their algorithms. [2]

Example of Sample Optimization Problem using Machine Learning at Birchbox

Source: Birchbox Unboxed Blog [2]

Birchbox uses machine learning in recommendation and personalization algorithms to predict when customers are likely to purchase and send them an email. They also use machine learning to personalize the online site experience for each customer. To tackle these problems, they embed their data scientists directly with their product development teams, to work hand in hand in improving their algorithms. As they solve these problems, they need to balance number of data points processed and the speed at which they can run the algorithms. [3]

Birchbox recognizes that machine learning isn’t always accurate. To deal with these inherent limitations of predictive machine learning, they let the user choose one of the samples, so no shipment is a total miss. They also ask customers for profile information on an ongoing basis to improve their data points. [3]

Machine Learning in Retail

Not only is Birchbox tackling machine learning challenges online, they are also leveraging machine learning in their retail strategy. In their new stores, customers will leverage iPads to help them find personalized products. Inputs will include skin tone, hair color, hair style, and more. Birchbox will continue to expand out these retails stores to provide additional places for their subscription customers to buy their full price product. [4] A challenge they will face is integrating all these online and online touch-points for machine learning purpose. [3].

Example of Birchbox in Walgreens 

Source: WallStreet Journal [7]

Looking further ahead, Birchbox is partnering with retailers such as Walgreens to create even more reach for their customers. An 11-store pilot will create Birchbox branded experiences within select Walgreens stores [5]. These branded experiences will allow customers to sign up within store for the subscription service and even “build your own box”. This will give Birchbox even more data, particularly from an in-store experience, as input in to their algorithms. [6] This creates new challenges as they have to connect the data and use machine learning to predict success not only in their online and owned brick and mortar business, but also in their growing partner retail channels.

Future Steps for Birchbox

One step the organization can take to further address improving their sampling offerings using machine learnings is to expand geographically into new markets where customers can provide more information. These could include East Asian countries or South America. That way Birchbox can have more data and more types of users to optimize it’s offering.

Birchbox also needs to take steps in the medium term to attract and retain talent in data science and machine learning. It is notoriously difficult to hire and retain talent in this field because the supply of these capabilities is low while demand is high. Talented machine learning professionals will work for high pay at large tech companies. Birchbox needs to think about how it can be attractive to these individuals either in compensation or by providing exiting machine learning problems.

Open Questions Looking Forward

How does a company with a very specific machine learning problem and limited budget continue to hire and retain top machine learning professionals?

How does Birchbox deal with solving machine learning problems such as sampling success as measured by full-size product conversion when they may not directly own the data as they scale more retail partners such as Walgreens?

(Word Count: 787)



[1] Chhabra, E. (2018). Key To Success: Beauty Box Company Birchbox Says It’s Not Just About The Box. [online] Forbes. Available at: [Accessed 12 Nov. 2018].

[2] Birchbox Unboxed. (2018). The Birchbox Problem – Birchbox Unboxed. [online] Available at: [Accessed 12 Nov. 2018].

[3] Gutierrez, D. (2018). Data Science & Analytics at Birchbox – insideBIGDATA. [online] insideBIGDATA. Available at: [Accessed 12 Nov. 2018].

[4] Heine, C. and Heine, C. (2018). Birchbox Aims to Bring ‘Artificial Intelligence’ to Offline Retail. [online] Available at: [Accessed 12 Nov. 2018].

[5] Dennis, S. (2018). Strange Bedfellows? Legacy Retailer And Disruptive Brand Partnerships Are On The Rise. [online] Forbes. Available at: [Accessed 12 Nov. 2018].

[6] Digital Commerce 360. (2018). Walgreens takes a stake in Birchbox and will begin selling its products online. [online] Available at: [Accessed 12 Nov. 2018].

[7] Kang, J. (2018, October 04). Walgreens Takes Stake in Birchbox. Retrieved from


Roche: Improving Drug Discovery and Development with Machine Learning


Autonomous Stock Replenishment at Online Retailer OTTO

Student comments on How Birchbox Stays Competitive using Machine Learning

  1. That’s quite the equation you posted! 🙂

    As a consumer of subscription based services, Birchbox included, I didn’t realize how difficult it was to stand out in this industry given fierce competition. I particularly enjoyed the insight you provided on converting customers to “full size” products. I definitely see how this serves as a competitive edge for Birchbox; however, how I struggle to see how they can compete with more established beauty businesses with a loyal following (e.g., Sephora, Blue Mercury). If someone finds a product they love through Birchbox, what prevents them from ordering it through one of the traditional retailers? What if a traditional retailer launches their own subscription based service?

    I bring this up since it is widely speculated that subscription based services are in a venture backed bubble [1]. One example of this is Blue Apron. They IPO’d in June 2017 with an original trading price of $9.34 per share; however, over since then, they’ve been on a steady decline to $1.22 (as of after-hours trading on Nov 13th).

    This perceived skepticism in the market brings me to an answer to your question about acquiring talent: they don’t need experts in ML, they need experts in branding. Given the saturation and steep declines in the market, Birchbox should focus on creating a brand millennials love, not optimizing an algorithm.

    What point is the algorithm if no one’s there to buy anything?


  2. I find it interesting that Birch Box started as an online subscription company and now is moving into retail! It will be cool to see if Birch Box finds different preferences in the store consumer vs. online consumer. I think one problem that they will run into moving to retail is the quality of data on consumer purchases and decision factors. Currently as an online store, they can match buyers by profiles. In Walgreens, they wont be able to match a repeat buy that purchased in Texas with one credit card and in Boston with another (I don’t think….). My guess is that retail will make it harder to draw conclusions on the success of their product selections and make it hard to scale their machine learning.

Leave a comment