How Enterprises Are Using Machine Learning To Deliver Incredible Results Today
The buzz around machine learning has built up to a loud, steady rumble over the past couple of years. You’ve undoubtedly heard about its successes in board games, cancer detection, or self-driving cars. At the same time, the enterprise narrative has increasingly been “if you don’t begin implementing ML in your organization, you will be obsolete in 5 years”.
There’s a clear disconnect here — yeah, sure, I’d love to use ML, but how do I even get started? It’s nice that computers can beat humans at video games, but how do I get from there to an impact my bottom line?
At SensAI, we believe strongly in piercing the hype and uncertainty to create concrete bottom-line business results using proven, practical machine learning methods.
To that end, I’d like to take you through some impressive case studies from companies just like yours. We’ll cover three areas that anybody could use a leg up in: marketing, sales, and customer service.
Importantly, none of these companies needed to use the most cutting-edge research; they were all able to deliver using tried and true techniques. Yes, maybe at a Google or a Facebook it makes sense to invest in R&D after they’ve tapped out on all of the existing techniques. But for the rest of you, the following illustrates big, quick wins you can implement using proven technology.
Marketing Case Study: Dole
In 2017, Dole Asia wanted to increase brand awareness of its Seasons fruit cocktail:
To do so, they built up a suite of in-house marketing creative — dozens of unique pieces of ad copy and visuals. The next step was to serve these ads across social, search, and email channels throughout Southeast Asia.
Conventionally, there are two options for deploying digital ads, and neither seem quite right:
- Marketing professionals set up hundreds of experiments on the ad platform, mixing and matching copy and visuals, target audiences, and more. Then, they must analyze the results of these experiments, re-allocate resources to the best-performing ones, and do it all over again. This drudgery is extremely costly from a human capital perspective, and it still leaves a lot on the table in terms of optimization.
- The ads can be deployed ad-hoc, randomly — each target user will see a set of ads, but not one that’s scoped any more specifically to them. This rudimentary approach certainly does not make the best use of the ad budget.
Dole went with a third strategy: employing a machine learning service called Albert. Albert took all of Dole’s marketing assets, and mixed and matched them to specific audiences — essentially performing the work of a marketing professional, but much quicker and cheaper. Albert took care of the targeting, rotation, and budget allocation for each ad set, with the goal of maximizing engagement.
The results from this 8-week experiment speak for themselves. Suppliers in the region quickly ran out of the product, and once they re-stocked, in-store business for the product grew a further 87% over the following 8 weeks.
To recap, by utilizing a machine learning tool to take care of the grunt work in their marketing process, Dole saw the following major benefits:
- Freeing up their professionals to focus on the creative side of marketing: generating unique and engaging assets
- Incredible results in sales for the product they were advertising
Sales Case Study: Vivint Smart Home
Vivint is a smart home technology provider in the US. In a fast-changing market, Vivint requires effective yet flexible sales techniques in order to continue growing. To that end, they decided to leverage the technique of personalized website sales copy.
While scoping the project, it quickly became clear that their goals were intractable for human labor. According to Jacob Parry, a Digital Manager at Vivint Smart Home:
“We needed to create highly localized copy for 12,000 unique web pages, which we calculated would take a team of human writers over two years to complete at a cost outside of our budget.”
So Vivint leveraged the power of Wordsmith, one of a category of tools utilizing NLG (Natural Language Generation). NLG helps improve sales efficiency by generating content that is unique and optimized for the source and location a website visitor originates from. Parry noted that, together, they were able to produce 30,000 pages over just a few months.
For an example, see the Vail, Colorado sales page: https://www.vivint.com/stores/city/vail-colorado. You’ll notice that the copy and images are tailored to that specific location. Zoom out to the Colorado state page and scroll down, and you’ll see that there are custom pages for almost every populated locale in the state: https://www.vivint.com/stores/state/colorado. The same goes for every state and province where Vivint operates.
In the year after these personalized pages went live, Vivint saw a 4x increase in local sales through organic search traffic.
By leveraging an NLG tool to create personalized website copy to sell more smart home systems, Vivint was able to drastically increase the output (30,000 pages vs. 12,000), reduce time (2+ years vs. several months) and cost (manually writing each page vs. generating content via machine learning), and improve results (4x greater local sales year over year).
Customer Service Case Study: Magoosh
Magoosh is an online platform for standardized test preparation (SAT, GRE, GMAT, etc.). Their team of 50 support staff (split into “community support” and educational tutors) was getting heavily bogged down as they scaled. They had a set of 900 macros in Zendesk for responding to common inquiries, and the staff spent an inordinate amount of time searching through these for the correct response. Additionally, there was too much dead work associated with attaching metadata to each support ticket: assigning it to a team, tagging it with an inquiry type, priority, pulling up relevant account information, etc. Customer satisfaction surveys run by the company showed that the lack of timely responses was a point of frustration.
To address this issue, Magoosh began an integration with a group called DigitalGenius. DigitalGenius got access to all of Magoosh’s historical support data and used it to train a machine learning model to aid in the above tasks.
First, DigitalGenius is able to suggest and surface response macros based on the content of a support ticket, rather than the agent having to go search through the database on their own. The algorithm is able to do this by comparing the content of the ticket with previous tickets in the database, and analyzing which responses were most commonly used to similar tickets in the past. This alone reduced the average support time per ticket by 30%.
Using the same pattern-extraction techniques, DigitalGenius was able to predict most of the metadata as well, including routing to the correct team, tagging, and assigning priority.
After a 6-month training period, DigitalGenius was able to handle 83% of incoming support tickets and its tag prediction had reached 92% accuracy.. This cut down on the response time for support tickets without compromising on the quality of support, leading to a happier support staff, and most importantly, happy customers.
Hopefully, these case studies begin to get some wheels spinning in your head. These three workflows are the cornerstone of the enterprise and there are certainly inefficiencies at your company that you would be interested in applying ML to.
While these results are impressive, they still barely scratch the surface of the power of machine learning. Keep in mind that the most important part of machine learning is the learning. At each of these companies, a flywheel effect has begun: as they gather more data from their customers, the algorithms will get better at their respective tasks, and the gap between them and their competitors will get wider and wider. So what are you waiting for?