Swedish house Mackmyra created a blend of single malt whisky, Intelligens, using AI with cloud-based compute and infrastructure services from Microsoft Azure coupled with data science expertise from Finnish services partner Fourkind. The whisky itself took center stage at a recent event marking its launch—it was delightful, with vanilla, citrus, and pear on both the nose and mouth.
HFS liked the potential applicability of this experimental foray using machine learning (ML) to generate the blend recipe. It demonstrates an approach to achieving mass customization across a range of domains with a design element at their core, such as food, drink, clothing, and shoes.
Framing the problem to be solved is the essential first step
ML excels at determining a model by ingesting large quantities of data. Patterns emerge, and then algorithms evaluate new data’s fit with established patterns. Conversely, anomalies and outliers not fitting the usual patterns stand out. AI operates almost instantaneously upon receipt of new data with pattern matching and outlier detection in disease identification, network optimization, fraud detection, and more.
But, the nature of the quest in creating a new whisky blend was different. The pattern matching element applied as a critic to newly generated recipes. The challenge was not to just create random blends. Nor was it to create the single most perfect, optimized whisky ever. The aim was generating multiple good options with a high probability of success.
Similarly, for many business challenges, multiple favorable outcomes are possible; for example, next best action functionality for contact center agents is not just suggesting one single action, it presumes a series of other actions will follow in time, too.
We see many enterprises in Exhibit 1 with significant intent to invest in AI and ML, it’s important that you take time to define and contextualize your problems before sourcing, prepping, and training data. Bad questions get bad answers.
Exhibit 1: AI and ML are seeing increased investment and focus in 2019
Source: HFS Research supported by KPMG, "State of Operations and Outsourcing” 2019 & 2018
Sample: Global 2000 Enterprise Leaders = 355 & 388
A two-stage process of generation and critique delivered a selection of blends to Mackmyra’s master blender
“All who study creativity agree that for something to be creative, it is not enough
for it to be novel: it must have value, or be appropriate to the cognitive demands of the situation.”
From “Creativity—Beyond the Myth of Genius” by Robert W. Weisberg
Fourkind built Mackmyra’s ML models on Microsoft’s Azure cloud platform using Machine Learning Studio, Microsoft’s browser-based visual drag-and-drop authoring environment for predictive analytics. Training data incorporated Mackmyra’s existing recipes (75), cask types (100), sales data, customer preferences, employee preferences, and industry or critic preferences (evidenced by awards). The datasets were small, but the combination of datasets was enough to enable the generation of millions of recipes. In the first stage, the model used recursion to limit the subspace on a loosely fit decision tree model, generating blend recipes with controlled exploration using Monte Carlo simulation.
The second stage of the model ran newly generated recipes through an AI “critic” using collaborative filtering and gradient boosting. The model scored newly generated recipes’ success using past recipes’ ratings derived from customers, employees, and industry accolades. Interestingly, the critic part of the model resembles strategies deployed in training AI to play a game to a high standard. The algorithm constantly tried to beat its previous high score, discarding blends with lower scores.
More complex options than this could have been deployed, like using reinforcement learning with policies for rewards established by human tasters.
AI made suggestions; Mackmyra’s master blender made the final decision
One hundred million iterations later, the model delivered a batch of contenders for master blender Angela D’Orazio’s consideration—unique recipes that fit the house style and scored well when critiqued against what has previously worked well for Mackmyra.
Some recipes were deprioritized according to Angela’s expertise—we could call this bias! The human element is key. Calling herself the “mentor” of the AI-infused whisky, Angela has taste buds, prior experience, and knowledge that allowed her to eliminate some recipes. AI cannot replace the human palette.
An entirely practical filter eliminated some proposed blends based on logistics; specifically, what cask types were available in the warehouse. Angela narrowed the set of recipes to five contenders, and then based on taste, selected the whisky blend later named Intelligens. The blend was duly mixed and flavors married in casks in Mackmyra’s whisky mine (warehouse) for a couple of months.
The Bottom Line: A head start beats a blank page. In commercial circumstances where the creative element is not for the sake of artistic merit, good options suffice. A guardrailed, data-based incorporation of design style that rapidly generates unique designs (here, a whisky blend) can customize offerings at scale.
Faster than a person launching the creative process, AI algorithms can sift through and perform calculations on vast amounts of data. AI can predict each option’s likelihood of success and rate and rank as many combinations as needed in less time than it takes to drink a whisky.
In 1987, Stan Davis coined the term “mass customization” in his book Future Perfect, personalization, hyper-personalization and markets-of-one are now on the ascent. This exotic application of ML is not necessarily practical in terms of effort, resources, and cost to create only one new possibility. The exciting part here is that the method Mackmyra used can create many options, enabling a faster time to market of lateral designs at scale.
Whisky has many faces, and some brands fit the mass-production model. Mackmyra is a boutique brand; 30-liter casks, either individually or group-owned, are very popular. Releases are small. This scarcity, combined with high demand for popular blends, has led to some sensational pricing in the past. Mass customization turbocharges the boutique model by producing lots of variations suitable for granular market segmentation—potentially per individual customer.
For more information on Mackmyra’s Intelligens, the AI-created whisky, click here.