But What Does It Mean!? How Enterprise AI is Going to (Further) Revolutionize Business

We keep hearing that artificial intelligence is going to change the face of business forever. I happen to agree – it’s why I started milk+honey – but what’s missing from most of these declarations is a clear explanation of how AI is going to revolutionize business. What, exactly, is going to be different? This piece is the first in a series I’ve written to explain what AI actually looks like in a business context. My hope is that business leaders running companies outside of the data-native technology sector will start to see what’s possible with enterprise-based AI (EAI), and why its impact is indeed revolutionary. From there, future pieces I publish will make sense, and the focus will be on how to get things done.

Accuracy & Precision

I think decisions are a form of executive currency. They have to be made, and there’s a lot at stake. The most successful business leaders push decisions “down,” deeper into their organizations, in order to free themselves up to make more nuanced, strategic decisions, and to develop organizational talent. The risk, of course, is that some decision somewhere along the line of decisions that get made every day in every organization, generates an unexpected result. At that point we’re all in reactive mode, racing to minimize fallout or capitalize on unanticipated opportunities.

It’s here, at the heart of the decision-making process, that AI is poised to change the face of business forever. The first thing to know about AI in an enterprise setting is that it transforms estimates and forecasts into much more precise, accurate predictions. It gets us a country mile closer to what we’ve always wanted: a crystal ball that removes enormous chunks of risk from the decision-making process. Suddenly opportunities aren’t so unanticipated, and there’s a lot less fallout to minimize. Suddenly everyone from the c-suite to the front line is free to make BIGGER decisions with greater potential to positively impact business performance. This radical increase in accuracy translates into higher degrees of confidence throughout an organization, which in turn drives fundamental changes in behavior. Billions of dollars are already entrusted to machine-learning (another two words for AI) trading systems every day, and the practice of “preventative” maintenance is already being replaced with “predictive” and “prescriptive” maintenance practices on shop floors around the world.

The second thing to know about EAI is that it’s replacing our reliance on aggregate statistics and averages with data points about every individual in a population. Businesses no longer have to settle for playing to the majority. Everything from products to promotional campaigns to pricing can be customized for each customer. Not only can we determine much more precisely how much demand there will be for something, we can know precisely where it needs to be ahead of time.

All this transformative awesomeness is available now, but it comes at a price that many business leaders (again, outside of the data-native technology sector) are hesitant to pay…TRUST. This isn’t surprising for two reasons. First, trust is famously hard to gain and easy to lose. It’s especially difficult to trust something we don’t understand. Secondly, I think business leaders know at an intuitive level that once AI technologies are present in the workplace, nothing will ever be the same. That’s a scary proposition because it requires a willingness to embrace fundamental change. We face a paradox, in which our greatest opportunity to minimize business risk requires us to risk a great deal of the Great Familiar.

As with so many things, the way through this conundrum is perspective. So here’s a framework for beginning to understand how AI works in business. There are two principal categories for enterprise AI application: automation and analytics. 

Automation

Think of automation applications as those designed to perform a repeatable task previously performed by a person or a number of people. Automation-based applications simulate a human’s ability to see, hear, sense and communicate.

Automation opportunities are not limited to obvious manufacturing and customer support examples. Among countless other applications, automation models are being used to visually inspect output on a manufacturing line, recognize handwriting on waybills in distribution, track customer activity in stores, interpret charge receipts to automate expense reporting, and detect emotion in facial expressions. Speech recognition models typically used to automate customer service call routing scenarios are now being leveraged to create simplified speech-driven interfaces for manufacturing equipment. Inside the legal system, testimony transcription is being automated. Voice-based biometric security and emotion detection are also just around the corner.

We’ve only scratched the surface of applications of natural language processing. Language translation models like the one used in Google’s highly publicized translating earbuds have numerous applications for many businesses. The legal system is also one of the most active early adopters of written language models, leveraging natural language processing (NLP) to automate trademark and patent searches, due diligence and contract evaluation. Content topic detection is already broadly used for personalizing content recommendations. Finally, content interpretation and summarization capabilities are also maturing quickly. Imagine asking Google a question and getting back a synthesized answer instead of a list of links for you to investigate on your own.

Analytics

If, at the highest level, automation-based AI technologies enable us to function more efficiently, then analytics-based AI helps us function more strategically. These applications do one of three things:

• predict numeric quantities
• calculate the probability of outcomes
• bucket things into distinct categories

Numeric Prediction Models

Think of machine-learning (ML) models that predict numeric quantities as a logical extension of the forecasting process. Any continuous, measurable quantity can be forecasted: dollars, items, measures of time, etc. Some ML numeric prediction models employ a regression technique similar to what statisticians have been using for decades. The key difference is that non-ML-based regression models require normalized, structured data sets, and typically can’t be executed in real time. ML models, on the other hand, can analyze HUGE volumes and a whole variety of data types, both structured and unstructured, all in real time. This is what allows us to move more confidently from estimates to extremely accurate predictions.

High frequency trading is among the most well-known examples of this type of machine learning application. Other celebrated examples of this type of machine learning include customer lifetime value prediction and time-to-equipment-failure prediction.

Another less obvious but incredibly powerful use case for numeric prediction models is real-time scenario testing. This is exactly what’s happening in the mapping applications on our smartphones. These apps leverage a combination of historical route times, current conditions and even real-time data from other users to estimate and recommend the best route at any given moment. The same approach can be used to optimize any process or system, including logistics routing, dynamic pricing and automatic investment profile management.

Outcome Prediction Models

The way these models are typically used is to first calculate the probability of a desired outcome for each item on a list of alternatives, then sort and output the resulting list. Applications that work this way are often referred to as recommenders or recommendation engines. In these models, an event refers to an action that is taken in hopes of generating a desired result or outcome. Examples of events include purchases, click-throughs, requests, and connections.

There are a couple different ways to leverage recommenders. The first is to personalize an experience for an individual user. High-profile examples include Amazon’s product recommendations, Netflix’s movie recommendations, and Facebook’s friend suggestions. Recommendation engines are being used to customize a wide variety of customer sales, support, or marketing interactions. They are equally effective at serving employees through personalized development and onboarding plans, and for enhancing productivity by targeting especially relevant content to a particular employee. 

Recommendation engines are helping HR managers sort through candidate lists, sales people optimize leads and deals, and operations managers decide on their next expansion. They’re also helping employees prioritize their work independently, ensuring the actions they take are those that will achieve the most meaningful and desired outcomes.

Think of recommenders as response flow lists on steroids. The key difference is that recommended actions are highly customized for the exact situation and even for the specific skillset and experience of an employee.

Classification Models

Classification models bin things into one of two or more categories. Categories are generally referred to as classes in the data science world, and the resulting applications are called classifiers or classification models.

The first use case for classifiers is to detect things. Maybe they’re flagging something that may even only subtly stand out as an anomaly, such as fraudulent financial transactions and network intrusions. Or maybe they’re identifying things in order to differentiate between them. Examples include spam vs. valid emails, positive vs. negative medical diagnosis, and whether a customer tweet is positive, neutral or negative. The same technologies can be applied to financial systems and networks in business, to identify unhappy employees.

The second use case for classifiers is to predict things. Credit approval models predict whether an applicant will repay or default on a loan. Churn models predict whether or not a customer is about to defect. The same technology can be applied to predict employee departures. HR teams are developing models to predict whether or not employees or candidates will be successful in a new role. Law firms are developing models that predict the outcomes of litigation. Sales departments are using models to predict imminent purchases, so they can anticipate and answer inquiries before they’re even asked. The possibilities are quite literally endless.

Predicting & Prescribing

The value of analytics-based machine-learning models doesn’t stop with making better real-time decisions. All three types of machine learning models can be leveraged to explore “what if” scenarios, so the impact of actions is known before they’re taken, effectively predicting the future. R&D groups can explore the impact of including or not including a given product feature. Sales departments can model the impact of raising and lowering prices, as well as the impact of a competitor’s change in price.

Machine-learning models are also being mined for insights and understanding about why things happen. When business leaders begin to influence the factors they’ve identified, they find themselves in a position to actually make things happen – effectively prescribing the future. For example, knowledge of traits that the most successful employees have in common can be used to actively recruit more people with those traits. Digging into the history of the manufacturing machines that fail most often could identify causes for the sub-par performance that could be avoided in the future.

Where To Next?

Hopefully I’ve provided a readable lay of the land where machine learning is integrated into everyday business practices, and communicated that there’s nothing everyday about it. From here, it’s a matter of exploring opportunities to benefit from machine learning in your particular organization, as well as what it will take to make it stick.