Summary.

   In order to realize AI’s giant potential, CMOs need to have a good grasp of the various kinds of
applications available and how they may evolve. This article guides marketing executives through the current
state of AI and presents a framework that will help them...more

Of all a company’s functions, marketing has perhaps the most to gain from artificial

intelligence. Marketing’s core activities are understanding customer needs, matching
them to products and services, and persuading people to buy—capabilities that AI can
dramatically enhance. No wonder a 2018 McKinsey analysis of more than 400 advanced
use cases showed that marketing was the domain where AI would contribute the
greatest value.

Chief marketing officers are increasingly embracing the technology: An August

2019 survey by the American Marketing Association revealed that
implementation of AI had jumped 27% in the previous year and a half. And a
2020 Deloitte global survey of early AI adopters showed that three of the top
five AI objectives were marketing-oriented: enhancing existing products and
services, creating new products and services, and enhancing relationships with
customers.

While AI has made inroads in marketing, we expect it to take on larger and larger roles
across the function in the coming years. Given the technology’s enormous potential, it’s
crucial for CMOs to understand the types of marketing AI applications available today
and how they may evolve. Drawing on more than a decade of experience studying data
analytics, AI, and marketing and advising companies across industries about them,
we’ve developed a framework that can help CMOs classify existing AI projects and plan
the rollout of future ones. But before we describe the framework, let’s look at the current
state of play.

Today’s AI
Many firms now use AI to handle narrow tasks, such as digital ad placement (also
known as “programmatic buying”); assist with broad tasks, like enhancing the accuracy
of predictions (think sales forecasts); and augment human efforts in structured tasks,
such as customer service. (See the sidebar “Well-Established AI Applications in
Marketing” for a list of some common activities AI can support.)
Firms also employ AI at every stage of the customer journey. When potential
customers are in the “consideration” phase and researching a product, AI will
target ads at them and can help guide their search. We see this happening at
the online furniture retailer Wayfair, which uses AI to determine which
customers are most likely to be persuadable and, on the basis of their
browsing histories, choose products to show them. And AI-enabled bots from
companies such as Vee24 can help marketers understand customers’ needs,
increase their engagement in a search, nudge them in a desired direction (say,
to a specific web page), and if needed, connect them to a human sales agent by
chat, phone, video, or even “cobrowsing”—allowing an agent to help the
customer navigate a shared screen.

AI can streamline the sales process by using extremely detailed data on individuals,
including real-time geolocation data, to create highly personalized product or service
offers. Later in the journey, AI assists in upselling and cross-selling and can reduce the
likelihood that customers will abandon their digital shopping carts. For example, after a
customer fills a cart, AI bots can provide a motivating testimonial to help close the sale—
such as “Great purchase! James from Vermont bought the same mattress.” Such
initiatives can increase conversion rates fivefold or more.

After the sale, AI-enabled service agents from firms like Amelia (formerly IPsoft) and
Interactions are available 24/7 to triage customers’ requests—and are able to deal with
fluctuating volumes of service requests better than human agents are. They can handle
simple queries about, say, delivery time or scheduling an appointment and can escalate
more-complex issues to a human agent. In some cases AI assists human reps by
analyzing customers’ tone and suggesting differential responses, coaching agents about
how best to satisfy customers’ needs, or suggesting intervention by a supervisor.

The Framework
Marketing AI can be categorized according to two dimensions: intelligence level and
whether it’s stand-alone or part of a broader platform. Some technologies, such as
chatbots or recommendation engines, can fall into any of the categories; it’s how they’re
implemented within a specific application that determines their classification.

Let’s look at the two types of intelligence first.

Task automation.
 
These applications perform repetitive, structured tasks that require relatively low levels
of intelligence. They’re designed to follow a set of rules or execute a predetermined
sequence of operations based on a given input, but they can’t handle complex problems
such as nuanced customer requests. An example would be a system that automatically
sends a welcome email to each new customer. Simpler chatbots, such as those available
through Facebook Messenger and other social media providers, also fall into this
category. They can provide some help to customers during basic interactions, taking
customers down a defined decision tree, but they can’t discern customers’ intent, offer
customized responses, or learn from interactions over time.

Machine learning.
 

These algorithms are trained using large quantities of data to make relatively complex
predictions and decisions. Such models can recognize images, decipher text, segment
customers, and anticipate how customers will respond to various initiatives, such as
promotions. Machine learning already drives programmatic buying in online
advertising, e-commerce recommendation engines, and sales propensity models in
customer relationship management (CRM) systems. It and its more sophisticated
variant, deep learning, are the hottest technologies in AI and are rapidly becoming
powerful tools in marketing. That said, it’s important to clarify that existing machine-
learning applications still just perform narrow tasks and need to be trained using
voluminous amounts of data.

Now let’s consider stand-alone versus integrated AI.

Stand-alone applications.
 

These are best understood as clearly demarcated, or isolated, AI programs. They’re

separate from the primary channels through which customers learn about, buy, or get
support for using a company’s offerings, or the channels employees use to market, sell,
or service those offerings. Put simply, customers or employees have to make a special
trip beyond those channels to use the AI.

Consider the color-discovery app created by Behr, the paint company. Using IBM
Watson’s natural language processing and Tone Analyzer capabilities (which detect
emotions in text), the application delivers several personalized Behr paint-color
recommendations that are based on the mood consumers desire for their space.
Customers use the app to short-list two or three colors for the room they intend to paint.
The actual sale of paint is then executed outside the app, although it does allow a
connection to order from Home Depot.
Integrated applications.
 
Embedded within existing systems, these AI applications are often less visible than stand-alone
ones to the customers, marketers, and salespeople who use them. For example, machine
learning that makes split-second decisions about which digital ads to offer users is built into
platforms that handle the entire process of buying and placing ads. Netflix’s integrated machine
learning has offered customers video recommendations for more than a decade; its selections
simply appear in the menu of offerings viewers see when they go to the site. If the
recommendation engine were stand-alone, they would need to go to a dedicated app and request
suggestions.

Makers of CRM systems increasingly build machine-learning capabilities into their

products. At Salesforce, the Sales Cloud Einstein suite has several capabilities, including
an AI-based lead-scoring system that automatically ranks B2B customer leads by the
likelihood of purchase. Vendors like Cogito, which sells AI that coaches call center
salespeople, also integrate their applications with Salesforce’s CRM system.
Combining the two types of intelligence and two types of structure yields the four
quadrants of our framework: stand-alone machine-learning apps, integrated machine-
learning apps, stand-alone task-automation apps, and integrated task-automation apps.

Understanding which quadrant applications fall into can help marketers plan and
sequence the introduction of new uses.

A Stepped Approach
We believe that marketers will ultimately see the greatest value by pursuing integrated
machine-learning applications, though simple rule-based and task-automation systems
can enhance highly structured processes and offer reasonable potential for commercial
returns. Note, however, that nowadays task automation is increasingly combined with
machine learning—to extract key data from messages, make more-complex decisions,
and personalize communications—a hybrid that straddles quadrants.

Stand-alone applications continue to have their place where integration is difficult or

impossible, though there are limits to their benefits. Therefore, we advise marketers to
move over time toward integrating AI within current marketing systems rather than
continue with stand-alone applications. And indeed, many companies are heading in
that overall direction; in the 2020 Deloitte survey, 74% of global AI executives agreed
that “AI will be integrated into all enterprise applications within three years.”
Getting Started
For firms with limited AI experience, a good way to begin is by building or buying
simple rule-based applications. Many firms pursue a “crawl-walk-run” approach,
starting with a stand-alone non-customer-facing task-automation app, such as one that
guides human service agents who engage with customers.

Once companies acquire basic AI skills and an abundance of customer and market data,
they can start moving from task automation to machine learning. A good example of the
latter is Stitch Fix’s clothing-selection AI, which helps its stylists curate offers for
customers and is based on their self-reported style preferences, the items they keep and
return, and their feedback. These models became even more effective when the company
began to ask customers to choose among Style Shuffle photos, creating a valuable source
of new data.

New sources of data—such as internal transactions, outside suppliers, and even

potential acquisitions—are something marketers should look for constantly, since most
AI applications, particularly machine learning, require vast amounts of high-quality
data. Consider the machine-learning-based pricing model that the charter jet firm XO
used to increase its EBITDA by 5%: The key was to tap external sources for data on the
supply of private jets and on factors that affect demand, such as major events, the

macroeconomy, seasonal activity, and the weather. The data XO uses is publicly
available, but it’s a good idea to also seek proprietary sources whenever possible,
because models using public data can be copied by competitors.