AI At Work: Brands Share Real-World Results

Around 200 senior executives came together last week for the invitation-only VB Summit 2017, where they learned about how artificial intelligence and machine learning are improving customer support, enabling visual search, and augmenting, rather than replacing, human workers.

Around 200 senior executives came together at Berkeley, Calif.’s Claremont Hotel last week for the invitation-only VB Summit 2017, which focused on “Riding the AI Wave.” (Note: CMO.com was a media partner.) Attendees, who included investors, CMOs from marketing analytics companies, product managers from AI and IoT startups, and operators of tech accelerators, came to hear how artificial intelligence and machine learning are improving customer support, enabling visual search, and augmenting, rather than replacing, human workers.

In one early session, provocatively titled “AI Will Drop Your Customer Support By 90%,” Gregg Spratto, VP of operations at Autodesk, outlined how the company had turned over much of its initial support interactions to a virtual agent.

“The types of things we get asked haven’t changed for five to 10 years,” Spratto said. That made it feasible to train its agent, named AVA (“Autodesk virtual agent”), to at least screen callers. It used to take the company up to 38 hours to reply to a simple request for an activation code, said Spratto; now the time is down to 5.4 minutes, and the cost per case is down from $15 to $200 to $1.

Spratto emphasized, however, that the successful implementation and use of AVA was not a push-button process. The company worked with subject-matter experts to train the agent to distinguish between questions that look similar but aren’t. AVA also has to be able to understand the social and psychological aspects of a conversation, he said. To that end, Autodesk employs “conversational engineers” to teach the agent how to glean the intent behind a user query.

Man vs. Machine

Autodesk’s need for conversational engineers is an example of another issue around AI: the question of how much can be passed off to a machine and how much still requires human intervention. According to Spratto, about half of what AVA does is just finding out what the customer’s problem is before passing it along to the right person.

“We get humans out of troubleshooting and into problem solving,” he said.

Rajat Mishra, VP of worldwide services strategy and innovation at Cisco, described the “Iron Man lesson” that Cisco learned from its experience with predictive analytics. In the action movie, the robot suit is impressive, Mishra said, but it only becomes great when paired with Tony Stark (played by Robert Downey Jr.). Along those lines, Cisco determined that the human connection will always be important and has set a target of having 30% of its work handled by humans. But by pairing AI and humans, Cisco is now able to predict outages 32 hours in advance.

Visual Search

One subject of particular interest to marketers is the rising importance of visual search. Li Fan, Pinterest’s head of engineering, discussed the site’s “shop the look” feature, launched earlier this year. The feature, which provides viewers with recommendations of buyable items similar to something in a pinned photo.

Developing “shop the look” used both traditional and neural-net machine learning to identify appropriate products. First, Pinterest relied on humans to properly tag a large training set of images. “The users, in some ways, labeled the data for us,” Fan said, through comments on photos like “my favorite sundress.”

From that set and through machine learning, “shop the look” could learn to properly identify objects in other photos. Humans still curate the selections to make sure they’re good matches, but the machine learning did the initial searching through the millions of images available.

Couresty of machine learning and AI, “shop the look” doesn’t rely on existing categories and judgments the way a human would, Fan said. In other words, they don’t need to distill your taste into a particular style. It can suggest articles of clothing that match, for example, without having to decide if your look is chic or punk rock.

Another presenter, Ramzi Rizk, explained his company EyeEm’s use of visual search to drive its amateur stock photo agency. The results of a photo search are 80% automated, he said, and the firm’s AI can recognize not just quality but marketability. Brands can search through aesthetics, not just labels and tags.

Marketing-Specific

The second day of the conference featured “boardroom sessions” focusing on particular applications of AI, including “Changing the Game in Advertising and Marketing with AI and Machine Learning.” One co-chair was Ophir Tanz, CEO of computer vision company GumGum, which offers “integrated advertising” that places marketing messages in line with content (similar to the popup promos during a TV show). Ad placement is according to analysis of the images they overlay, not by text on the page. But, Tanz said, that raises the need for context awareness.

Talk also turned to AI’s ability to generate creative material, which, co-chair Massimo Portincaso, partner and managing director at the Boston Consulting Group, pointed out is already happening. Regarding AI’s ability to understand and follow brand guidelines, Portincaso said he would give a creative agency a “sandbox” they could work within—broad parameters for the creative that AI could generate, and from which the marketer would then choose.

“What creative needs to learn is to switch their tools,” he said. “The job of you as a creative is more to be selective.”

Getting Started

The road to effective use of AI, many presenters agreed, starts with clean, existing data. You train AI and machine learning by feeding it properly coded data and then reviewing the choices it makes based on that. The good news is, most companies already have lots of data to work with. Autodesk, for example, started with the simple questions that its customer service team gets asked over and over. Over time, it built that out to handle more complex cases, with five employees dedicated to training AVA to handle them.

To develop a model that consumers will want to use, you need an in-house team, Lin advised. A company can get a generic AI model to work maybe 80% of the time, she said, but that’s it. She recommended installing a new model and training it, but keeping the traditional model around and comparing to see which is more efficient.

And don’t limit yourself by preconceptions, Cisco’s Mishra added. He advised taking a “first principles” approach to scouting and nurturing AI and machine-learning initiatives (what he called the “Moneyball” lesson). Be prepared to defy tradition: “We looked everywhere,” he said, to find projects worth developing further.

All photos by Michael O’Donnell/O’Donnell Photography