Want your company’s A.I. project to succeed? Don’t hand it to the data scientists, says this CEO

Aible's Arijit Sengupta says too many business A.I. projects lose sight of the bottom line

Portrait photo of Aible CEO Arijit Sengupta

Aible CEO Arijit Sengupta says that data scientists too often lose track of the bottom line. Photo courtesy of Aible

Arijit Sengupta once wrote an entire book titled Why A.I. is A Waste of Money. That’s a counterintuitive title for a guy who makes his money selling A.I. software to big companies. But Sengupta didn’t mean it ironically. He knows firsthand that for too many companies, A.I. doesn’t deliver the financial returns company officials expect. That’s borne out in a slew of recent surveys, where business leaders have put the failure rate of A.I. projects at between 83% and 92%. “As an industry, we’re worse than gambling in terms of producing financial returns,” Sengupta says.

Sengupta has a background in computer science but he also has an MBA. He founded BeyondCore, a data analytics software company that Salesforce acquired in 2016 for a reported $110 million. Now he’s started Aible, a San Francisco-based company that provides software that makes it easier for companies to run A.I. algorithms on their data and build A.I. systems that deliver business value.

Aible makes an unusual pledge in the A.I. industry: it promises customers will see positive business impact in 30 days, or they don’t have to pay. Their website is chock full of case studies. The key, Sengupta says, is figuring out what data the company has available and what it can do easily with that data. “If you just say what do you want, people ask for the flying car from Back to the Future,” he says. “We explore the data and tell them what is realistic and what options they have.”

One reason most A.I. projects fail, as Sengupta sees it, is that data scientists and machine learning engineers are taught to look at “model performance” (how well does a given algorithm do with a given data set at making a prediction) instead of business performance (how much money, in either additional revenue or cost-savings, can applying A.I. to a given dataset generate).

To illustrate this point, Aible has run a challenge in conjunction with UC Berkeley: it pits university-level data science students against high school 10th graders using a real-world data set comprised of 56,000 anonymized patients from a major hospital. The competing teams must find the algorithm for discharging patients from the 400-bed hospital that will make the hospital the most money, understanding that keeping patients in the hospital unnecessarily adds costs, but so does making a mistake that sees the same patient later readmitted. The winner gets $5,000. The data scientists can use any data science software tools they want, while the high school kids use Aible’s software. The high school kids have beaten the data scientists—by a mile—every time they’ve run the competition, Sengupta says.

Subscribe to unlock this article and get full access to Fortune.com

Arijit Sengupta once wrote an entire book titled Why A.I. is A Waste of Money. That’s a counterintuitive title for a guy who makes his money selling A.I. software to big companies. But Sengupta didn’t mean it ironically. He knows firsthand that for too many companies, A.I. doesn’t deliver the financial returns company officials expect. That’s borne out in a slew of recent surveys, where business leaders have put the failure rate of A.I. projects at between 83% and 92%. “As an industry, we’re worse than gambling in terms of producing financial returns,” Sengupta says.

Sengupta has a background in computer science but he also has an MBA. He founded BeyondCore, a data analytics software company that Salesforce acquired in 2016 for a reported $110 million. Now he’s started Aible, a San Francisco-based company that provides software that makes it easier for companies to run A.I. algorithms on their data and build A.I. systems that deliver business value.

Aible makes an unusual pledge in the A.I. industry: it promises customers will see positive business impact in 30 days, or they don’t have to pay. Their website is chock full of case studies. The key, Sengupta says, is figuring out what data the company has available and what it can do easily with that data. “If you just say what do you want, people ask for the flying car from Back to the Future,” he says. “We explore the data and tell them what is realistic and what options they have.”

One reason most A.I. projects fail, as Sengupta sees it, is that data scientists and machine learning engineers are taught to look at “model performance” (how well does a given algorithm do with a given data set at making a prediction) instead of business performance (how much money, in either additional revenue or cost-savings, can applying A.I. to a given dataset generate).

To illustrate this point, Aible has run a challenge in conjunction with UC Berkeley: it pits university-level data science students against high school 10th graders using a real-world data set comprised of 56,000 anonymized patients from a major hospital. The competing teams must find the algorithm for discharging patients from the 400-bed hospital that will make the hospital the most money, understanding that keeping patients in the hospital unnecessarily adds costs, but so does making a mistake that sees the same patient later readmitted. The winner gets $5,000. The data scientists can use any data science software tools they want, while the high school kids use Aible’s software. The high school kids have beaten the data scientists—by a mile—every time they’ve run the competition, Sengupta says.

The teens, Sengupta says, are able to keep their eyes on the bottom line. They’re not concerned with the particular model that Aible suggests (Aible works by training hundreds of different models and finding the one that works best for a given business goal), whereas the data scientists get caught up on training fancy algorithms and maximizing accurate discharge predictions, but losing sight of dollars and cents.

Sengupta’s point is that ignoring, or not actually understanding, the business use of an A.I. system can be downright dangerous. He describes what he calls the “A.I. death spiral,” where an A.I. system maximizes the wrong outcome and literally runs a business into the ground. Take for example an A.I. system designed to predict which sales prospects are most likely to convert to paying customers. The system can achieve a higher accuracy score by being conservative—only identifying prospects that are highly likely to convert. But that shrinks the pool of possible customers significantly. If you keep running this optimization process using only the small number of customers who convert, the pool will just keep shrinking, until eventually the business winds up with too few customers to sustain itself. Customer win rate, Sengupta says, is the wrong metric—the A.I. should be trained to optimize revenue or profits, or maybe overall customer growth, not conversion rates.

Sidestepping these pitfalls requires a little bit of machine learning understanding, but a lot of business understanding. Sengupta is not alone in hammering home this theme. It’s a point that a lot of those working on A.I. in commercial settings—including deep learning pioneers such as Andrew Ng—are increasingly making: algorithms and computing power are, for the most part, becoming commodities. In most of the case studies on Aible’s website, customers used the startup’s cloud-based software to train hundreds of different models, sometimes in less than 10 minutes of computing time. Then the business picks the model that works best.

What differentiates businesses in their use of A.I. is what data they have, how they curate it, and exactly what they ask the A.I. system to do. “Building models is becoming a commodity,” Sengupta says. “But extracting value from the model is not trivial, that’s not a commodity.”

With that, here’s the rest of this week’s A.I. news.

Jeremy Kahn
@jeremyakahn
jeremy.kahn@fortune.com

A.I. IN THE NEWS

OpenAI moves towards commercializing its image-from-text generating A.I. DALL-E. The San Francisco-based A.I. lab, which is backed by Microsoft, has said that it has invited 1 million people from its waitlist to begin using DALL-E, an A.I. system that can generate all kinds of images, including photorealistic ones, from a text description. The system has been used by digital artists and to provide inspiration to design teams and could be used for commercial image generation. According to a blog post, OpenAI is giving users a certain number of free credits each month, but an additional 115 credits can be purchased for $15, signaling that OpenAI intends to turn DALL-E into a commercial product. This would be its second product after GPT-3, the language generation A.I. system.

Google fires researcher who made claims about “sentient” A.I. chatbot. The technology giant has fired Blake Lemoine, the A.I. researcher who claimed that Google’s LaMDA chatbot had become sentient, according to the newsletter Big Technology. The company had previously placed Lemoine on paid administrative leave. The company says it fired the researcher for continuing to violate its confidentiality policies, potentially jeopardizing trade secrets, and for continuing to undertake provocative actions, such as claiming he hired a lawyer to protect LaMDA’s interests. Lemoine had leaked a number of internal company documents—including copies of his dialogues with LaMDA—to the press and Congressional staffers. The company says it has investigated Lemoine’s claims that LaMDA is sentient and found them to be false. The vast majority of A.I. researchers say that Lemoine’s claims are incorrect and that a system designed in the way Google says it designed LaMDA could never acquire sentience.

Walmart’s automatic re-stocking system is clogging stores with too much inventory. That’s according to a report in Insider, which says employees are complaining about having to stock vast piles of items in nursing rooms and supply cupboards. The retailer told the publication that “As inventory has become an issue for all retailers, we’ve increased our focus on creating a safe working environment.” The problem, according to several experts Insider talked to, is that after experiencing unexpected shortages during the pandemic due to unusual demand trends and supply chain disruptions, Walmart has seemingly now set its automatic ordering system to ensure that stores have “just-in-case” inventory on hand. But it may have now erred too much on the side of overstocking, especially as buying behavior normalizes and supply chain disruptions abate. Demand may even now be dropping due to inflation and recession fears. It is not clear from the Insider story how “smart” Walmart’s restocking system is or if it uses any A.I., or whether is a more static, rule-based piece of software. 

A chess robot grabbed and broke a child’s finger after the child moved a piece too fast on the board. The chess-playing robot was taking part in a demonstration in Moscow where it played multiple games against multiple opponents simultaneously. In a game against a top-ranked seven year-old. the robot suddenly grabbed the child’s finger and held it fast, forcing several adults observing the match to rush to boy’s aid. They freed his finger, but it was broken—although it isn’t clear if the injury occurred because of the pressure the robot exerted or if it fractured as the adults wrested it free from the robot’s grasp. The incident seems to have occurred because the boy made a move before the robot, which had just taken one of the boy’s pieces, according to The Guardian. The incident demonstrates how hard it can be to predict how A.I.-enabled robots will react to unusual situations that they have not encountered during their training—and why we need to be especially cautious about deploying such robots into situations where they are in close physical proximity to people. 

DataRobot CEO resigns amid stock sale controversy. Dan Wright, the CEO of Boston-based A.I. software company DataRobot, resigned last week amid employee outcry over the revelation that he and other top DataRobot executives had quietly sold $32 million worth of company stock last year, when the private company’s valuation peaked at $6.3 billion, while the rest of the company’s 1,200 employees were not given the same opportunity. The company has since faced a string of negative financial forecasts and announced lay-offs. The revelations, first reported in The Information, lead the company’s chief A.I. evangelist Ben Taylor to resign in protest.

EYE ON A.I. TALENT

DataRobot, the Boston-based A.I. company, has named Debanjan Saha its new chief executive officer,  the company said in a press release. Saha had been the company’s president and chief operating officer. He had been a former general manager of data analytics at Google before joining DataRobot earlier this year.

Global law firm BCLP (Bryan Cave Leighton Paisner) has hired Dan Surowiec as its chief information officer, according to a story in trade journal The Global Legal Post. Surowiec had previously been CEO at the Chicago-based international law firm Baker McKenzie

EYE ON A.I. RESEARCH

Sony Gran Turismo A.I. bested all comers at the car racing video game, and learned “etiquette.” In a feature in MIT Technology Review, writer Will Douglas Heaven looks at Gran Tourismo Sophy, an A.I. system that Sony built to play the car racing video game Gran Turismo, which is used for many e-sports competitions. At first, Sony just tried to train the A.I. to be able to stay on the track at faster speeds than the vast majority of human video gamers could handle. That alone led to some revelations. For example, human opponents were “struck by the way GT Sophy took corners, braking early before accelerating out on a much tighter line” than they normally would. ‘It used the curve in a weird way, doing stuff that I just didn’t even think of,'” one human racer told the magazine. For example, GT Sophy would often drop a wheel onto the grass and skid into turns, a move that the human drivers considered far too risky and difficult to pull off without crashing.

But, according to Heaven’s story, Sony found that even these super-human abilities weren’t enough to allow GT Sophy to win consistently in real competitions, where human referees can deduct points from cars that drive dangerously. GT Sophy drove so aggressively it frequently risked hitting other vehicles, forcing the human racers to abandon their lines. The A.I. racked up too many penalty points to win. So then Sony focused on training the A.I. to exhibit the same kind of “racing etiquette’ that human race car drivers employ—the subtle knowledge of knowing when to pass aggressively or make an intimidating swerve to convince an opponent not to attempt a pass, and when to back off throttle and be more cautious.

Peter Wurman, head of Sony AI America, says this kind of etiquette is an “example of the kind of dynamic, context-aware behavior that robots will be expected to have when they interact with people,” in lots of real world situations. “An awareness of when to take risks and when to play it safe would be useful for AI that is better at interacting with people, whether it be on the manufacturing floor, in home robots, or in driverless cars,” Heaven writes. Wurman says, “I don’t think we’ve learned general principles yet about how to deal with human norms that you have to respect, But it’s a start and hopefully gives us some insight into this problem in general.”

FORTUNE ON A.I.

Commentary: A new generation of data scientists could be our best weapon against climate change—by Weiwei Pan

A.I.-powered robotic lawnmowers help you reclaim your summer days—by Chris Morris

GM wants to test a self-driving car that doesn’t have a steering wheel, pedals or mirrors—by Anurag Kotoky and Bloomberg

BRAINFOOD

Indie genre fiction writers are increasing turning to A.I. writing assistants to help them churn out books as fast as fans demand. That’s according to a story in tech publication The Verge. The piece looks at the world of authors who make money self-publishing science fiction, fantasy, romance, mystery, and horror, novels through sites such as Amazon’s Kindle Direct Publishing, Apple Books, and Kodo. To generate more income and to keep up with demand from fans, some of these self-published authors are turning to A.I. software, some of it based on OpenAI’s GPT-3 or other large language models, that can automatically compose long passages of mostly coherent text, often in particular style. The author might need to edit this text a bit to make it work or stitch several smaller A.I.-generated passages together, but it’s easier than starting from a blank screen. Yet the technology raises all kinds of questions over the meaning of authorship and creativity and art. Joanna Penn, who the Verge identifies as “an independent novelist and one of the most outspoken proponents of AI writing,” says that she has been subjected to a lot of criticism from those who see the A.I. software as a kind of cheating that devalues writing. But she says:

The reality…is that AI is advancing regardless of whether novelists want it to, and they can choose to use it or be left behind. Right now, she uses Sudowrite as what she calls an “extended thesaurus.” (There are only so many ways to describe a crypt, she said.) But she foresees a future where writers are more akin to “creative directors,” giving AI high-level instruction and refining its output. She imagines fine-tuning a model on her own work or entering into a consortium of other authors in her genre and licensing out their model to other writers. AI is already being used in photography and music, she said. “Writing is possibly the last art form to be disrupted because it’s so traditional.”