For about a year, Sam Fox-Hartin had worked for an on-demand concierge startup called GoButler as a “Hero,” the company’s term for employees who field users’ requests, via text message, and then complete tasks such as booking tables at restaurants, scheduling appointments, or ordering food for delivery on their behalf. Most of these tasks, like the ones I watched Fox-Hartin maneuver when GoButler invited me to visit its New York headquarters last year, were fairly routine. But he also wrote poems. Convinced couriers to deliver dry ice. And in response to one particularly odd request, drew “some horses hanging around a campfire.” As a comedy writer and former philosophy major, he brought creative enthusiasm to the job. “I’m by no means a gifted visual artist,” he says of the horse drawing, “but I tried my best.”
In late February, he learned he was about to be replaced by an algorithm that he’d unwittingly helped build.
GoButler’s CFO called Fox-Hartin at home to tell him the news (he wasn’t at work on the day the announcement was made). The company had decided to move to “a more automated product,” he said. Instead of offering an “everything and anything” service, GoButler would reinvent itself as a fully automated discovery and booking tool. The first category would be air travel, and Priceline would pay GoButler a fee for every ticket it sold. GoButler no longer needed heroes, or Fox-Hartin.
Nowhere is the potential for job automation so obvious as it is in the on-demand economy, where many startups have grown fat with venture capital despite poor unit-economics. Uber is spending heavily to hasten the development of driverless cars. Instacart, Postmates, and other delivery-heavy startups are unlikely to stick with humans once machines–which don’t take sick days, need bathroom breaks, or threaten to unionize–can do the same jobs.
But even if you don’t work in the on-demand economy, chances are high that you or someone you know will eventually be in the same position as Fox-Hartin. Machines already exist that can flip burgers and prepare salads, learn and perform warehouse tasks, and check guests into hotels. Companies like WorkFusion offer software that observes and eventually automates repetitive tasks done by human workers. And automation has also crept into knowledge-based professions like law and reporting. When in 2013 researchers at Oxford assessed whether 702 different occupations could be computerized, they concluded that 47% of U.S. employment was at risk of being lost to machines.
Fox-Hartin knew, tacitly at least, that being replaced by a robot was a looming possibility. But somehow, when he had long conversations with GoButler engineers about natural language processing technology, their work didn’t seem like an immediate threat to his job. In early tests, the automated features often categorized requests incorrectly. And none of the features could automate a unique poem or a horse drawing.
The heroes of GoButler treated the possibility of their robot successors as a joke. “You make jokes about brain aneurysms because brain aneurysms are really frightening,” Fox-Hartin says. “It was sort of at that level. The sum total of it all never really appeared to be an immediate cause for alarm.”
When Navid Hadzaad, GoButler’s CEO, started the company, he didn’t know whether it would be a fully automated or partially automated product. In either case, Fox-Hartin’s job would probably have been toast: The human portion of a hero’s job in a partially automated concierge service would have likely been outsourced to a country with a lower cost of living than New York City.
GoButler had recruited its first human concierge agents in expensive cities like New York and Berlin was because these workers were, in a way, part of the same team as GoButler’s engineers.
Just as much as he was hired to interact with clients, Fox-Hartin was hired to “train” GoButler’s algorithm. As he responded to differently worded inquiries and narrowed requests to their most vital information (How many tacos? What kind? Where is your delivery address?), GoButler’s algorithms “learned” from his patterns. Every time he interacted with a customer or categorized a conversation, the artificial intelligence got better. “The better the people we had to create messages, the better our product would be,” Hadzaad says. “Data was the heart of our product. So for us, the people we had weren’t just customer service agents. They were part of building the product.”
As GoButler’s AI improved, Hadzaad faced a choice. He could continue providing the type of bespoke, human-layer service that kept Fox-Hartin busy writing poetry and hunting down dry ice, which would mean hiring people in the Philippines and charging for a service that had previously been free. Magic, a GoButler competitor with a $100 per hour premium option, relies heavily on automation, but it has chosen to keep humans in the loop so it can offer a more comprehensive, personal service. Or, Hadzaad could focus instead solely on automated, commodity tasks like ordering food. He went with the automated free service, believing that it would be easier to create a product that people would incorporate into their daily lives. “My general view is that people will always want convenience, but they’re not willing to pay premiums for it,” he says. “And then when you think about the human labor behind it, I think that will become a luxury more and more.”
GoButler’s new, automated service launched on Thursday. It looks more or less the same as it did when humans were involved—just a message thread—but now the responses and follow-up questions are automated. A typical conversation, as demonstrated in screenshots on its app download page, proceeds as follows:
User: “Hey! I need a nonstop flight from NYC to London June 18-24.”
GoButler’s algorithm: “Got it! Looking for flights from JFK to LHR on 2016-06-18…returning 2016-06-24”
Efficient. Practical. Rather obviously automated.
Fox-Hartin, who plans to spend the next couple of weeks watching college basketball as he searches for a job as a copywriter, is surprisingly gracious toward his replacement. As much as he loved his job, there were times when fulfilling a user’s request could be difficult. “There’s no frustration in an algorithm,” he says. “It just tries the next thing. You take out the human stress from it, and it probably becomes a better service.“
Still, he remembers, whenever his team of “heroes” called an airline’s support number on behalf of a customer, they would attempt to surpass the automated menu as quickly as possible. “We’d pray for the moment,” he says, “when a human would pick up the phone.”