Companies adopting AI need to move slowly and not break things
AI is not a computational condiment: You can’t just spread it on an existing product or process and expect improvements, as if you’d asked ChatGPT to summarize some notes.
As the surge of interest in artificial intelligence applications continues unabated since last fall’s launch of OpenAI’s ChatGPT, one can trace it spreading from the tech industry itself to every corner of corporate America. McKinsey’s new State of Organizations 2023 report has identified applied AI as one of the ten major shifts to which companies and leaders need to adapt, revealing that between 50% and 60% of its survey respondents have adopted AI in at least one business unit, with approximately two-thirds sharing that their organizations are likely to increase their investment in AI.
Companies, of course, are not weekend tinkerers fooling around with the latest generative AI development to light up a corner of Twitter. If they’re going to apply AI, they need to do it responsibly and sustainably. They also need to look at how they apply the organic sort of intelligence—the humans who procure, manage, and monitor AI systems. Many people will need to develop new skills; some will need new job titles or roles. “Successful deployment poses many of the same organizational challenges seen with workplace digitization,” the McKinsey report notes. “It requires changes to create a culture that enables the responsible use of AI, as well as to cultivate AI-savvy leadership and talent.”
To see what applied AI looks like in practice in the heat of the AI boom, consider the experiences of two companies—the billion-dollar market cap earth observation company Planet and Intercom, the cloud-based customer experience platform relied upon by more than 25,000 companies. They point to another important factor: a mindfulness of AI’s limits.
Planet: Staying grounded in ground truth
The satellite-imagery firm Planet is no stranger to having more data than humans can readily interpret. Back in 2017, it was taking 1.4 million images of Earth a day, and the San Francisco public benefit corporation (a 2022 Fast Company Most Innovative Companies honoree) is now approaching four million a day.
“We now have something like 50 petabytes of total data,” says Kevin Weil, president of product and business. “We’re well beyond what a human could do with their eyeballs.”
As a result, over the past few years Planet has been exploring ways in which AI could make sense of that data, with help from such partners as Microsoft’s AI For Good Lab.
For example, Microsoft’s Global Renewables Watch tracks the deployment of solar and wind power by using AI to spot their buildout. Earlier this year the startup Synthetaic used an AI image model to map the Chinese spy balloon’s path across the Pacific. And a “Queryable Earth” demo lets people ask plain-language questions about topics like environmental damage and have them answered with data plucked from Planet databases.
There is the risk that AI could serve up “rosy answers,” Weil admits, but Planet has the advantage of having the source material a few clicks away. “With any model that we build,” he says, “we’re validating ground truth at the end of the day.”
In Planet’s view, drawing conclusions from the data remains the role of humans. “Let’s use computers for what they’re good at, which is understanding what’s different, understanding what’s happening. And let’s use humans for what they’re good at, which is the why,” says Weil.
Increasingly powerful AI has not yet led to organizational changes at Planet, though. “We’ve been working on AI for a long time now,” Weil says, citing such longstanding use cases as detecting clouds in imagery and identifying roads. But having “more generalized AI models” has made these capabilities more accessible.
“It’s democratizing the use of AI within the org,” he says. “It’s making it easier for any engineer across the team to try things out.”
That, in turn, has also spurred some self-directed studying, as “Planeteers” work to master these new tools. “I think you’re seeing a lot of people, a lot of engineers, take themselves through that in evenings,” he adds.
Engineers are not the ones newly intrigued by AI’s capabilities.
“Our executive assistant team took a class on ChatGPT for EAs,” Weil mentions. “It is a thing, and apparently it was super eye-opening.”
Intercom expands its experiments with bot banter
Intercom, which is also based in San Francisco and whose tools are used by everyone from Amazon to H&R Block, has a different kind of too-much-data problem: extracting insights from a company’s knowledge base to help its customers solve their problems.
Bots have been part of Intercom’s story “since kind of 2015, 2016,” says Paul Adams, chief product officer. After all, “AI-Powered Customer Service that saves time and money” is the company’s opening pitch on its site. Yet that background and framing still didn’t have the company quite ready for what the arrival of ChatGPT in late 2022 could mean to its business. “This was going to be just a massive, massive transformation,” he recalls.
But Intercom’s first GPT-based tools, launched in January, did not have this conversational AI doing the talking to customers just yet.
Instead, Intercom put AI to work as a sort of amanuensis that can expand a customer-service reps’s shorthand, rephrase a reply, summarize a customer-service interaction for the next rep, or create a help-desk article from support reps’ notes.
After the debut of the more sophisticated GPT-4 language model in March, Intercom released a GPT-4-based bot called Fin that can now handle live customer interactions.
Adams calls GPT-4 “way, way better than 3.5,” saying “it doesn’t hallucinate nearly as much.” But, he adds, it shares a failure point: insufficient data to field its replies. “Knowledge management is actually potentially the biggest barrier,” Adams says. He sees the need to solve this problem as a reason not to fear an AI-driven job apocalypse.
“A lot of companies are already talking about redeploying part of their customer service team,” he says. “There are new types of jobs—knowledge management, knowledge creation.” (Library-science degree holders, take note.) But Intercom, which made Fast Company’s 2022 list of Best Workplaces for Innovators, decided early on to avoid having its chatbots try to imitate people, and it’s sticking to that principle. “It must be really clear when you’re talking to a bot,” Adams says.
This increased investment in AI has led the machine-learning team to double in size, but it has also spawned a new job title: Conversation Designer. “Their job is to design how the bots work,” Adams says. One big responsibility is deciding when Fin ought to hand a customer over to a human rep.
The company has learned that timing that handoff isn’t always obvious. Human customer-service reps can give out incorrect information while bots—at least, GPT-4-based ones, Adams stipulates—can be programmed to stay within knowledge boundaries.
The biggest difference: “Bots don’t have empathy.” That may make the skills of human reps who can empathize with customers through a screen even more valuable, especially with loyal, high-value customers that any company should want to keep around.
Adams’ vision is for AI to do for customer support what robots have done in factories—taking on mundane, boring, or uncreative work—and letting humans solve more interesting problems.
But as with any application of technology, there’s always the risk that it will be wielded instead to cut cost centers down to size. As Planet’s Weil says, “If you’re using AI to optimize something, what exactly is it optimizing for?”
(16)
