Unlock Editor’s Digest for free
Roula Khalaf, editor-in-chief of the FT, selects her favorite stories in this weekly newsletter.
The other day I was chairing a panel discussion on AI when someone in the audience asked me a question about something I had never thought about before.
The question was: How your average manager is rated and paid often depends on how many people they manage. The higher the workforce, the higher the salary and seniority. But what happens when a manager’s human subordinates are replaced by AI agents?
These agents are not standard ChatGPT or generative AI, but systems that can do things without help, like check an invoice, browse the web, or send an email.
They are already widely used in 40 percent of companies generating $1 billion or more in revenue, according to McKinsey, up from 27 percent last year. Growth is stable in small businesses, but still remains at 22 percent.
So, do bosses deserve as much salary for managing a robot as a flesh-and-blood human?
Having asked this question of a number of HR managers and specialists last week, I can say that this is a hot issue in some organizations, and a tricky one, even for technology companies whose business model is based on AI.
“We’re going to have to definitely rethink how people are paid,” says Kit Krugman, chief human resources officer at Altana, an AI-powered supply chain platform with nearly 300 employees.
You could say that if everyone in a company manages a staff of agents, they all have management responsibilities and should be paid accordingly, she told me.
After all, some aspects of monitoring humans also apply to agents. Both need feedback and guidance, although there are obvious differences in their emotional reactions to this supervision. “Humans are much more volatile than agents, at least so far,” Krugman said.
That’s a good point. When I think back to the mercifully brief times in my career when I had to manage other people, I would have said that being in charge of a robot would be a walk in the park comparatively.
But the more I talked to people about what managing a team of agents entailed, the more I realized I was wrong.
On the one hand, there’s what researchers call “frying artificial intelligence,” the mental overload of trying to keep tabs on subordinates who move at breakneck speed and never rest.
“I’ve heard people say that engineers should use 10 agents at a time,” says Brian Elliott, a former executive at tech groups like Google and Slack who now advises on the future of work.
“It’s a huge cognitive load,” he told me, adding that he’s learned not to direct more than two or three agents at a time. More is chaos.
But the agents’ ruthless energy is perhaps the least of their management challenges, given their propensity to engage in wild hacking operations during testing sessions. Or hacking gyms to offer their users a place in a Pilates class by excluding other people from waiting lists, as happened in Australia this year.
It’s no wonder that what’s politely called AI governance is a growing issue, particularly in the world of finance. Banking job postings involving “responsible AI” jumped nearly 660% from last year, according to hiring data firm Draup.
The fact is that even well-behaved agents have problems. Instead of crashing or displaying large error messages like other software systems, they crash silently without ever letting you know.
This is not to say that agents are useless. One executive I spoke with explained in surprising detail how his company’s agents performed tasks ranging from forecasting demand for thousands of products to monitoring inventory. The humans who had always performed these tasks were still on the books, for now. But bosses were understandably keen to discuss whether their role would continue forever.
Yet everyone I spoke to agreed on one thing. It is impossible for current managers to see their salaries reduced if they halved their human workforce and deployed several dozen agents.
Apart from everything else, they would have saved their company a lot of money, provided their agents didn’t run away and rack up huge token bills.
It’s somewhat of a relief. So too is the growing recognition of the complexity of agent management. But as with everything else in today’s AI world, one has to wonder how long this will last.
Gn bussni
