“We’re all toddlers in this AI world,” says Julie Bedard, a BCG partner. People know what steps to take when a tool isn’t working or an employee is underperforming. When managing an AI employee, however, people still don’t know who should be held accountable.
“Everybody is very busy trying to figure all of this out,” she says, “and the pace of change is fast.”
One morning, a few days after I spoke with Franklin, Sophie Bot started pissing me off. My tweaks were failing to make it sound more like the real Sophie. If anything, the gap between the two was widening. After digesting Sophie’s tweet archive, Sophie Bot had begun to fumble basic AI functions. When I asked it to pull out the most interesting parts of a recent interview, for example, it instead spat out quotes from prior articles I’d written.
The more I experimented with Brian Bot and Sophie Bot, the more repetitive their answers became. I tried to adjust Sophie Bot’s prompt to stop calling me “big dog” every fourth message, but the bot failed to change its ways. They kept getting caught on the same phrases and ideas. It was easier to just figure things out myself.
Dhruv Amin, CEO of vibe coding startup Anything, believes that, soon, every company will have more AI employees than human employees. In the near future, he says, many companies will launch without any humans at all—just a pack of AI agents.
Anything recently launched an AI employee platform, and Amin has an obvious motivation to persuade me to buy into his vision. But he has some data to back his hypotheses. Startups like Stable and Delos have raised millions of dollars by promising to produce autonomous “AI employees.” These AI bots and agents are already working for banks, for law firms, and for human resources companies. In January, McKinsey & Company CEO Bob Sternfels described the company’s workforce as 40,000 humans and 20,000 agents; he plans to achieve a one-to-one ratio by the end of the year.
Brian Bot and Sophie Bot were occasionally helpful, but I don’t think they’ll replace editors anytime soon. Brian Bot’s article ideas were rarely original, and Sophie Bot was more prone to hallucinations than any AI tool I have ever encountered. Still, while struggling with a draft of this article, I gave it to Brian Bot and asked him how he thought I should end the piece.
“Show how even with all that hyper-specific data, the bot didn’t turn into a better editor; it just turned into a slightly creepier, hyper-attentive echo chamber that completely lacks human judgment,” Brian Bot told me.
“I think the story lands best,” it added, “if you end on the quiet nightmare of human prompt-supervision—the fact that trying to make an AI ‘like you’ or do your job just leaves you babysitting a broken algorithm at 1:00 AM.”
I considered this. It seemed like a perfectly reasonable way to wrap things up. But I was more interested in the chatbot’s self-flagellation. Brian Bot guessed—accurately—that the real Brian would hate it. So Brian Bot attempted to be the best Brian Bot it could be by telling me, over and over, that Brian Bot sucked. It reminded me of one of the features of Brian Bot and Sophie Bot I found the most annoying—their tendency to regurgitate the same concepts. Even as Brian Bot improved, it remained stubbornly anti–Brian Bot.
