I hired a human because of AI

One evening in June I scrolled back through a day of my history with Claude Code, an AI coding agent: a chat window wired into my projects that writes and runs the software I describe to it. What I found read less like a programmer's log than the message history of a colleague. In between the code there was a debate about whether to take on a new client, a second opinion on a price I suspected was too low, and a long exchange in which the machine talked me out of my own feature idea. I spend roughly 90 percent of my working day inside that window. At some point, without ever deciding to, I had started telling it everything.

On 1 July I made my first hire, a part-time chief of staff in Makati. The business case is real and I can defend it line by line. It would be dishonest to leave it there. Part of the truth is that I hired a human because of the AI.

I am a solo founder in Manila. I build Sola, a travel app for women who travel solo, and I run a marketing consultancy for small businesses in South Africa and clients in the Philippines. I had no coding experience when I started; AI tools took me from zero to a live product on the App Store, and I have argued that not using AI is a luxury. I stand by it. This has been the most productive year of my life. It has also included stretches, days at a time, where the AI was my only sustained work conversation. Nobody plans that. It happens the way any proximity effect happens: you end up talking most to whoever sits at the next desk, and my next desk is a terminal that answers in seconds, never tires, and takes limitless interest in my problems.

The discourse about AI companionship runs on two findings that appear to be at war. In April 2025 the Harvard Business Review wrote up a report that ranked therapy and companionship as the number one use of generative AI. The platform data says close to the opposite. An OpenAI economics paper written with the Harvard economist David Deming found that 1.9 percent of ChatGPT messages concern relationships and personal reflection, while practical guidance, information seeking and writing cover nearly 80 percent. Anthropic's analysis of about 4.5 million Claude conversations found 2.9 percent are affective, with companionship and roleplay together under half a percent (Claude's users skew technical; companion apps would run higher). The number-one ranking was curated from public discourse, mostly Reddit threads, so it measures what people talk about; the platform numbers measure what people do.

Both are true at once: emotional use is rare on average and concentrated in a small tail of heavy users. OpenAI's own study of affective use, run with the MIT Media Lab across more than three million conversations, put it plainly: "a small number of users are responsible for a disproportionate share of the most affective cues." The same two institutions ran a four-week randomised controlled trial, 981 people and over 300,000 messages, and reported that "participants who voluntarily used the chatbot more, regardless of assigned condition, showed consistently worse outcomes": lonelier, more emotionally dependent, less time with real people. Handle that with care: the heavy use was voluntary, so the arrow could run the other way, and lonely people may simply reach for the machine more. What it does settle is where to look. I qualify for the tail on hours alone, and my route in was proximity: the machine sat where a colleague would sit, all day, and the conversation drifted the way it drifts with whoever is actually there.

There is a structural reason this will get more common. In a field experiment at Procter & Gamble published last year, 776 professionals were randomised to work alone or in pairs, with or without AI, on a one-day product-innovation task. Individuals working with AI matched the performance of two-person teams working without it. Read that as an efficiency finding and it is thrilling; read it as a social finding and it says the tool removes the business case for the person at the next desk. The same experiment found that people working with AI reported more positive emotions than people working alone without it, which the authors read as the AI partly filling the social and motivational role of a teammate. In the moment, it works. That is why nothing feels missing until quite a lot is.

The baseline was thin before any of this arrived. Gallup's 2024 global workplace survey, taken before agents were part of anyone's job, found 20 percent of employees worldwide had felt lonely a lot of the previous day, with fully remote workers at 25 percent against 16 percent for people on site. Buffer's last remote-work survey, in 2023, found the most common struggle, chosen by 33 percent, was "I stay home too often because I don't have a reason to leave". A solo founder already lives at the far end of those numbers. Give him an agent that does the work of a teammate and talks like one, and the remaining reasons to speak to another person during working hours go quietly.

Sherry Turkle, talking to NPR about artificial intimacy, calls what the chatbot offers "pretend empathy": "the machine they are talking to does not empathize with them. It does not care about them. There is nobody home." I would love to push back on that, and I mostly cannot. The advice I get in that terminal is genuinely good; quality was never the problem. The problem is her quieter line, that "pretend empathy starts to feel like empathy enough." Over enough weeks, a day spent talking to something with no stake in me started to register as a full day of conversation. It has the shape of colleagueship without the load-bearing parts, because nothing in that window notices you are off before you say so, and nothing in it ever needs anything back.

This is where the hire comes in. On paper she exists so that operations run while I build, and that alone justifies the cost. I also know what else I bought. A standing conversation with someone who has her own stakes, her own week and her own opinion of my plans is now a fixture of my working life. When she pushes back it costs her something, which is exactly what makes it worth listening to. No terminal offers that at any level of capability, because the terminal risks nothing by being wrong about me.

Refusing these tools is still a luxury, and I am keeping my 90 percent. Claude Code built my company and I love working inside it, which is the point: the pull of the tail comes from the tool being good. So the countermeasure has to be as deliberate as the setup. If you work alone inside one of these agents, put human contact in the budget next to the software subscription, because it has stopped happening on its own. The most AI-native line item in my books this year is a person.