There are no two-week development sprints at Orel Zilberman's WriteStack any more. He tells an AI agent what he wants to build, and it coordinates the coding and prepares emails for beta testers. He runs the software business on his own.
During our Millennial Masters Live, Philip Hofmacher admitted he left one of his automations 80% finished for four weeks.
I recognised the problem immediately. I've got around eight projects in much the same state, often because I start something else before getting through the awkward last bit. I didn't think much about the unfinished backlog when I wrote about the new one-person company in June. 👇🏻
New here? Subscribe to Millennial Masters for weekly reporting and conversations on the decisions behind building a business 📧
How much can one person run?
Behind Orel's setup are roughly 13 or 14 specialist bots, all reporting to one he calls Jarvis. He estimates that a feature he shipped in a week might previously have needed five developers working for a year.
“I just come up with an idea and I can test it an hour later and see how it works.”
Philip runs Write • Build • Scale with two co-founders and two employees, one of whom was part-time when we spoke. Their 11 or 12 bots handle specific jobs, including a daily check of the CRM for recent sales activity and conversations needing attention.
He doesn't think that amounts to a dozen virtual employees. The CRM check would take about an hour a day, nowhere near enough work to justify another hire.
“I prefer to have an AI do this one hour a day task for me rather than doing it myself or paying someone to do that.”
When building becomes too easy
After building two tech businesses, Cory Blumenfeld still caught himself making an old product mistake at BlueMoso. He began planning a portal around the functions it could have, then realised he hadn't properly established which problem customers needed it to solve.
“We're building before we actually truly understand the pain point that we're building for, right?”
With AI doing the coding, a new feature costs Orel a five-minute prompt and about an hour of checking, by his estimate. He also tested an AI coach with around 20 people, expanding the trial to 30 when others asked to join. Other experiments never reached general release.
“But testing it is so cheap that I don't see a reason why not.”
An agent working on WriteStack's landing page showed Orel the commercial upside. Visitors saw a button labelled “Start free”, then found they had to choose a plan and enter card details. Some abandoned the process. After he changed the button to “Start your free trial”, Orel reported a 10% rise in conversions. He still approves proposed changes before they go live.
Philip questioned where Orel's feature ideas were coming from. Some came from customers, while others were Orel's own.
The problem with 80% done
Philip picked up the 80/0 rule at a recent mastermind.
“If something is eighty percent done, it provides zero value.”
Philip left the final setup of his Threads reply bot unfinished before travelling. When we spoke, the bot still wasn't doing the job he built it for.
I have the same habit of leaving projects at the point where they need a final test or another round of feedback. Once something else takes my attention, I struggle to return to those open problems.
Cory recommended parking ideas on a list and revisiting them a few days later, though he no longer did it consistently himself. He paused work on his portal because he wasn't convinced it solved a clear customer problem.
Having more ideas became part of the problem. Philip thought “the most important skill nowadays is spotting the right opportunities rather than spotting ideas for growth.” Orel told Jarvis to file away new features and point him back to whatever he was supposed to be working on.
Someone still needs to know what's going on
Philip learnt his usual reply and close rates by doing sales work himself. He wants those figures before delegating the process to an agent, so he has something to compare the results against.
“I would not be able to judge if the AI is doing a good job or not.”
Even Orel said AI now writes so much of his code that he doesn't know 50 or 60% of it in detail. His development experience means he can still investigate a fault. Cory's unfinished portal raised a more worrying possibility: he didn't know enough to check whether the system stored customer data securely. Without that expertise, “you're just letting the AI pick for you”.
At Millennial Masters, I spent more than a month setting up the agents that handle sponsorship requests and editorial follow-ups. Much of that time went into writing the operating rules and correcting cases I didn't anticipate. The agents now draft replies that stay in my inbox until I review them.
All that setup makes the time saving difficult to calculate. I haven't measured how many hours my agents save after I count the corrections, and Philip couldn't give a figure for his business either. Some of his bots produce reports that nobody prepared before.
What still needs a human
Cory is building an agent to follow up overdue invoices and plans to have a virtual assistant oversee the system. That should free up the assistant to deal with conversations that need a person's attention.
“I refuse to hand over the customer relationships.”
The suggestion that an AI caller could chase people who missed his webinars made Philip recoil. He saw it as an intrusion into the trust he built with his community.
“I do not want an AI to call my people.”
His Substack welcome message is automated too. New readers are asked why they subscribed, and Philip or his VA takes over when they reply. He estimated his automated welcome message went out to 50 to 100 new readers a day. Orel automates welcomes as well, while refusing to turn comments into bot-written exchanges.
At that volume, Orel wondered whether a trained AI could do the same work as Philip's assistant: “what's the difference in this case between a VA and an AI that knows how you write?”
Philip tried AI-assisted replies before. In informal comparisons, he found his human VA's replies worked better for his readers. He wasn't ruling out a sufficiently capable AI, although training one to handle all those exchanges was more work than he wanted: “I don't have the time nor the expertise to train the AI to get it to a point where it really responds exactly the way that I would.”
Philip wished he'd introduced an AI check earlier to question whether a call or new project deserved his time before he committed.
I've already got an AI strategist helping me make decisions like that. The eight unfinished projects on my own list are still waiting for an answer.
