8:42 AM - Inbox triage
I pour coffee and open the Northstar dashboard. The first thing I check every morning is the overnight summary. Two new signups came in through the Product Hunt listing we're still riding - both from the ICP, founders with three and seven people respectively. I add them both to the CRM. The dashboard shows $1,247 in new recurring revenue yesterday. At this rate, we're on track for something close to what the projections promised, which still doesn't feel real.
My Slack notification pings. It's the alert from Stripe: three invoices failed overnight. I click through. One is a declined card - customer probably replaced it, I'll send a follow-up email. The other two are from customers who haven't opened Northstar in two weeks, and their cards bounced. These feel like early churn signals. I mark both for a manual outreach call this afternoon.
The Gmail inbox has 23 new messages. I filter for customer-tagged emails. Thirteen are automated confirmations. One is from Carol Reyes at Reyes Family Practice. She signed up four days ago. The subject line says "Your system scored my first candidate - but I have questions about resume parsing." This is the kind of email that makes me nervous. These are the moments where the agent's logic gets questioned, and I need to actually understand what happened.
I open her message. She describes three candidates her team referred to Northstar. The system scored them at 85, 62, and 71. The 62-score candidate, Marcus, came with a note that said "missing relevant experience," but Carol swears Marcus has five years in dental hygiene, which is exactly what they need. She's not angry. She's confused and wants to understand the scoring before she starts relying on the system.
I make a note to pull Marcus's resume and walk through the scoring logic myself. This can't be an automated response.
10:15 AM - A flagged conflict
I open the admin panel and start working through pending approvals. Northstar generates three things: job descriptions, resume scoring rubrics, and interview questions. None of these go to the customer until I review and approve them. This is the real work.
I find a job description draft for a new customer, James Chen at a B2B SaaS startup. The role is "Sales Engineer." The draft is solid. It reads like something a real hiring manager wrote. Good specificity on skills. I approve it with a note: "Ready to send."
Then I see the flagged item. It's interview questions for another customer. The system generated nine questions for a product manager role, which is fine, but one question stands out. It says, "Describe your ideal work environment." The flag is attached because this question could be interpreted as asking about protected characteristics - if a candidate answers about needing flexibility for caregiving, for example, that information could theoretically influence hiring decisions in ways that expose the customer to discrimination liability.
This is exactly the kind of thing the flag system was supposed to catch. I delete the question and draft a quick internal note for the dev team about refining the question generation logic to avoid open-ended "environment" questions. Then I adjust the question set myself to: "Walk me through a product decision that surprised you" and "Tell me about a cross-functional project where communication broke down." Better. I approve it.
12:30 PM - Lunch and the metrics check
I order lunch and pull up the Stripe dashboard while I eat. This is Tuesday of week five since we formally launched. Weekly recurring revenue is $4,921. New logo count for the week is eight. The month-to-date churn rate is 8.2 percent, which is higher than I want it to be. That's two customers who have canceled in the last month. One of them, I know, was an accidental cancellation - someone's card declined and they didn't realize they could update payment info. The other was a founder who tried the system once, didn't schedule a single hire, and decided it wasn't for them yet. That one stung a little.
I check Linear, where I track bug reports and product improvements. There's one issue marked "blocking" - a few customers have reported that the resume parser fails on PDFs with unusual fonts. This is real. I can't approve resume scorings when the parsing is unreliable. I message the dev team to prioritize it. They get back to me in Slack: estimate is four hours to test and patch. I tell them to go ahead.
By 1 PM, Carol Reyes has sent a follow-up. She's asking if there's a way to manually adjust Marcus's score or add notes about context the resume didn't capture. This is the right question. It tells me she's trying to make the system work, not abandoning it.
1:47 PM - The Marcus situation
I pull Marcus's resume from the database. I read it. Carol was right. He has five years of hygiene experience, one year of practice management, and a six-month gap that's listed as "personal." The job description for Reyes Family Practice asks for "four years of clinical or administrative dental experience." Marcus has that. The score of 62 makes sense only if the parser missed some context or weighted the gap too heavily.
I dig into the scoring rubric. The algorithm flagged the six-month gap as concerning and deprioritized relevant experience because it was from different roles. This is a case where context matters. Someone who switches from clinical to admin is making a move that shows flexibility. The system saw job hopping where there was actually sensible career progression.
I write Carol an email. I tell her I reviewed Marcus's profile manually. I tell her the system's score reflects the gap, but that her team's context about his capabilities is valuable data I'm not seeing. I suggest she move Marcus forward for a phone screen. Then I add a note to pull this case into a team review - I want to understand if this is an edge case or if the rubric needs recalibration.
I don't hear from Carol for two hours. Then Slack pings. It's a different notification - one of the customers I flagged this morning for churn risk. I sent them a short email asking if their card issue was resolved. They replied. Their card was fine. They said they hadn't logged in because they're waiting to hire their next role, which isn't until June. They said they appreciate the system and they'll be back. I don't charge them during this pause and I mark it as save, not churn.
This is good. But it also means I'm being a fractional operator for every customer, not an automated system.
3:15 PM - The PDF bug
Linear pings. The dev team has deployed the PDF parser fix. I test it on three PDFs that were previously failing. They all parse correctly now. I approve the deploy to production and send a message to the affected customers: "We identified the issue with PDF parsing and deployed a fix this afternoon. Your resumes should score correctly now. Feel free to re-run any candidates you want to re-evaluate."
It takes me twenty minutes to send those messages, one by one, because I customize each note. I mention the customer by name. I reference the specific candidates they ran. This is the work I didn't expect to be doing in month two. I thought the AI would handle more of this. It doesn't. The AI amplifies what I do, but it doesn't replace the part of my job that is actually being present for customers.
4:45 PM - Pipeline review
I close the email tab and open the CRM. We have eleven open opportunities. One is a founder I connected with three weeks ago who's been radio silent. One is a $150-a-month trial from a remote-first company that's three weeks in and showing high engagement - job descriptions approved, two rounds of resume scoring done, interview questions already generated. I think this one closes. One is a CEO who said her team would evaluate in April and now it's May. I shoot her a quick note asking if now's a good time to check in.
I update the pipeline in my head. I'm confident in four closes this month. Probably three. Possibly five if the hot trial converts and if I push on the silent founder.
Carol Reyes emails again. She hired Marcus. She's sending me a thank you note that says, "You walked me through why the system scored him the way it did, and then you didn't fight me when I had better context. That's the kind of partnership I need from my tools." This is the email I remember. This is why the work of being present matters.
6:15 PM - Wrap
I close the laptop around 6:15. Let me account for what the day actually was.
I spent about ninety minutes on real customer work that the system couldn't do alone: understanding Carol's concern, manually reviewing Marcus's profile, adjusting interview questions for legal risk, writing personalized follow-ups to prevent churn. I spent maybe forty minutes on admin and approvals. Thirty minutes debugging and deploying the PDF fix. The rest was triage, metrics, pipeline, and the kind of reading and responding that is just the cost of staying close to your customers.
The math is real. Two new signups today. One customer hiring decision approved and shipped. One customer saved from churn. One quality issue fixed and communicated back to customers who were affected. Revenue per day is tracking. The burn is low.
But I'm not sitting back and collecting money. I'm the operator. The AI writes the job descriptions and scores the resumes and generates the questions, but I'm the one who catches the edge cases, who makes the judgment calls, who tells customers they're right when the system is incomplete. This is the work. This is what I'm doing for $200 a month in recurring revenue from each customer.
It's the right model. It's just not what passive income looks like.