8:42 AM - Inbox Triage
I open the Marigold dashboard at my kitchen table with a fresh coffee. The numbers from yesterday are already there: 7 new signups, $127 in subscription revenue. The signup funnel looks clean. I've learned that Tuesday mornings are when parents get their billing statements and think about what to do differently this month.
I flip to Gmail. There are 23 new messages since I went to bed. Most of them are from Marigold - the system sends me a daily digest of tutor-generated responses to parent inquiries. I need to review these before they go out. This is the part that surprised me most when I bought the engine: the AI does probably 85% of the communication work, but I read almost every customer-facing message before it's sent. It's not because I don't trust the system. It's because I'm the human on the hook if something goes wrong.
The first draft is from Elena's tutor. A parent named Carol Reyes emailed yesterday asking if Marigold would help her daughter with her IB Biology coursework. Elena's tutor has drafted a response that explains how well the model adapts to advanced biology prep, mentions that we support independent study, and offers a specific starting point: "We can begin with photosynthesis breakdown, since that's a common bottleneck in first-year IB." It's good. I flag it green and hit send. Carol will see this in her inbox within the hour.
I work through the stack. Most of them are good. One, from a parent asking about our refund policy, feels too formal. The draft talks about "contractual terms" and mentions "force majeure." This tutor needs an adjustment. I make a note: tutor style drifting into legal language. I'll batch these tweaks into the system prompt this afternoon.
Then I hit one that stops me. A parent named Marcus Webb has written to ask if Marigold's AI has ever made a mistake in explaining something. He says his son came home confused about mitochondria and had to be corrected by his teacher. The draft response from the tutor is technically solid, but I know what Marcus is really asking. He's scared. He wants to know if he's throwing $17 a month at something that might make his kid worse at science.
I click reply. I write Marcus directly, skipping the AI draft. "Hi Marcus, thanks for flagging this. The mitochondria explanation may have been simplified for his level - that's actually intentional for retention - but I want to check our interaction log for your son's session. Can I get back to you with exactly what was said. I want to make sure our tutor is calibrated right for him. I'll follow up this afternoon." I send it. Marcus gets a real human response within five minutes of his concern.
10:15 AM - A Flagged Conflict
My Slack channel for system alerts pings. Red dot. It's a note from the Marigold backend: "Billing conflict detected - Student active in two concurrent subscriptions." The system is smart enough to flag it but not smart enough to fix it.
I search the admin panel for the offending account. It's a user named Jasmine Okonkwo. She has a $15 monthly plan that started three weeks ago. But there's a second one - a $19 yearly plan that auto-renewed yesterday. I can see her tutor activity: she's been consistently using the service about 20 minutes a day, seems engaged, making progress in math.
This is where the real work lives. I pull up her Stripe dashboard. The transaction happened. Looking at the timestamps, she probably thought her free trial had ended and signed up for the monthly plan, but the system still had the yearly plan in the queue from a promotional email two weeks ago.
I make the fix. I refund the overlapping day on the yearly plan. The backend can't do this without explicit permission. I send Jasmine an email: "Hi Jasmine, we caught a double charge on your account and have refunded the overlap. You're all set on your $15 monthly plan. Your tutor will still have all your learning data. Let me know if you see anything weird." Professional, direct.
Slack pings back with confirmation that the charge was reversed in Stripe. One less thing.
12:30 PM - Lunch and the Metrics Check
I take lunch at my desk. I've learned not to step away from the laptop for more than an hour because things move fast in month two. The business is small enough that I can see everything.
I open the metrics dashboard. Monday-Tuesday this week shows 14 signups. Last week's same two days: 11. The Facebook ad spend is up marginally, and the attribution is tracking cleanly. Week-to-date signups are at 42. That's ahead of last week's 36.
Revenue week-to-date: $3,246. That's 182 active subscriptions times an average of $17.85 per month, divided down. Month one, I was at $1,800 week-to-date. Month two is tracking 80% ahead.
But I notice something. Churn this week is at 8%. Last week was 6%. I flag the cohort analysis. The churn is concentrated in cohorts that signed up in the first 10 days, right when I was still tweaking the tutor prompt. Some kids had a worse experience early. The improvements I made to the prompt are working - seven-day retention on new cohorts is up to 78% - but the early cohorts are still experiencing the old version in their memory.
I make a note to run a re-engagement campaign on the churned users from week one. But I don't do it yet. I want to see if the new cohorts sustain higher retention first.
2:08 PM - Customer Escalation
My email pings. A parent named David Chen has sent an escalation: his daughter has been using Marigold for two weeks, her engagement is great, but the tutor keeps recommending topics that are "too advanced" for the curriculum she's actually in. He's frustrated.
This one I need to think about. The tutor's job is to adapt to the student's learning pace, not to be constrained by the curriculum. But David is right that if the tutor is consistently overshooting the actual course, the student might get confused in class or feel discouraged.
I call David. Not email. Call. It takes me seven minutes to find his number in the system. He picks up on the third ring. I introduce myself as the founder. His tone shifts. He wasn't expecting a human voice at 2 PM on a Tuesday from a subscription service.
We talk for eight minutes. I learn that his daughter is in 7th-grade math, she's strong, but David wants her focused on getting A's in the class she's in, not getting ahead and creating gaps. That's a legitimate parenting choice. I explain that the tutor can be reset to a specific curriculum track. "I don't want her held back," David says. "I just want her learning the right things for right now." I get it.
I walk him through the reset process. I tell him I'll personally watch the next few interactions and make sure the tutor is tracking to the 7th-grade standard he's defined. He's calmer. He says, "Thanks for picking up the phone. I didn't expect that." I tell him I'm glad he called instead of just churning.
I make a note in his account: Curriculum-constrained student. Tutor needs grade-level guardrails.
4:30 PM - Pipeline Review
The Facebook ads dashboard shows me the next tier of data. A lot of searches for "AI homework help" and "math tutor for middle school." The cost per click has stayed stable at $0.87. The click-to-trial conversion is 23%. The trial-to-paid conversion is 14%, which is down from 18% last week.
I drill into the trial experience. What's the first interaction a student has with their tutor. I pull up the email sequence we send trial users. The first email is good. The second, sent on day 3, is explaining features. Maybe it should be explaining results instead. I make a note: test a version of day-3 email that includes a specific example of a Marigold session helping a kid understand something.
I draft the new email, run it past the tutor system to make sure the tone is consistent, and schedule it for next week's trial cohort.
6:15 PM - Wrap
I close the Stripe dashboard. All the customer funds are accounted for. Nothing weird in the ledger. I close the admin panel.
I think about Marcus Webb's concern about mitochondria. I think about David Chen picking up the phone and realizing I care enough to actually be the person on the other end. I think about Jasmine Okonkwo getting a refund and a real explanation instead of a form letter. These moments are the reason I bought this. They're also why I can't let the system run entirely on its own.
The metrics are trending right. The unit economics work. But so much of this is still me - reading tutor drafts, jumping on a call with a concerned parent, catching a billing conflict, tweaking the system prompt based on what I'm seeing. The AI amplifies what I do, it doesn't replace it. I thought that's what I was buying. It turns out it was exactly right.
I close the laptop at 6:31 PM. Same time tomorrow.