Day 1 of building a marine F&I brokerage with an AI stack — March 25, 2026 — had no AI in it at all. Just a Google Doc, the names of three Texas marine dealers I wanted to email, and a half-page draft for each. I sent zero of them that day.
For context: I came out of the dealership world to start Lakeside Lender — marine F&I brokerage and AI consulting practice, Conroe, Texas. No team. No clients. The plan was build the brokerage AND the AI stack at the same time, because the AI stack was going to be how a one-person operation could serve dealers at the same caliber as a finance department at a multi-store group.
But Day 1 was three half-page drafts and no AI. Then six days of nothing on paper. Life, work, dog walks on the Lone Star Trail. The build was already showing me its shape: weeks where you ship a lot, weeks where you don’t. Both count.
On March 31 I set up the first thing that resembled a system: a daily scheduled task in Claude. Eight AM. Scope: scan MRAA and NMMA directories, business filings, social signals, trade press — anything that surfaces a new marine dealer in Texas. Output to a dated markdown file. No exceptions.
I considered other paths first. A virtual assistant on Upwork around $600 a month — would never know boats well enough to read between the lines on a dealership’s social feed. A spreadsheet I’d update manually — I’d have quit by week 2. Google Alerts — too narrow, can’t synthesize across sources, doesn’t tell you what the signal means. The scheduled task was the only option that scaled without me sitting in front of it.
That was the first real lesson of the build, and the one that’s shaped everything since: automate the recurring thing before the one-time thing. The daily-briefing task from March 31 is still running today. Everything around it has been torn down and rebuilt twice — different LLM tier-routing, different output paths, different storage backend — but the briefing itself has run every weekday morning since the day I set it up.
The reason this lesson matters isn’t just operational efficiency, though that’s real. It’s that the recurring thing compounds. Every morning the daily briefing fires, my picture of the Texas marine dealer landscape gets sharper. Every weekday the competitor-watch fires, I see a market that other consultants are guessing about. After two months of that, the asymmetry is real — I’m not smarter than the next consultant, I just have a system that’s been ingesting the same beat for two months while they were building landing pages.
The one-time things — the proposal documents, the client onboarding flows, the email campaigns — those get built when a client engagement demands them. You can absorb a one-time spike of work because it has an end date. The recurring thing has no end date, so if you’re doing it manually, every day it’s running is a day stolen from building.
Day 1 had no AI in it because I didn’t yet know what the AI was for. By day 7, I knew: the AI’s job is to do the things that have to happen every day, whether I’m focused on them or not, so that I can be focused on the things that only happen once.
That’s the shape of the build. The next post covers April 1 through April 3, when the second piece of recurring infrastructure went in — a daily PDF generator for my business partner, built before any paying client existed.