AI Runs My Back Office

When most people hear the word "AI," they picture a chatbot. You ask a question, it answers. That is genuinely useful, but it is not the part that changed how I work.

The part that changed how I work is the unglamorous part. The back office. The invoices, the follow ups, the inbox, the daily task lists, and the reports that nobody enjoys building and everybody depends on. So I made a deliberate choice. Any process that is a high latency loop, meaning slow, repetitive, and predictable, gets handed to a machine. Not eventually. Now. What follows is an honest account of what that looks like once you build it for real.

The hardware: a command center that never sleeps

Everything runs on three Macs that communicate over a private network.

At the center is a MacBook Pro that lives in Florida and never powers down. I call it the Command Center. It hosts a web application I built, a local AI model, a vector database, the three assistants I will describe shortly, and every scheduled job I rely on. If a task must run every single day at six in the morning, whether I am awake or not, it runs here.

The second machine is the MacBook Air I travel with. It is the steering wheel. I move between Florida and Indiana constantly, and I can operate the entire system from wherever I happen to be sitting, whether that is an airport, a coffee shop, or a hospital in Indianapolis.

The third is a stationary iMac. It holds backups and a complete copy of my code, so that nothing of consequence ever lives in a single location.

The element that makes the arrangement feel almost effortless is a mesh network called Tailscale. It binds all three machines into a single private tailnet, where they locate one another by name rather than by physical address. When I edit something on my laptop in Indiana and trigger a deployment, the code travels to the backup machine, then to the Florida server, and the server restarts itself. I am updating a computer twelve hundred miles away as if it sat on the desk beside me. I have done it many times, from many places, and it has never complained.

The system also observes real backup discipline: a local Time Machine drive, an encrypted offsite copy in the cloud, and a nightly job that mirrors everything to the Indiana machine at two in the morning. I learned the hard way that a backup you do not verify is not truly a backup. So the system now verifies itself.

The three assistants: a staff, not a chatbot

This is the part I find most interesting. I did not build a single assistant. I built three, and each holds a distinct role.

Jeeves is the assistant. He operates through my Telegram, and I message him the way I would address a chief of staff: look this up, retrieve that figure, tell me what is on my calendar, prepare a draft. He has web access and can read my files, so he is never guessing. He works from the actual record.

Florence is the strategist. When the task calls for thinking rather than fetching, that is her domain. She helps me plan the week, stress test an idea, and map out a build. She is the one I debate before I commit to a decision.

Benson is the operator. He carries the authority to act. When something genuinely needs to be executed across my files, Benson does the work while the other two reason and advise.

An assistant, a strategist, and an operator. They run around the clock. They recover on their own when the network falters. And none of them may send an email, publish anything, spend a dollar, or delete a file without my explicit approval. That restriction is not a shortcoming. It is the entire design. The machines handle the preparation. I retain the decisions.

What the system actually runs on its own

This is the portion people tend to doubt until they see it, so let me walk through the real back office.

Mornings. Each weekday at 7:15, a brief arrives in my inbox: my calendar, the weather wherever I am, the urgent mail that genuinely requires me, my open tasks, and the plan for the day. I also receive a short personal newspaper every morning, and on Sundays it includes a shopping section, a small touch I will admit I enjoy.

The inbox. Newsletters and automated replies are filed every fifteen minutes, so the clutter never accumulates. A sweep runs nightly across all of my accounts and surfaces the few items that actually warrant attention. I stopped drowning in email, and I never hired anyone to do it.

Invoicing. This is the process that convinced me. On the operations side of the healthcare company, monthly invoicing once consumed the better part of a day: pull the log, format each client differently, assemble the PDFs, verify the figures, and repeat whenever something did not reconcile. Roughly eight hours. Today I paste in the monthly log, and the system returns finished, submission ready invoices for every client, each in that client's required format, with the math already checked. Ten minutes. Eight hours down to ten, every month, indefinitely. That is not a productivity tip. That is a week of my year returned to me.

The relationship engine. Planned giving is relationships at scale: hundreds of people, each owed the right contact at the right moment. So I built an outreach engine. It schedules calls, emails, and messages on a rotation, honors a daily limit so that no single day is overloaded, and carries forward anything I miss rather than letting it lapse. At three in the morning it produces a printable sheet listing precisely whom I need to reach that day and by what means. A "top three" indicates where to begin when the list runs long. There is even a severe weather mode that, when a storm strikes a region, surfaces the people who live there so that I can check on them first. Underneath all of it, a nightly synchronization ensures the information I work from is never a day out of date.

The financial view. Bills and statements reconcile themselves. I submit a PDF statement, and the system matches the charges, flags the recurring items, and isolates anything new. Revenue, investments, and goals all appear on a single dashboard I can open on my phone.

The ventures. The site you are reading stages my notes and drafts my newsletter each week. My other ventures carry their own automated infrastructure: a small collection of affiliate publications, a growing catalog of applications, and a performance scrubs label I recently brought to market, conceived, branded, and launched with the same toolkit. Each one runs on plumbing the system maintains for me.

I could continue. An overnight process reviews everything I touched during the day and synthesizes it while I sleep. Another monitors the health of my drives. Another exists solely to confirm that the rest are functioning. None of these is the point on its own. The point is that the tedious middle of running several operations has been delegated, one loop at a time, until what is left on my desk is the work that actually needs a person.

The rule I will not break

None of this functions without one firm principle. The AI prepares. I decide.

Nothing is sent autonomously. No email, no post, no payment, and no deletion, ever, without my reviewing it and granting approval. The machines are fast, tireless, and occasionally wrong. In donor work or medical billing, the cost of one wrong autonomous action is far higher than any efficiency it would buy. So I keep the steering wheel.

As it happens, this is also the most productive configuration. The slow part of my work was never the typing. It was the gathering, the formatting, the remembering, the preparation. Once that is handed to the machine, what remains is the judgment, which is the part I want to be doing in the first place.

Credentials live in a secure vault, never exposed and never typed by a bot. That principle is not optional either. You do not earn the right to automate your entire life and then grow careless with its keys.

What holds it all together

Every one of these AI sessions begins from the same foundation. I keep a single document that is the front door to my entire world: who I am, what I am working on, where everything is kept, the operating rules, and the open threads. Each night, the system automatically refreshes the "where things stand today" portion.

So when I open a new session, it does not begin cold. It reads the front door and already understands me. It knows the deals in motion, the machines, and the requests I have made repeatedly. I do not re explain my life each day. That single habit, recording everything in one place and keeping it current, is very likely the highest leverage practice in the entire system. An assistant is only as capable as the context it is given, so I made the context excellent and its upkeep automatic.

It breaks. That is expected. I reinforce it.

I want to be candid, because the polished version of this story is seductive, but it does break.

Last night, I learned there were bad storms in Indiana, and the backup machine had been knocked offline by a power outage. The week before, the system nearly lost a piece of work because one machine was holding changes that the others were unaware of. Things drift. Networks drop. Drives fail.

The distinction is in the response. Each time something breaks, I do not merely patch it; I resolve the underlying cause of the problem. When work was nearly lost, I built a safeguard that refuses to deploy when the machines are out of sync, along with a tool to recover stranded work. It proved itself on its very first run. When a backup drive failed quietly for six days before I noticed, that became the basis for a new monitor. Permanent solutions over temporary patches, every time. The system grows more resilient with each failure rather than more fragile.

Periodically, I audit the entire thing. Last night, tired as I was, I directed a small team of agents to examine every layer of my own system: whether the code was sound, the jobs were running, the data was consistent, and the outage had done any harm. The report came back clean. There is real satisfaction in pointing the machines at the machines and getting back a clean bill of health.

Why I do this

I am one person. I manage a portfolio of clinical trials, oversee operations for a healthcare company, hold a position in planned giving and nonprofit fund development, manage several personal ventures, and I am working toward another graduate degree. On paper, none of that should be possible without a staff.

The machines are the staff. Not because they are intelligent enough to replace judgment, but because they are tireless enough to absorb the loops. The slow, repetitive, predictable middle of every role is precisely what they do well, and it is precisely the part that was wearing me down.

That is the wager: go AI first in my own businesses before it is the obvious move. Eliminate the high latency loops. Keep the judgment. Keep the keys. And build the system so that it grows stronger each time it cracks.

The back office runs itself now. I am free to spend my time on the work I genuinely care about.

MG