Marcus has done the startup arc more times than most people change jobs. (His name, and his assistant's, are changed here at his request.) A founder several times over, he has built companies, seen them through to exits, and is about to start the next one. In the gap between two of them, he did what builders do with free time: he started building.
The thing he chose to build was an AI chief of staff, an agent that could sit between him and the constant stream of messages demanding his attention. He was pulling recent GitHub repos, repurposing them, wiring up the infrastructure himself. Someone close to him suggested he just wait six months and see which products survive. His answer: "that's not fun."
Then, a few weeks into the build, Marcus stopped.
Quick overview
- Marcus is a founder several times over who, between companies, started building his own AI chief of staff to sit between him and a life running through too many channels.
- A few weeks into the build he stopped, deciding that a full-time team would solve the infrastructure faster than he could alone, and bet on a Vellum assistant instead.
- He chose Vellum because it is open source: he could stop owning the plumbing underneath and spend his building energy extending the assistant on top.
- His assistant, Otto, has its own email address and a dedicated card, joins threads he forwards, books appointments online, and learns in public while he keeps the final say.
- Where it is heading: every channel in his life funnels through Otto first, so what reaches Marcus is a short list that actually needs him.
Thousands of people are building the same agent
His life runs through more channels than any one inbox can hold. Several email addresses, phone calls, voicemails, a work messaging app, texts with everyone else. Each one demands its own checking ritual, and together they fragment the two things he actually wants his attention on: building, and his family.
An agent that funnels all of that into one place is a real piece of software. Building one yourself means scaffolding, infrastructure, integrations, and maintenance, and that is before the agent does anything useful. A few weeks in, Marcus ran the math that most builders avoid. As he put it: "I wouldn't say I've given up, but I've accepted the fact that there are enough people now trying to solve this problem that me doing it on my own is not the best use of my time."
His read on the market made the decision for him. There are thousands of people right now gluing together the same chief-of-staff agent from the same repos, each one solo, each one redoing work a dedicated team has already done. Someone whose entire company is this problem will build it better and faster than any one person working alone. The move is to bet on one of them, let them carry the infrastructure, and spend your own building time on the things only you can build.
Marcus placed his bet on a Vellum assistant.
Why he chose Vellum
He found Vellum through a post in his network, recognized everything he had been trying to automate, and signed up fast. He had tried other assistants in the category. What kept him on Vellum, in his words: "no one else is doing it better."
The deciding factor was that Vellum is open source. For a builder, that is the whole game: he could stop owning the plumbing underneath and start building on top instead. The infrastructure a full-time team maintains becomes his foundation, and the skills system above it is open for him to extend, so his technical energy goes into shaping what the assistant does for him rather than keeping the lights on.
Marcus found his solution in Vellum:
- A native Mac app he lives in daily, with web app and iOS alongside it, no scaffolding to assemble
- An assistant with its own email address and identity, so it can work with other people directly
- Same-day start, with the infrastructure, integrations, and maintenance carried by a full-time team
- A skills system that leaves room to build, so a tinkerer extends his assistant instead of rebuilding it
- Credential handling he trusted enough to hand over a credit card
Meet his assistant Otto
Marcus named his assistant Otto. Otto has his own email address, and his main job so far is handling Marcus's email: drafting it, sending it, and replying on Marcus's behalf. The proof that it works is simple. When Otto started emailing the Vellum team directly, the replies came back saying the work spoke for itself.
The way Marcus delegates is by forwarding. When something lands in his inbox, he loops Otto into the thread and tells him what to do, and Otto takes it from there. The people on the other end know they are emailing an AI agent, not Marcus, and he is fine with Otto learning in public. Early mistakes are part of the deal, and the people he works with find the whole thing entertaining.
Otto also handles appointments. Marcus gave him a dedicated payment card and told him to book an appointment online, and Otto signed up and booked it himself. Marcus set this up the way someone who has run companies would: the card is a separate account with fraud protection, the machine Otto runs on never touches anything more sensitive than that, and if a charge ever looks wrong, a phone call to the card company sorts it out.
He has also given Otto access to his Mac, so Otto can see incoming texts and phone calls alongside the email. That single point of access, every channel visible to one assistant, is the foundation for where this is heading.
One funnel for everything
The end state Marcus is building toward is the chief of staff he originally tried to write himself. Every channel in his life flows through Otto first. Every hour or two, Otto pings him: here are the things you need to respond to, tell me what to do. Marcus stops checking anything directly. The noise gets absorbed, the decisions still belong to him, and his attention stays on building and on his family, whether the interruption was from his boss, an investor, or a relative.
The part he is most excited about is the part that sounds least like productivity software. He wants Otto to notice things like "you haven't called a family member in a week, here's a message, want me to send it?" and fire it off from his own texts when he says yes.
What he got back in the meantime is his time. The weeks that were going into scaffolding now go into building things suited to him, on top of an assistant that a full-time team keeps improving underneath him.
How to set up an AI chief of staff (without building your own)
Marcus's setup reflects a builder's instincts, but the pattern works for anyone whose attention is spread across too many channels. Here is the shape of it.
- Pick the assistant you will bet on: Judge the options by one question: does this get you to your end state faster than your own scaffolding would? Pick the team that works on this problem full time, then commit.
- Give it an identity: Give your assistant its own email address and a name. It can join threads, work with other people directly, and take blame gracefully while it learns. A named assistant makes the delegation visible instead of hidden.
- Start with one channel: Loop the assistant into your email first. Forward it a thread, tell it what you want done, and let it handle the follow-through. One channel is enough to learn how it thinks and to build the trust the next steps need.
- Expand the funnel as trust grows: Add the calendar, the messages, the booking tasks, one at a time. Grant access deliberately, keep anything truly sensitive off the machine the assistant can see, and use dedicated accounts with their own protections for anything involving money.
- Set the cadence: The goal is a digest, a ping every hour or two with the short list of things that actually need you. You make the calls, the assistant does the legwork, and you stop checking five apps to find out nothing happened.
Main takeaway
The build-versus-buy math on an AI chief of staff comes down to one comparison: your weekends and evenings against the velocity of a team that eats, breathes, and sleeps this one problem. The agent you are sketching in your head is already someone's full-time job, and they will out-build you. The winning move for a builder is to bet on one, then spend your building energy on top of it.
That is the quiet promise of an AI chief of staff. The noise of your life goes through it first, and what reaches you is the short list that actually needs you.
If your nights are going into agent scaffolding, hatch your own at vellum.ai instead.
FAQ
What is an AI chief of staff?
An AI chief of staff is a personal AI assistant that sits between you and your communication channels, reads what comes in, and surfaces only what needs your attention. Where a simple AI tool answers questions, a chief of staff manages flow: it triages messages, drafts replies, handles follow-through, and checks in with a short list of decisions only you can make.
Should I build my own AI assistant or use an existing one?
If your goal is the learning, build. If your goal is the working assistant, the math favors buying. A dedicated team ships infrastructure, integrations, and fixes full time, while your version only improves on your weekends. The rule of thumb: bet on the team that gets you to your end state fastest, then build on top of it.
What can an AI chief of staff actually do day to day?
A capable assistant joins email threads you forward and handles them, books appointments online with a dedicated credit card, and works under its own email address so colleagues can deal with it directly. An assistant like Vellum picks up new channels and tasks as your trust grows, which is how a single email loop becomes a full funnel.
How is an AI chief of staff different from a regular AI chatbot?
A chatbot waits for you to ask it things. A chief of staff watches your channels and comes to you, with a digest of what arrived, drafts ready to send, and questions only when a decision needs you. The difference is direction: you feed a chatbot, while a chief of staff feeds you.
Is it safe to give an AI assistant my credit card?
Treat it like any delegation of spending. The safe approach is a dedicated card on its own account, so there is a clean line around what the assistant can spend, full fraud protection behind every transaction, and nothing else exposed if something goes wrong. Start with low-stakes purchases and widen the lane as the assistant earns it.
Can an AI assistant manage multiple inboxes and apps at once?
That is the end state, and the right way to get there is one channel at a time. Start with a single inbox, learn how the assistant handles it, then add calendars, messages, and other accounts as each one proves out. The goal is a single funnel where everything flows through the assistant and reaches you as one prioritized list.
How much does an AI chief of staff cost?
A human chief of staff or executive assistant runs thousands of dollars a month. Personal AI assistants land around the price of a software subscription: Vellum starts free with a Pro plan from $50 a month. The cost comparison that matters for a builder is different, though: a subscription against the hundreds of hours a self-built agent quietly consumes.
Do I need to be technical to have an AI chief of staff?
No. The builders were just the first ones in, because they could see the end state. Setting up an assistant like Vellum is connecting accounts and describing what you want in plain English, and the daily loop is forwarding things and answering check-ins. The technical depth is there if you want to extend it, and invisible if you do not.
Can an AI assistant read my texts and phone calls?
Only if you give it that access, and that is a deliberate power-user choice. Granting an assistant access to your Mac lets it see messages and calls as part of your funnel. The principle to follow: the assistant sees only what lives on the machine you give it, and anything truly sensitive stays somewhere it cannot reach.
What should I delegate to an AI assistant first?
The repetitive thing you check most. For most people that is one inbox: let the assistant read what comes in, draft responses, and bring you the short list that needs a decision. First jobs build the trust that everything else depends on, and they prove value in days rather than months.
Will an AI chief of staff send messages without my approval?
Only if you set it up that way. The pattern that works keeps you on the send button: the assistant drafts, you approve, it learns from your edits. Over time you may let it send low-stakes replies on its own, but that is a lane you widen deliberately, message type by message type.
