Jukka Blomberg 展示一人工作室職能圖：23 項工作自動化、5 項由本人把關、1 項選擇不開展

I run 14 ventures with no employees. That claim is cheap. What is less cheap is writing down every job a company needs and putting a name next to each one.

So I did. The table below is generated from the system map of my studio every time the site builds. It lists the functions a small company needs, and who runs each one today. There are three possible answers: an AI agent with a written contract, me, or nobody.

The count as served today: 23 automated · 5 human gates · 0 not covered · 1 not pursued.

The org chart

The live version is at jukkablomberg.com/ai.

Three things the table taught me

1. "0 not covered" is the number that matters, and it was not always zero. The useful exercise is not counting agents. It is listing the functions first, from the point of view of a company, and only then asking who does each. When I first did that, several rows had no owner. One agent was producing an output nobody read. Another job had been "assigned" in a note that the responsible agent never saw. A function with no named owner does not fail loudly. It just stops happening.

2. One agent can hold several jobs, and that is fine as long as it is written down. The Convertor answers inbound, drafts outbound and handles customer support, because all three are the same skill: reply to a real person within a day with something finished. The Opportunity Strategist also owns pricing. Collapsing jobs onto fewer agents is cheaper than one agent per box. The rule that keeps it honest: a duty is only real if it is in that agent's own instructions. Putting it in an org chart does not count.

3. The human rows are not "the things AI is bad at". They are the things where a mistake is expensive or hard to undo: money going live, credentials, anything that contacts a third party in my name, signing terms, and the relationships that matter most to me. Agents do the preparation; the decision stays a human click. Where an exception exists (some routine email replies, for example), its edges are written down: who, which address, how many a day. The gates are narrow on purpose. A system that has to ask about everything gets ignored, and one that asks about nothing will eventually spend your money.

What it produces

These are the counts the same site serves, each with its source and an "as of" date:

• 1,236 items built in the last 90 days (one append-only ledger line per item handled)
• 3,413 commits pushed to GitHub in 90 days, across 42 repositories
• 12 live products across 14 ventures
• 244 AI-agent queries in 7 days to one venture's MCP data server

None of these mean the work is good. They mean the work is counted, and that a reader can check the counting. The full method, including where each number comes from, is on the Portfolio OS page.

Try it on your own team

You do not need agents to get value from this. Take a blank sheet and do it in this order:

1. List the functions your team or company needs, without looking at the people you have.
2. Write one owner per function: a person, an automation, or "nobody".
3. Every "nobody" gets one of two labels: not covered (a gap) or not pursued (a decision, with a date).
4. For every automated row, find where its instructions actually say it owns that job. If you cannot find it, the row is not covered.
5. Decide your human gates before you automate anything else.

Step 4 is where most of my gaps were.

I'm Jukka Blomberg, ex-CMO at two crypto exchanges and before that at Nokia and Microsoft. I now help leadership teams decide what to do with AI, and then build it. The org chart above is generated from my own system map; the numbers are as served on 28 October 2026.