To sell second brains as a service, offer one bounded engagement: a paid audit of the documents a business runs on, a fixed-fee build that turns them into a sourced markdown knowledge base its AI tools read, a retrieval test that proves it answers correctly, and a handoff the client team can run without you.
How do I start my own AI automation agency?
Start with one narrow, repeatable offer instead of a menu of automations. Building a business its own AI knowledge base is a strong first one: every company has scattered documents, the result is concrete, and every later automation works better on top of it. Practice on your own material first, then sell a small paid audit.
Here is the whole engagement, in the order you run it. Nothing in it depends on one AI model or one app; the deliverable is a folder of plain markdown files.
Step 1: Build one for yourself first
You cannot sell a process you have never run. Build a second brain from your own work: your notes, your decisions, your sources, with Claude Code or another AI doing the filing. The Claude Code + Obsidian setup guide walks through it in about twenty minutes, and the practical second brain guide covers the habit that keeps it alive. Running your own for a few weeks gives you the vocabulary, the failure cases and a demo.
Step 2: Pick who you sell to
The best buyers are businesses whose know-how lives in people's heads and scattered files: service firms, agencies, contractors, practices with written procedures nobody can find. The signal to listen for is a team that already uses ChatGPT or Claude and complains that the answers are generic, or that they keep re-explaining the business to it. That complaint is a knowledge problem, and it is the one you solve.
Step 3: Lead with a paid audit
Sell something small before you sell the build. The audit inventories the client's sources, flags where documents disagree (an old price sheet next to a new one is the classic), sorts material into public, internal and private layers, proposes a structure, and ends in a written report with quick wins. Charge a flat fee and credit it toward the build if they proceed within a set window. A prospect who will not pay for an audit rarely pays for a build.
Step 4: Build from their real material
Turn the sources into short notes, one rule, decision or process per note, each with a Source line pointing at where it came from. Keep the folder under version control and log every structural change with its reason, so the client can see what changed and why. Resolve conflicts once, write the resolution down, and archive the losing version instead of deleting it.
Step 5: Test retrieval before you call it done
Write the questions the client's team actually asks, including a few the knowledge base should refuse to answer, and run them through the AI tools that will read it. Record what fails, fix the notes, and re-run until the suite passes. The second brain retrieval test shows the scoring pattern. This test is also your completion gate: the build is finished when it passes, not when you run out of hours.
Step 6: Hand it off
The handoff package is what makes this a professional service instead of a favor: a changelog, an access register that lists who can reach what (with your own access revoked), a maintenance guide naming the client's owner for weekly and monthly checks, and a walkthrough of adding a new source. Before you leave, have someone on their team add one source alone. Then offer optional monthly upkeep for the notes most likely to drift.