Second Brain
as a Service

Every company is sitting on a pile of knowledge. Proposals, call notes, SOPs, pricing, the stuff locked in their best people's heads. None of it is organized, so when they point AI at their business, they get generic junk back.

Here's the one line that explains the whole business: AI without context just scales your mistakes. AI grounded in a company's real knowledge becomes a machine that prints leverage.

You are the person who organizes that knowledge so the AI can use it. That's it. You're not building software. You're onboarding the AI to the company the same way you'd onboard a brilliant new hire who otherwise knows nothing about the business.

The money has two parts:

• $2-5K upfront to build the brain. You collect their knowledge, organize it, and set the rules for how the AI reads it.
• $500-1K a month to keep it fresh. New stuff gets added, old stuff gets retired, and the AI keeps getting smarter.

That monthly fee isn't a fake retainer. Keeping the brain fresh is the part that actually drives results, and it's the single biggest reason these projects work or die. More on that in Section 6.

Quick credit so you sound smart when someone asks: the technical pattern comes from Andrej Karpathy's "LLM wiki" idea. Garry Tan turned it into a tool called gbrain. Corey Ganim is the one who framed it as a business you can sell. Nobody owns the definitive how-to yet. That's the opening.

Most people hear "organize a company's knowledge for AI" and immediately reach for the complicated stuff. Vector databases, embeddings, RAG pipelines. Skip it. For almost every small business, you don't need any of that.

Anthropic says it straight: if a company's knowledge is under about 500 pages, you can just hand the whole thing to the AI. No fancy database needed. The team that built Claude Code tried the complicated version first and dropped it because letting the AI just read and search organized files worked better, and was simpler and more private.

So here's the wedge that makes you different: you sell organized files plus one smart instruction file. Not a software platform. The company owns it. It's simple. And it beats the expensive enterprise tools for the kind of business you'll actually be working with.

The beginner stack (pick one, no code)

• Google Drive + NotebookLM. Start here. Organize the files in Drive, point NotebookLM at them. It cites the exact sentence behind every answer, which kills the "is the AI making this up" fear instantly. Basically free.
• Claude Projects. Best when they want deeper thinking over the knowledge as an internal tool. Works on the free plan.
• A Custom GPT. Good when they want a shareable assistant with a link. Note the trap: a Custom GPT caps out at 20 files. That ceiling is exactly the kind of problem you get paid to solve.
• Notion AI. Only if they already run the whole company in Notion.

For context, the enterprise version of this (a tool called Glean) starts around $50 per person per month with a 100-seat minimum. That's $60K a year. You're the answer for every company that was told they need that and don't.

What "set the rules" actually means

This is the deliverable. Four pieces.

1. A folder structure. Two levels deep, max. Numbered so it always sorts the same way.

2. Naming rules. No "Final FINAL v3 (1).docx" chaos. Lowercase, dashes instead of spaces, dates written as 2026-06-17 at the front of anything dated. So 2026-06-17-onboarding-sop.md, not onboarding final.docx.

3. A label block on every file. A few lines at the top of each note that tell the AI what it's looking at: title, type, who owns it, when it was last checked, and whether it's confidential. Those last two become guardrails. The AI can refuse to share secret stuff and warn you when something's gone stale.

4. The navigation file. This is the actual product. One file, named CLAUDE.md, that tells the AI how the whole thing works and how to behave. Keep it under 200 lines. Anything longer and the AI starts ignoring it. Here's the template you can steal.

Once the folders and the navigation file are set, you don't hand-write every article. You let the AI build the wiki for you. Drop in all their raw files, then paste this.

The $2-5K build plus $500-1K a month isn't a made-up number. Similar "internal AI tool" builds sell for $2,000 to $7,500. Full small-business AI projects run $10-15K. You're pricing the simple, fast version of something companies already pay real money for.

The real risk isn't charging too much. It's letting the job balloon while you stay cheap. So scope it tight.

How long one takes

The number of files barely matters. The mess is what eats your time. Cleaning up bad files is 30 to 50 percent of the whole job.

• Clean, ~20 docs: 17-30 hours. This is the job your $2-5K price assumes.
• Medium, 100-200 docs: 38-57 hours. That's really a $6-16K project. Don't sell it for $5K.
• A giant messy dump: 60-110+ hours. Walk away or charge a lot.

Scope rule: only take their most important files. Top few hundred, one or two functions. Write down what you're not doing right in the proposal. No "organize everything," no fancy connectors, no public-facing bot. That one paragraph protects your whole margin.

What the monthly fee actually buys

If you can't answer "what am I paying for every month," neither can the client. Here's the real work:

• New stuff in, old stuff out. Add new files, retire the outdated ones, fill the gaps where the AI couldn't find an answer.
• A monthly report. How much it got used, top questions, hours saved, what's missing. This one email is what makes the client keep paying.
• A quality check. Run a handful of real questions every month and make sure the answers are still right after changes.

You don't have a portfolio yet. That's fine. Your first job becomes the portfolio.

Do one tightly-scoped pilot, free or cheap, for a testimonial. One team, one process, two to four weeks. In exchange, lock in three things at the start: a written and video testimonial, permission to use an anonymized before-and-after, and one or two warm intros. That pilot is the Trojan horse. You expand it into the paid version once they feel it.

Start with the easy clients: agencies, coaches and online educators, and software companies. They're documentation-heavy, they already get the ROI, and they're not buried in regulations. Local service is great for simple pilots but needs more hand-holding. Skip healthcare and law as your first niche, the privacy rules are a headache.

The pitch (never mention the tech)

Don't say RAG, embeddings, or vector database. Ever. Say this:

Three pains to press on, all of which cost time and money: generic AI fails on real company questions, knowledge walks out the door when people quit, and onboarding new hires takes forever.

Handling the obvious objections

• "We already have ChatGPT." Right, and it's a smart hire you never onboarded. I onboard it.
• "We tried a chatbot, it was useless." That was a brittle FAQ bot. This is grounded in your real documents and cites its sources, so you can trust the answer.
• "Our files are a total mess." That's exactly why this exists. Cleaning it up is half the job, and it's included.

Here's the stat that should scare you into doing this right: about 95% of company AI projects fail to make any real money. Not because the AI is bad. Because nobody designed for the boring stuff. Avoid these and you're already ahead of almost everyone.

Privacy, the part you cannot get wrong

You're handling a company's private data. One rule above all: never use the free consumer chat versions for client data. Use the business or API versions of Claude, ChatGPT, or Google Workspace. The consumer versions can train on what you put in. The business versions don't. Get an NDA signed, and for sensitive clients keep their data inside their own accounts, not yours.

One thing to understand and explain: "we don't train on your data" is not the same as "we keep nothing." Most providers still hold onto prompts for a short window to watch for abuse. Know that so you can answer it honestly.

The four ways these die

• The brain goes stale. This is the number one killer, and the entire reason the monthly fee exists. Old info in, wrong answers out, trust gone.
• Nobody uses it. If the owner doesn't adopt it, the team quietly goes back to pasting company data into free ChatGPT. Win one visible person in week one.
• Garbage in. If you skip the cleanup and load messy files, the answers are messy. The cleanup is the job.
• No proof. If you never set a baseline, you can never show the win. Measure something on day one so your monthly report has teeth.

Notice that every failure points back to one thing: the work after the build. That's not a bug in the business model. That's the business model. The build gets you in. The maintenance is where you become impossible to fire.

Don't overthink it. The fastest way to learn this is to build one before you sell one.

• Build a free one for yourself or a friend's business. Pick the messiest pile of files you can find. Organize it, drop in the navigation file, run the build prompt. Now you've done it once.
• Pick your industry. Choose the one from Section 4 you already understand. Going deep on one beats being generic across all five.
• Send three messages. Use the pitch from Section 5. Offer one scoped pilot in exchange for a testimonial. That's your first client.

The window on this is open right now because almost nobody is doing it yet. That won't last. The people who move this quarter own the niche before it has a name.