I was listening to an old interview with Jeff Bezos last week, and one thing he said stopped me cold.
Someone asked how he set Amazon's strategy in the early days. He said he asked himself one question:
What will always be true about what a customer wants?
He landed on three things.
Lower prices. Faster delivery. More convenience.
Then he built the entire company around those three answers and never deviated. That focus made Amazon the most valuable company in the world and made him, for a stretch, the richest man alive.
So I started asking myself the same question about AI. What is always going to be true for businesses in an AI-first world?
I keep landing on the same two answers.
1) Every business will need a way to make its own data accessible to AI
2) Data is the foundation for every successful AI deployment
Bad, unstructured, scattered data makes every AI project a slog. Clean, structured, easy-to-access data makes every project a hundred times smoother.
My business partner built a simple graphic to show how we should be prioritizing AI infrastructure at the company level:
Data sits at the foundation.
Everything stacked on top, the agents, the automations, the fancy workflows, is irrelevant if the data layer underneath is a mess.
This is true for every company in the world, from 1 person solo businesses to the enterprise.
But fewer than 1% actually have a properly structured data layer.
My friend Adam Sandler (not that Adam Sandler) has gone all in on this exact gap. He builds structured knowledge bases for small businesses, and he's crushing it.
The beauty of Adam's model is the sequence.
He sells an audit on the front end. The audit tells him exactly where a client's information lives today (Google Drive, Dropbox, local machines, somebody's inbox) and what the structure should look like. That's the deliverable. It also happens to be half the work of the build.
Then he sells the build itself as the upsell, somewhere between $3,000 and $5,000.
Here's the part that made my sales brain light up:
Once the knowledge base is built, he can literally ask the client's own data where the leaks are. Now that everything lives in one structured place, the data tells him exactly where the next opportunity is.
Maybe inbound leads take 22 hours to get a reply. There's the speed-to-lead build. He doesn't have to sell the next project. The data sells it for him.
The objection I'm sure I'll hear: "Aren't a hundred startups already building this?" Sure. But they sell an off-the-shelf, do-it-yourself tool.
They don't sell the consulting, and they don't get to learn how your specific business actually runs. A solo operator who sits with a client and sees how the sausage is made delivers something no off-the-shelf product can.
I had Adam on the Build With AI podcast last week. He shared his screen and showed us his entire process, beginning to end.
If you want to copy it for yourself (or your clients), here's how Adam builds one, start to finish.
Step 1: Lead with the cleanup, not the AI.
Your foot in the door is not "let me build you an AI agent." It's "your company knowledge is scattered across Drive, Dropbox, and three inboxes, and I'm going to organize it into one living asset."
Step 2: Audit where the knowledge lives.
You don't need access to everything. A verbal map is enough. "Brand stuff is in Notion, client info is in Drive, transcripts are in Fathom." That alone lets you sketch the schema and deliver the audit.
Step 3: Use the seven note types as your structure.
Every knowledge base Adam builds ladders up to seven durable note types:
- a snapshot of the business
- the people and contacts
- preferences and rules (how they work)
- project history
- decisions and rationale
- open loops (what's in flight)
- links to the source of truth for each
Find the "spine" too, the one schema everything connects back to. For most clients it's their annual goals.
Step 4: Interview the client to fill it in.
Ask a tight set of discovery questions.
- Who are the stakeholders and what does each care about?
- What have you promised?
- What decisions did you make and why?
- What's worked and what clearly hasn't?
You can run this yourself, hand them a prompt that walks them through it, or even point a voice agent at it.
Step 5: Build it with markdown, not software.
Run one prompt that turns those answers into seven markdown files plus a simple index file that acts as the table of contents. That's the whole knowledge base. It lives on the client's machine. No fancy database or Obsidian required.
Step 6: Make it self-maintaining with two skills.
Paste a set of global instructions into the client's Cowork so it knows where the base is and how to use it. Then add an ingest skill (drop files in, or pipe in Gmail, calendar, and CRM, and it files them automatically) and a curate skill that runs once a week to flag stale notes, contradictions, and gaps. Now the base ingests and cleans itself.
On pricing, roll the audit into the build so it feels free. A $999 audit becomes a $3,500 core build or a $4,700 premium build, plus a couple hundred a month to maintain.
That's the whole play.
Knowledge bases got popular in the last few months, but I think we're about a year out from this exploding and everyone piling in. Build one for yourself first this week. Once you feel the unlock, you'll sell it with real conviction.
Reply and tell me what you'd put in your knowledge base first. I read every response.
- Corey
P.S. Watch my full breakdown with Adam, where he walks through the entire build live on screen.
A few notes/tweets/cool things sourced from the AI community.
1)
Use Codex to vibe code on the move.
2)
My team is testing this model (it works).
3)
How to set up the best open source model in the world.
Thanks for reading the Corey's Notes newsletter. I'd appreciate it if you sent it to a friend that might find it interesting.
Be sure to check out the Build With AI podcast on Apple, Spotify, and YouTube. If you're a non-technical entrepreneur who wants to learn how to make money with AI, you'll love it.
Be back next week.
-Corey
