Notes
Week 6. The first surface.
By week six, the AI started showing me things I'd forgotten I knew.
By Samer Azar, Fractional CFO · 2026-05-16 · 4 min read
Key Takeaway: The asset that compounds in AI workflows is your private archive, not the model. Models commoditize on a 60-day cycle. The archive grows. Inflection lands around week six. Reliable by month three. Undeniable by month five. The work isn't prompting better; it's logging consistently while it looks like nothing is happening.
Hi Reader,
I went into a call yesterday thinking I knew the man across the table. My AI knew him better than I did.
The call was for a founder I work with. Real money on the line. I had a window between meetings to prep, maybe 30 minutes.
I opened the work expecting to do the usual: scan the deck, skim the email thread, scribble a few talking points on a sticky note.
Instead three things landed in the first ten minutes that changed the call.
Beat one: the buried shareholder
He'd put a small check into the company years ago. I'd forgotten he was on the cap table. My archive hadn't.
That panel is the archive talking. Months of input it had been quietly chewing on, surfaced in one paragraph the moment I needed it.
Beat two: the public writing mirror
He'd written publicly about how he thinks: long essays, specific frames he returns to. The work surfaced which of his published phrases would land in a live conversation, and which of his pet themes the founder should mirror back to him.
Beat three: the numbers and the honest answers
It told me which three numbers he'd press on. Then it told me the three honest answers I was afraid to give. The ones that build trust faster than the polished ones do.
What happened on the call
The call ran fifty minutes. He opened by saying the bridge should be twice what we'd planned. Walked out with a Friday deadline to send him a model so he could underwrite it himself.
Here's the thing.
I didn't write a better prompt yesterday. I didn't use a better model.
What actually worked
What worked was the archive. I started building it about five months ago, when I connected every meeting transcript and every email into one place the AI could actually read. Five months of dumping context into the system without expecting anything back. Emails, meeting transcripts, board decks, financial actuals, hallway conversations. Things I thought I'd never re-read.
For the first six weeks, the AI felt like a slightly faster search engine. Then something shifted. Around week six, it started referencing things I'd forgotten. Not in answer to a direct question. In answer to a question that was about something else, and it pulled in the forgotten thing as context.
By month three it was happening reliably. By month five, yesterday, it walked me into a call knowing things about the person across the table that I'd genuinely forgotten.
The buried shareholder. The published essay nobody else on the team had read. The KPI hiding in a six-week-old export.
The bet most founders are making
The bet most founders are making right now is on the next model: better reasoning, longer context, faster inference. That bet's already lost. Models commoditize on a sixty-day cycle.
The bet that compounds is on your private archive. It's boring work, slow to pay off, and there are no screenshots in it for Twitter. Five months before you really see what it can do.
It is also the only investment in AI right now that changes what you can actually do.
Where to start
So if you're trying to figure out where to put your time, the answer probably isn't a new prompting framework or another tool subscription.
The answer is to start dumping context today, knowing it'll look like nothing is happening for a month.
It is. And then around week six your AI tells you a thing you should have known and didn't.
That's the first surface.
A question for you
Three months in? Six weeks? Just started? Reply with where you are. I read every reply.
Below the line
Below the line: the two output schemas to build for yourself (KPIs to defend, what to handle carefully), with the prompt you actually paste in to generate them. The five rules I follow to keep the archive growing. And the benchmark behind why this isn't just my experience. I've watched the same pattern across three different archives I maintain.