Four beliefs · May 2026
What we believed. What happened.
We wrote these down when we switched it on, which is the only reason we can be this exact about being wrong. The full map of what the thing reaches is in brief 015; this is the part about people.
01
What we expected vs what happenedThe wrong assumption
We expectedThe non-technical people need this most. Sales, admin, managers drowning in mail.
What happenedEngineers use AI to write code, each their own way. People from admin and sales built the first tasks that run on their own every morning. Who picks it up runs on curiosity, not on job title.
We now look for the curious person in each department, whatever the title says.
02
What we expected vs what happenedThe founder would carry it
We expectedOne person keeps the shared core alive for a year, everyone else consumes it.
What happenedCounted this morning: 237 changes since 15 May, from six people. Two of them have made more than the founder has. He is third.
The day somebody else fixes your core without asking, it stopped being your project.
03
What we expected vs what happenedSharing does not happen by itself
We expectedSolve something once, leave the recipe, the next person starts where you stopped.
What happened7 of the 77 who answered in September had added something. What moved people: someone sitting next to them for an hour, a demo, a first-steps page with seven things to try. Not the folder.
A shared brain needs pairing. A folder is not a habit.
04
What we expected vs what happenedEverybody would know it exists
We expectedShow it, tell a few people, word of mouth does the rest in a company this size.
What happenedOf the 77 who answered in September, 26% had not heard of our AI push at all. Five months, and a quarter of the answers had missed it. Telling people is a feature we forgot to build.
Adoption is not a tool problem. It is a telling problem.
Counted this morning · 7 October 2026
Today's numbers, taken today.
Everything here was counted on the repository this morning, in the same shape as the September count so the two can be compared. None of it is a survey answer, so none of it depends on anybody remembering anything.
237
changes to the shared core since 15 May 2026, from six people.
September's count was 211 from mostly one person. The split today: 61, 55, 47, 46, 15, 11. The founder's 47 puts him third. Three of the six are not engineers by job title. This is the number that changed our minds about what the thing is: a core that only its author edits is documentation, and a core other people edit is infrastructure.
Counted · our own repository, 7 Oct 2026
608
pages of written knowledge: 73 company-wide, 535 across 21 departments.
Three weeks ago it was 582. The growth is almost entirely departmental, which is what we hoped for and did not expect this early: the company-wide layer barely moves while each department writes its own. Recipes — a task written down once so anyone can run it — went from 135 to 138, of which 19 are company-wide and 119 belong to a department.
Counted · our own repository, 7 Oct 2026
43%
had it running in September: 33 of the 77 who answered our monthly survey.
The company is about 130, so this is 43% of the answers, never of the company. From the same 77: 36% see no need yet, 26% had not heard of the push at all, and 49% say they have no time to learn. We publish the survey in full every month in brief 014, including the parts that do not flatter us.
Survey · our monthly survey, 77 answers, Sep 2026 · brief 014
Three editions · our monthly survey
It only compounds if people add to it
Five people review what goes into the company-wide recipes. A department's recipes are its own call. Percentages are of the people who answered each month, never of the whole company.
For your company
What you need to start
- Your procedures written somewhere the assistant can read them. A shared folder is enough.
- Logins to the three to five tools people already use every day. Their own logins; a shared admin key only where the tool forces it, and you list those.
- One department and one curious person in it, whatever their title.
- One workflow, ninety days. The same rule as in brief 009: one repeatable process first, not a company-wide program.
- What it cost us: subscriptions of $20 to $100 per person per month, and time nobody clocked, so we have no hours to quote and will not invent any. The real price is learning: 49% of the 77 who answered say they have no time for it.
- The rule that keeps it honest: each person's assistant sees exactly what that person already sees, nothing more.
Five months in
The part we are enjoying, and the part we are not
What is genuinely good. People who were never going to be the audience for this built the first things that run on their own every morning. The core has six authors. Departments write faster than the centre does. Nobody was told to.
- Still true and still annoying: a quarter of the answers in September had not heard of the push. We have been building, not telling.
- Still unmeasured: we have no hours-saved figure we would defend. The monthly survey asks people to estimate, and people are poor judges of their own speed. We publish it anyway, labelled as an estimate.
- Still a risk: every connected tool widens what one compromised instruction could reach. Access is per person and per login, which limits the blast radius; it does not remove it. That is the subject of brief 017.
- Still unfinished: five months is not a verdict. Ask us again at a year, and if the core has stopped growing we will say so here.
We are writing this up at five months rather than at a round number because it is the first point where the answer to “whose project is this?” changed. That is a better milestone than an anniversary.