AI native

We build the company
the way we build
the software.

Navon sells racks, power and compute. An agent builds none of that. It builds almost everything around it.

ONE ENGINE EVERY MACHINE ROUTED THROUGH THE SHAFT A MOTOR ON EVERY MACHINE THE FLOOR FOLLOWS THE WORK, NOT THE POWER
5Skills written down and checked into the repo
51Components under one machine-checked contract
4Checks every change passes before it merges
54Pages held to that contract by two people
01 · Why now

The factory did not change
when the motor arrived.

It changed about forty years later, when someone moved the machines. Most companies are living in the first half of that gap.

1890s MOTOR AVAILABLE 1920s FLOOR REDRAWN, OUTPUT JUMPS FORTY YEARS OF ALMOST NOTHING

Steam factories were built around one driveshaft. The first electric ones kept the shaft and swapped the engine, so the power source changed and the work did not.

Bolt AI onto the company you have and you get the same shape.

One clever tool, a licence for everybody, every process still routed the old way. The steps get faster, which feels like progress until you notice you have sped up a process that should not exist.
$49bn 2005 under $10bn 2020 US NEWSPAPER PRINT ADVERTISING · PEW RESEARCH CENTER

Light, information and compute all collapsed in price once they became abundant. The journalism was the same. The scarcity was gone.

Everyone rents the same model. What you build around it is the difference.

02 · The shift

A chatbot answers.
An agent finishes.

The difference is whether anything happened without you.

A CHATBOT YOU CHAT BY HAND THE WORK REPO INBOX DEPLOY NOTHING MOVES UNTIL YOU MOVE IT AN AGENT YOU AGENT THE WORK REPO INBOX DEPLOY IT WORKS WHERE THE WORK LIVES

A chatbot ends every exchange with a human carrying the result across a gap. An agent holds context between steps, reads and writes real files, and comes back with the job done.

01

A goal

Clear enough that someone else could tell whether it was met.

02

Skills

The procedure written down, in the order you actually do it.

03

Tools

The repository, the data, the deploy, the inbox. Real access.

04

Context

What your business knows and the model does not.

Miss one and you are back to doing the work yourself

03 · The system

Five layers.
One of them is yours.

Read it bottom to top. Open a layer to see why four of the five are the same for everybody.

04 · The pieces

Memory, skills,
software.

Three ordinary things make up the layer nobody can copy. Writing them down now buys you labour, not just onboarding.

Memory

Decisions and the reason behind them, the shape of a customer, what went wrong in March. Held in a file rather than in one person's head, which is also the only version an agent can read.

Org chart name: culture and expertise

Skills

How we quote a colocation deal, how we answer a scoping request, how a page ships. Written plainly, with the failure modes named, because the failure modes took three attempts to learn.

Org chart name: standard operating procedure

Software

Tools built for how you work rather than how a vendor assumed you would. This used to be the expensive option. The agent writes most of the code now, so it often is not.

Org chart name: software
05 · The moat

Write it down lower,
repeat it cheaper.

Moving knowledge down this staircase is how a company stops depending on who is in the room. An agent can work at every step.

Culture

How people here decide when nobody is watching.

Cost to repeatSlowest

Expertise

What the person who has done it forty times knows.

Cost to repeatOne person at a time

Procedures

The same knowledge written down, in order.

Cost to repeatCost of reading it

Software

The procedure turned into something that runs.

Cost to repeatFree

Implicit at the top, explicit at the bottom

A project lands only if everybody does their part at the right moment. Each extra person is another independent chance of that failing.

59% Chance the project lands with 5 people
That team59%
One person directing agents90%

Agents share one context, hold no competing priorities and do not lose the thread at a hand-off. Arithmetic, not a law of management: real teams carry slack and cover for each other.

06 · At Navon

What this looks like
in an infrastructure company.

The racks are real and an agent pours none of them. Everything around them works like this.

Pull the work in

Take what already happens in inboxes, chats, repositories and databases instead of asking people to write it somewhere new.

Airbyte · Composio · Unipile

You can start this with a folder of text files. We did. Buying a platform first is how people end up with an empty one.

07 · Start

A folder of notes
is enough to begin.

No platform, no procurement, no migration. The first version costs an afternoon.

01

Pick one repeated process.

The one you explain to new people most often. Not the hardest thing you do.

02

Write it down plainly.

In the order you do it, with the two or three places it usually goes wrong.

03

Hand it over with access.

Goal, skill, tools, context. Then let it finish instead of checking every step.

04

Correct the writing.

Fixing the output solves today. Fixing the instruction solves every run after it.

Everybody ends up managing agents. The work moves toward deciding what is worth doing, handing it over with enough context, and reading what came back. The memory and the work become the same act.

The point

Do not hand AI to the organisation
you already have. Build the
next one around it.

Start with one process, write it down properly, and let the compounding do the rest. The design system is public and the checks are in the open.

Sources

  1. Paul A. David, “The Dynamo and the Computer”, American Economic Review, 1990. The gap between electrification and the productivity gain, and the factory layouts behind it.
  2. Pew Research Center, State of the News Media. US newspaper print advertising fell from roughly $49bn in 2005 to under $10bn by 2020.
  3. Navon figures come from this repository. Run npm run ds -- --check to verify the component, pattern and asset counts.