The Arch, a method by arkhen

Most AI projects don't hold.

Not for want of a model. For want of order.

The problem

Why this method

AI is offered as an answer before it has met a question. Companies buy in because the market demands it, because the board put it on the agenda, or because a competitor announced it. Rarely because a specific problem was stated, measured, and judged solvable this way.

Seen from the inside for three years, across three companies with nothing in common: three sectors, three sizes, three levels of maturity. Everywhere, the same questions, the same fears, the same mistakes.

This method does not describe what ought to be done. It describes what was done. No step entered it without a field case to support it.

The method

Seven stones, in order

An arch does not hold because the last stone is beautiful. It holds because each voussoir presses on the one before it, and because a central stone locks the whole.

The Arch is the method. arkhen is the practice that publishes it, documents it and puts it to work. One is read; the other keeps it current.

The foundations

What you give the system, before discussing the system itself.

01

Investigate

The work as done, not as described

The move
Draw out what sits in people's heads: the explicit knowledge, and above all the tacit knowledge.
Why
No model guesses what was never written down. As long as the knowledge lives in the heads of those doing the work, it is beyond the reach of any machine.
02

Quantify

The pain before the tool

The move
Qualify and quantify the problem and the expected outcome, before any technical choice.
Why
Without a baseline, there is no proof at the end. A project that cannot say what it costs today will never be able to say what it returned.
03

Verify

What the data is actually worth

The move
Measure the real state of the data in scope, before building anything on top of it.
Why
Conventional processing applied to false data produces a visibly false result. A generative system produces a result that is false and credible, and all the harder to challenge for being well written.

The build

What you construct, where you are allowed to place it, and what the human keeps in hand.

04

Connect

Context before model

The move
Make the data answer one another, rather than stacking it up.
Why
A system reasons only on the context it is given. Data that is accurate but isolated yields nothing but partial answers.
05

Place

Where the error is paid for

The move
Decide whether AI serves, and where it is allowed to sit.
Why
The question is not whether the machine can do it, but what an error costs in that exact place. The answer determines the right to be there, and walking away is one of the possible outcomes.
06

Frame

Human decision where it commits

The move
Define what the system proposes and what the human settles, use case by use case.
Why
A system that proposes and a system that decides do not carry the same liability, and the law does not treat them the same way.

The lock

The central stone, alone in its movement. Without it, nothing that came before holds.

07

Govern

The keystone

The move
Establish who decides, against which thresholds, and under what supervision.
Why
This is the central stone: it is not bolted on at the end of the chain. Remove it and everything else collapses, including what seemed to stand on its own.
Striking the centring

Re-measure

Reopen the baseline measurement and conclude. This is not an eighth stone: it is the moment the temporary frame comes away and you find out whether the structure stands.

Nothing is proven until the centring is removed.

The tools

Five supports, one per step that calls for it

Each support can be used without assistance.

Investigate

The interview guide

A question framework that gets people to describe the work as it is actually done, not as it is documented. It covers the organisation, the roles, the processes, the tasks, the pain points and the objectives.

Quantify

The pain sheet

One sheet per problem: it states the problem, puts a number on it, and sets the expected outcome before any tool is chosen.

Verify

The critical attribute check

A short test applied to the single piece of data that carries the most weight in scope, to establish what it is actually worth.

Place

The cost-of-error matrix

A two-axis grid crossing the severity of an error with the place the system sits, and deducing what is permitted and what is not.

Re-measure

The re-measurement protocol

The procedure that reopens the baseline once the structure is in place, and makes it possible to conclude.

How they work, how to use them and where they stop being reliable are in the Arch guide, which comes with the diagnostic.

The sources

What this rests on

The method was tested against what already exists, always at the original documents and never at second-hand accounts.

What each one supports, and where the method departs from it, is set out in the guide.

The master builder

Clément

behind the Arch

AI practitioner in business.

Three years on AI projects across three unrelated sectors: industrial distribution, healthcare, energy. In turn product owner, AI coordinator and consultant.

This method is what I took from it. This site is the verifiable trace.

Where do you stand?

Run the diagnostic on where you actually are, not where you ought to be. The report goes out by email, and the Arch guide with it if you ask for it: the method voussoir by voussoir, how to use it and where it stops.