NORTH And The Trust Ladder

NORTH, and the trust ladder

Five silent killers of adoption. None of them are in the code.

---

The dashboard nobody opens

A model scores well. It ships. Six weeks later adoption sits in single digits.

The leadership question is always the same. The model is perfect. Why will they not use it?

The answer is usually standing in the doorway, arms folded, remembering the one time it got something badly wrong and who had to fix it.

That is not a technology conversation. It is a trust conversation, and trust has structure.

N · Narrative

When a leader says this will improve efficiency by thirty percent, the room hears we are cutting thirty percent of the workforce.

That translation is instant, it happens at every level, and it is unaffected by what the memo actually said.

There is one question that decides which story people tell themselves.

Is this a weight rack, or a wheelchair?

A weight rack makes you stronger. A wheelchair replaces your legs. People need to know which one this is, and they need to hear it from someone they trust rather than from a slide.

The practical move is not communications. It is participation. People do not sabotage systems they helped design. Have the people whose work is affected build the test cases. When they define what correct means, they own whether it is correct.

O · Ownership

Software is owned by the technology function. Agents are owned by the business.

That is structural, not semantic. When an agent approves something it should not have, the question is not what was wrong with the prompt. It is who authorised this agent to make that decision.

Before any production deployment, one named person accepts five things:

1 · They understand which decisions the agent makes without asking. 2 · They have reviewed and approved its risk tier. 3 · They have defined the escalation path for when it fails. 4 · They will be notified within a stated window when a circuit breaker fires. 5 · They have committed to reviewing all of the above on a fixed cadence.

If nobody can be found who is willing to sign that, the agent is not ready.

This is not a bureaucratic gate. It is the organisational immune system, and an organism without one does not notice it is sick.

R · Readiness

You cannot build a reliable agent on unreliable knowledge. This is obvious and it is ignored constantly.

Five signals, each of which is a specific piece of work rather than a warning:

What you observe	What it means for the agent
Subject matter experts spend a fifth of their time answering the same questions	The knowledge base is not captured. The agent will fill the gap by inventing
Critical data lives in files named FINAL_v3_DO_NOT_TOUCH	There is a hygiene problem that has to be resolved before deployment, not during
API documentation predates the last migration	Integration will take roughly three times the estimate
The compliance function has not been briefed on agents at all	The governance gate is the bottleneck. Bring them into discovery, not review
Frontline staff cannot say what the data rules mean for an AI output	There is a training gap that raises exposure at the point of use

The failure this catches is specific: an agent that scores badly in its first week not because the model is weak but because the knowledge base held three contradictory answers to the same question and nothing chose between them.

T · Trust

Binary trust kills programmes. Somebody says turn it on, the first public failure lands, trust goes to zero, and rebuilding takes years.

The ladder prevents the collapse by making autonomy something that is earned in stages, the way it would be earned by a new colleague.

Rung	What it does	What earns the promotion
1 · Draft	Drafts. A human approves everything before any action	A sustained period at high acceptance with zero critical errors
2 · Shadow	Acts in parallel with a human. Outputs compared, never used	A sustained period showing accuracy within a defined band of the human baseline
3 · Bounded live	Handles defined low-risk transactions on its own. A threshold separates them from the rest	A sustained period below a defined error rate, at real volume
4 · Full scope	Handles the defined workflow. Humans handle exceptions	Periodic review showing sustained accuracy and unit cost

Autonomy increases as the rung number increases. That direction is worth stating plainly, because the opposite convention exists in circulation and two teams using opposite conventions will both say we are at rung three and mean opposite things.

Three properties of a ladder that is actually working:

The promotion criterion is written before the rung is occupied. A criterion invented at promotion time is a rationalisation.

Demotion exists. A ladder with no way down is a ratchet, and a ratchet is how an agent ends up at full scope after an incident nobody logged.

The rung is on the card. If autonomy is not a recorded field, it is a belief.

This four-rung ladder describes one agent's journey. The five-level Seniority Ladder describes what kind of agent it is, which is a different question and a longer scale. An agent can be a level 3 system sitting at rung 1 while it earns its way up.

H · Habits

The underestimated risk is not resistance. It is over-reliance.

After six months with a system that is usually right, people stop checking with the same rigour. That is the mechanism by which silent drift becomes a financial problem.

Training people to use AI takes half an hour. Training people to supervise it is a practice. The distinction is between auditing an output and accepting one.

What the practice looks like:

Week 1. Review everything. Flag anything that feels wrong, including the ones you cannot explain.

Week 2. Bring the flagged cases to a short team review. Find the pattern in what got flagged.

Week 3. The pattern becomes a new test case. The person who flagged it just improved the system.

From month two. Review a random sample weekly. The sample rate falls as accuracy holds.

Continuously. Every new edge case a human finds goes into the test suite.

The loop is the point. Humans improve the agent, and the agent expands what the humans have time to do. Either half without the other is not supervision.

The scorecard

Score each dimension from one to five, before kickoff, with input from frontline supervisors rather than only from leadership.

N. Have we told the affected team what this does and what it explicitly does not do? One: no. Five: yes, and they helped write it.

O. Is there a named business owner who has signed? One: no owner identified. Five: signed, briefed, review cadence scheduled.

R. Have we confirmed clean access to the agent's primary knowledge sources? One: spreadsheets. Five: documented interfaces with defined schemas.

T. Is there a written ladder with a defined criterion per rung? One: planning to turn it on. Five: shadow mode scheduled before any live traffic.

H. Is there a structured auditing practice for the people who will supervise it? One: not designed. Five: designed, training scheduled.

20 to 25. Proceed.

13 to 19. Close the gaps first. They will not close themselves later.

Below 13. Stop.

And say why plainly, because it is the most useful sentence in the framework:

You have a management problem, not a technology problem.

One thing the scorecard does not measure

When an agent takes over work a person was doing, that person's job changes.

That is not a side effect to be managed afterwards. It is a design responsibility, and it belongs to the same function that owns the agent.

Which means a reskilling path is defined before the agent enters shadow mode, not after the results come in. It also means a value report says where the recovered hours went, not just how many there were.

The measure of a working centre of excellence is not how many agents it deployed. It is how many people moved to better work.

_{Derived from the AI CoE and Agent Factory Playbook, based on The Augmented Enterprise framework. Ladder numbering reconciled against the agent seniority ladder held in the practice canon.}

_{Back to Home · The framework spine}