The Method
Everyone already knows what the problem is.
Nobody says what to do on Monday.
"The bottleneck is not the technology. It is the decision about how much autonomy to hand over, and how much control has to remain on the human side."
That sentence is not ours. It is the conclusion BCG, McKinsey, Bain, Deloitte and KPMG reached in 2025 and 2026, through independent methodologies, looking at different companies.
BCG puts a number on it: seventy percent of the value of AI comes from people and process, not from the model. KPMG measures the other side: organizations that can see what their own AI costs are five times more likely to prove a return. Fifteen percent against three.
The diagnosis is settled. Nobody needs to be convinced of it.
Five global consulting firms measured the same thing through independent methodologies and arrived at the same place: the bottleneck is not the technology. There are maturity reports, sector benchmarks and readiness indexes. All of them say where your company stands.
Govern before you scale. Redesign the process before you buy the tool. Name someone accountable. All correct, and all at the level of principle, which is the level at which the recommendation stops helping.
How far can this particular process run on its own? What control has to remain on my side? And who answers when it gets something wrong? No report answers that, because the answer changes with every process.
"The right step is not the highest one the technology can reach. It is the highest one the reach of the error can bear."
A system that writes a draft and a system that executes a half-million transaction cannot live under the same permission, the same monitoring and the same off switch. Written out like that, it looks obvious. It is the most common mistake there is, because no maturity ladder has a column for the size of the damage.
The method starts there. The reach of the error sets the autonomy ceiling of that specific process. And the control left on the human side has to follow that ceiling: the more irreversible the effect, the closer to the decision the human has to be.
It is not a single policy for the whole company. It is a yardstick applied case by case, and it gives back different answers for different processes inside the same area.
Autonomy is not an on-off switch. There are eight steps, grouped into four classes, and what separates them is how much of the work still runs through a person. The scale measures that, and only that: it does not measure who chooses the path, nor how capable the system is.
Equivalence with market vocabulary, August 2026: 1 chatbot · 2 copilot, inline AI, smart alert · 3 and 4 agent, agent mode · 5 agentic workflow · 6 autonomous agent · 7 multi-agent system · 8 orchestrator. The market calls “agent” things that sit at four different steps here, and the scale exists to undo that ambiguity.
1 · Conversation — question and answer, outside the workflow.
2 · Initiative to propose — it watches what you are doing and decides, on its own, that it is time to speak up. It does not execute: it proposes.
3 · Approved execution — it carries out complex tasks, with approval at every step.
4 · Delegated execution — it completes the flow on its own; the human reviews the result.
5 · Triggered execution — chained steps, starting from a trigger someone defined.
6 · Self-initiated execution — it decides on its own when to act, with no trigger defined by anyone.
7 · Multi-agent — several long-running agents in parallel, synchronized.
8 · Orchestration — a central instance delegates and consolidates the output of an entire team of subagents.
Framework base: Mike Taylor and Every, 2024-2025. Adaptation, grouping into classes and what comes next: Prædictor.
No maturity ladder has a column for the size of the damage. This is the column that was missing: classify how far the error reaches, and the autonomy ceiling of that process appears on its own.
Before setting the band, ask how long it would take for someone to notice. A reversible error that no one detects is, in practice, irreversible — it accumulates without triggering any correction. If the answer is "only if someone complained", the band goes up one step.
This is the most common classification error: calling a process internal because it runs on an internal system, when it decides about people. Employment, credit, health, and benefits are not R2 because the database is yours. They are external, because whoever is affected is outside.
A process can be formally at level 3, with human approval, and in fact operate as level 4. The difference is whether the signature was an examination or a rubber stamp — and it only shows on the day of the incident. This is what we call approval theater.
What is the reach of the error in this process, how much control is still left on your side, and who answers when it fails. Three questions, one answer per process.
What was measured, in whom, and who paid for the research. Three reputable sources have published, on the same question, numbers that differ by seventy-five percentage points. None of them lied.
Could the person who signs off on what the AI produced have produced it themselves? If they could not, the signature is not control. It is the appearance of control, and the difference only shows on the day of the incident.
It makes five things explicit before any autonomy — the same reason why, on the yardstick above, approving is not examining.
That is what a method is: a yardstick that answers per decision, not a principle that applies to everything and therefore decides nothing.
It was built on forty years in the executive chair at IBM, Mastercard, Johnson & Johnson, and Kenvue, and on the public AI adoption scale by Mike Taylor and Every. What it adds is the part that was missing: the ceiling conditional on the reach of the error, and the remaining control that has to go with it.
Yardstick published on . The four levels of reach, the conditional ceiling, and the remaining control are Prædictor's own extension of the public scale by Mike Taylor and Every.
The first step costs nothing: it is fifteen questions.
Apply the yardstick to my case
The Diagnostic is free and asks for nothing beyond your time. It returns where you stand today, and what changes from there.