For most of my career, judgement was never the constraint. There was always something slower in front of it.
Procurement was slower. Delivery was slower. Environments took weeks. Approvals queued behind other approvals. An architect could take three days to think about a decision because the decision could not move for three weeks anyway. The system was full of mechanical delay, and judgement lived comfortably in its shadow.
AI-assisted delivery has been steadily draining that shadow. The build is fast. The artefacts are fast. With runtime governance, even the controls are fast. And then you discover what was always standing at the back of the queue: a human being who has to decide, and who has not become any faster at deciding well.
The previous field note ended exactly here. Remove every mechanical bottleneck and the system starts surfacing decisions faster than the people inside it can responsibly make them. Judgement becomes the bottleneck.
I want to take that seriously rather than dramatically, because the usual responses to it are both wrong.
Two bad answers
The first bad answer is to speed up the human. Compress the review, skim the artefact, trust the summary, approve in the gap between meetings. This is how judgement degrades into throughput. The reviewer is still present, the approval still happens, but the thing the approval was supposed to contain, actual scrutiny by someone carrying actual accountability, has quietly left the room. AI makes this failure mode worse because the artefacts arriving for review are polished. Elegant nonsense survives a skim. It rarely survives a serious read, but a serious read is precisely what the accelerated tempo is pressuring you to abandon.
The second bad answer is to automate the judgement. If the human is the bottleneck, let the machine decide. Sometimes this is dressed in respectable language: risk-based auto-approval, confidence thresholds, human-on-the-loop rather than human-in-the-loop. I am not against any of those mechanisms. I use versions of them. But there is a category difference between automating a decision and automating the appearance of a decision having been governed. The machine can rank, filter, flag, and recommend. It cannot carry consequence. It will not sit in the incident review, face the client, or own the residual risk. Authority that cannot carry consequence is not authority. It is liability laundering.
Both bad answers share an assumption: that every decision deserves the same judgement. That assumption was always false, but slow systems never forced anyone to confront it. Fast systems do.
Designing the judgement budget
So the honest response, the one I am trying to build into my own practice, is to treat judgement as a finite, designable resource. A budget, in the same way the agents in my system run inside cost envelopes. The question stops being how do I review everything faster and becomes: what does this decision actually require, and from whom?
In practice that means stratifying decisions by what I think of as their blast radius and their reversibility. A reversible decision with a small blast radius can be governed by the constraint environment alone: if it goes wrong, the runtime controls contain it, and the telemetry tells me it went wrong. An irreversible decision, or one whose failure reaches a client, a contract, or a person's trust, gets the full human read, at human speed, regardless of what the tempo wants. The discipline is not reviewing less. It is refusing to spend scarce judgement on decisions the system can safely absorb, so there is judgement left for the ones it cannot.
This is where the loop from two notes ago quietly earns its keep. Cost, value, and integrity telemetry are not just operational hygiene. They are what makes triage defensible. Without signals, every decision looks the same size, and the reviewer drowns. With signals, the system itself can tell you where the blast radius actually is, and the architect can stand at the narrow point where judgement matters instead of being spread thin across everything that moves.
I notice the role changing shape again as I do this. The architect as control plane was about turning intent into executable structure. This is the next altitude: the architect as the designer of where human attention is mandatory. Deciding what may never be auto-approved. Deciding which categories of change always summon a person. Deciding, in advance and in writing, what the system is not allowed to decide. Those meta-decisions are the most consequential artefacts I now produce, and almost nobody outside the work would recognise them as architecture.
There is one more layer of honesty this note owes. The judgement budget is not just organisational. It is personal, and it depletes. A day of reviewing AI-generated artefacts is a strange new kind of cognitive labour: everything is plausible, everything is polished, and the cost of attention is paid on every paragraph because the one bad assumption looks exactly like the two hundred good ones. I finish those days more tired than I ever was producing the artefacts myself. Vigilance, it turns out, is more expensive than authorship.
Which means the constraint behind the constraint is not judgement in the abstract. It is the condition of the person doing the judging. Their clarity, their energy, their honesty with themselves about when their attention has degraded below the standard the decision deserves. No operating model I can draw fixes that. The loop, the worksite, the telemetry, the triage: all of it ultimately rests on a human mind being in a fit state to exercise authority.
I built an entire delivery ecosystem before admitting that its most fragile component was me.
That admission has a name in my system. The next field note turns the architecture inward, to Ordo Animi, and asks what decision governance looks like when the system under governance is your own mind.