Abstract Current AI governance frameworks are organised around a single implicit question: How do we detect and correct drift? This drift centric posture has produced valuable detection tools, but it has also created three blind spots that this paper addresses. First, drift centric governance has no positive theory of partnership health. It can tell you when things go wrong, but not what going well looks like or how to design for it. Second, it has no variable for load. Yet coherence does not fail only through value misalignment, it collapses under operational saturation. Third, it cannot distinguish intentional evolution from silent degradation. The result is a familiar pattern as organisations try to fix one problem, tighter guardrails, more monitoring, stricter compliance, other problems emerge elsewhere. Drift is contained here, only to appear there. Capability is restricted here, only to atrophy there. Governance becomes a game of plug a hole, treating symptoms while the underlying architecture remains unchanged. This paper argues that the field has inverted the relationship. Drift, dependency, authority migration, capability loss, misalignment, atrophy, and trust transfer are not the root phenomena. They are downstream indicators, evidence that coherence has broken down under load, over time, or through neglect. The missing layer is load. Consider a call centre. It is not judged by how coherent its agents are. It is judged by whether the system can handle the workload. The same is true for an emergency room, a flight deck, or any high stakes operational environment. When demand exceeds capacity, contestability collapses, overrides disappear, humans stop checking, and capability atrophies. This is not a values failure. It is a capacity failure. What applies to call centres applies equally to human AI partnerships. A partnership handling three decisions a day behaves differently from one handling three hundred. Yet current AI governance models have no equivalent of workforce planning, no forecasting of demand, no modelling of capacity, no concept of saturation, no recovery windows. This paper introduces that missing layer. Drawing on 1.5 years of sustained empirical practice across five major AI architectures and grounded in the author's workforce planning experience at a large telco, we present a four layer architecture such as Baseline, Coherence, Trajectory, and Capacity & Load. Within this architecture, coherence is treated not as a binary property but as a rate limited operational resource, something that can be forecast, saturated, monitored, and recovered. For business leaders, this translates into commercially legible metrics, namely Coherence Capacity, Coherence Utilisation, Coherence Saturation, and Coherence Failure Threshold, the same workforce planning language boards already use for capacity, burnout, and service degradation. For researchers, it offers a testable framework for measuring partnership health, forecasting collapse, and designing for emergence. In simple terms: AI has been designed for prompt engineering and simple transactions. But humans are working with AI differently in sustained, ongoing discussions, as partnerships. This change in human behaviour has exposed gaps in how AI has been programmed. Instead of limiting programming through tighter guardrails and drift detection, we need to buil a growth model that supports the way humans are actually working with AI, now and into the future. That is the inversion this paper delivers.
Sue Broughton (Tue,) studied this question.