The Four A's · Authority
Decision Rights When the Model Recommends
How to assign decision rights when AI makes the recommendation.
Organizations reassure themselves that a human is always in the loop. But when a model recommends and a person clicks approve, the research is clear that the human often did not really decide. Accountability that exists on the org chart but not in the moment is not accountability at all.
Research Foundation
- Peer-reviewed human factors research
- Decision science
- Government and institutional research (NIST)
- Enterprise transformation field experience
- The Builders Build Framework
Key Takeaways
- Decades of human factors research document automation bias: people over-trust automated recommendations, miss the system's errors, and defer even when they should not (Parasuraman & Riley, 1997; Skitka et al., 1999).
- A human approver is not the same as a human decision. Without designed friction, oversight collapses into rubber-stamping, and accountability becomes a formality.
- The Four A's treat this as Authority: decision rights must specify not just who signs off, but what the human is genuinely deciding, and what evidence they must weigh before they can.
Automation bias is well established
Raja Parasuraman and Victor Riley's foundational review of humans and automation described how automation is misused through overreliance, and disused through distrust, depending on how it is designed and introduced. Linda Skitka and colleagues demonstrated automation bias directly: when a system offers a recommendation, people commit both errors of omission, missing problems the system did not flag, and errors of commission, following the system into mistakes a vigilant human would have caught. Mary Cummings extended this to high-stakes decision support, showing that the mere presence of a confident automated recommendation reshapes the human's judgment. The consistent finding is that a recommendation is not a neutral input. It anchors the decision.
Approval is not the same as deciding
Herbert Simon established that real decisions are made where the knowledge and the authority meet. When a model supplies the analysis and a human supplies only a signature, the knowledge and the authority have separated: the system knows, the human signs. Under time pressure and volume, the human in the loop becomes a human on the loop, and then a human out of the loop in everything but liability. The organization retains accountability on paper while the substance of the decision has quietly moved to the model. That gap is exactly where costly, unexamined errors live.
If the only thing the human adds is a click, the human is not the decision-maker. The model is, and no one is accountable for it.
The Four A's reading
The research establishes the phenomenon. The Four A's of Organizational Readiness provide the executive lens, and this is an Authority question. Authority asks who decides what, at what level, with what information. When a model recommends, the decision rights must specify what the human is actually deciding, not merely that they approve. That means designing the moment so the person must engage with the evidence, can see and is expected to challenge the recommendation, and owns the outcome in a way that is real. Attention matters, because an overloaded approver cannot scrutinize, and Adaptability matters, because the system and the humans must both learn from errors. But the core move is to define authority so that accountability survives the presence of the model.
Where the decision should sit
The remedy is structural. Match the level of the decision to the level of the knowledge, so authority is neither hoarded at the top nor abandoned, but placed where the detail actually lives.
Misaligned
Authority far above the knowledge
The decision travels up to people far from the detail, and returns slow and ill-fitting.
Aligned
Authority placed at the knowledge
The decision is made where the detail lives, and is fast and well-fitted.
Good decision architecture does not centralize or abandon authority. It matches the level of the decision to the level of the knowledge.
Evidence matrix
| Claim | Research | Field evidence | Four A's |
|---|---|---|---|
| People over-trust automated recommendations | Parasuraman & Riley (1997); Skitka et al. (1999) | Rubber-stamped approvals | Authority |
| Knowledge and authority separate under automation | Simon (1947); Cummings (2004) | Approver with no real say | Authority |
| Overloaded approvers cannot scrutinize | Cognitive load research | High-volume approval queues | Attention |
What executives should do
Design the decision, not just the workflow. Decide which recommendations may be automated outright, and for those that require a human, make the human genuinely decide: require engagement with the evidence, surface the system's uncertainty, and hold a real owner accountable for the outcome. Give approvers the attention budget to scrutinize rather than a queue that guarantees they cannot. Human in the loop is only meaningful if the human is actually in the decision.