An AI product ships an output. A draft, a recommendation, a summary, a decision. The dashboards say it worked: latency held, the model returned, the feature was used. None of those measures the thing that actually determines whether the product created value – whether the person on the other side can understand the output, exercise informed choice and judge when to rely on the system.
That is the gap in most product instrumentation. We measure whether the signal left the building. We do not measure whether it landed. The receiving end is not unstudied: human–computer interaction research and work on appropriate reliance on automation have examined it for years. That work rarely reaches the product dashboard.
Output is not the bottleneck
Generative systems have made output abundant. The constraint has moved downstream, to the receiver – the human inside the system who has to absorb what the machine produced, form meaning from it, and act. When that receiver is overloaded, mis-pitched, or unable to judge when reliance is warranted, the output is intact and the outcome still fails.
Four conditions at the receiving end
Participant Coherence concerns whether a person can receive information, make sense of it, exercise informed choice and rely on the system appropriately. It reads the downstream end of the chain across four conditions.
Signal Landing: relevant information and material uncertainty are accessible in a usable form.
Meaning Formation: the participant can understand the claim, its basis and its limits, and articulate a different interpretation. Understanding does not require agreement with the organisation or the model.
Action Coherence: the participant can act, seek clarification, defer or refuse through a legitimate route. Informed non-action can be a successful outcome.
Trust Calibration: confidence and actual reliance are proportionate to evidence of reliability, uncertainty, context and consequences. Trust may appropriately rise or fall.
These are conditions to investigate, not model metrics or a validated universal scale. The Coherence Lens – agency, reciprocity, alignment and signal integrity – provides the analytical dimensions; these conditions describe the receiving-end experience. Their relationship is many-to-many. Adaptation may improve these conditions, but it can also obscure evidence or constrain choice; its effects must be tested.
Why it belongs on the product
Operating-model coherence asks whether the organisation can hold what AI accelerates. Participant Coherence asks the same question one layer out: whether the product's output holds for the person who receives it. One substrate, two ends of the same chain. Build for output alone and you scale noise. Build for coherence and the product compounds.
The four conditions are developed further, with the evaluation logic that would test them and the prior work they draw on, in the position paper Making AI Ethics Executable (September 2026).