Notes on working capacity, recovery and how a coach that reads real signals behaves in practice. No case studies we cannot evidence, and no numbers we have not measured.
Long read
The full argument with the method written out: the signals, the baseline, the gate on the coaching model, and the evidence behind it.
Corporate wellness is dead. It has just not been buried yet. Three of the largest peer-reviewed evaluations ever run found no meaningful effect, and burnout kept rising while spend kept climbing. The problem was never that wellbeing does not matter. It is that perks cannot fix structural overload, and a dashboard nobody reads is not a coach.
The problem is not that we have too few coaches. The problem is that we have a million bad ones, none of them talking to each other.
Burnout is produced by working conditions, not by individual weakness, so the interventions that work are structural: reduce sustained load, protect recovery, restore control over the working day, and shorten the gap between a problem appearing and someone doing something about it. Surveys are too slow to help. The signals that matter, workload density, hours drift and recovery decline, are visible weeks earlier if anyone is looking.
Privacy by design means building a system so that it is structurally unable to misuse workforce signals, not writing a policy promising it will not. In practice this means individual data staying private by default, metadata rather than message content, and reporting to leaders that is technically limited to anonymised aggregates.
Not every wearable metric is useful for understanding a working day; sleep, heart rate variability and resting heart rate trends carry the most reliable signal, while step counts and generic activity scores carry very little. Consistency of wear and comparison against a person's own baseline matter more than which specific device someone uses.
A wellbeing app typically asks someone to log in and self-report, then offers general content such as meditations or articles. An AI performance coach reads a person's actual working signals continuously and offers specific, timely suggestions grounded in their real working day, aimed at performance rather than wellbeing in the abstract.
Sustained pressure usually shows up weeks before anyone would describe it as burnout, as small, compounding changes in sleep, working hours and calendar density against a person's own normal pattern. Catching this early is best framed as protecting performance and judgement, not as a wellbeing intervention.
It is not the total number of hours in meetings that erodes deep work, it is fragmentation: short gaps between commitments that are too small to do meaningful work in but too disruptive to ignore. Reading calendar structure, rather than just counting meeting hours, is what makes it possible to protect focus time before a week fills up.
Team capacity can be understood well from anonymised, aggregated patterns, such as overall meeting density or shifts in typical working hours, without ever exposing an individual's data to a manager. The technical separation between individual coaching and group reporting is what makes this trustworthy rather than intrusive.
Heart rate variability is a proxy for how recovered someone's nervous system is, and recovery level is closely linked to attention, judgement and emotional regulation at work. A single HRV number means little; what matters is the trend against a person's own baseline over weeks.
Agentic performance intelligence is software that notices changes in a person's own normal working pattern from consented signals, then offers a timely, specific suggestion rather than a static report. It differs from analytics dashboards because it acts, within limits it does not overstep, and it differs from wellbeing apps because it is grounded in someone's actual working day.
