A fleet of narrow agents holds a live model of your hospital: its units, flows, stock, books and outcomes. Each one watches its own corner against your own baselines, and none of them writes an order.
Ambient documentation writes the note while you talk, and every serious platform now does it, ours included. What none of them hold is a model of the institution those consults happen inside. Institutional agents do.
Every institution gets an entity graph built from what it already runs: departments, wards, theatres, service lines, stock, cohorts, and the paths patients and money take between them.
The agent measures that graph against the institution’s own rolling baselines, never a generic industry benchmark, and wakes only when something moves outside your own band. You can read and correct the model it holds of you. A model you cannot inspect is a model you cannot trust.
Schematic. Layouts are simplified, and every name, number and result shown is invented.
Not one assistant. A fleet of narrow agents, each with its own scope, its own wake condition and its own measure of success. Narrow is the point: an agent that watches one thing can be audited on it.
Diagnosis to order-set adherence, by service line and roster.
Order-set adherence drifts outside the service line’s band.
Antimicrobial and formulary use against your declared policy.
Restricted-class starts outrun their cultures.
Complication and readmission spread, adjusted for cohort.
One list’s complication rate leaves the cohort band.
ICD-10 completeness and specificity before the claim.
Unspecified codes cluster before claims go out.
Queues, wait percentiles, bed turnover, theatre utilisation.
A queue’s p90 wait breaches its weekday band.
Stock-outs ahead, expiry exposure, PO lead against consumption.
Projected cover days cross the reorder horizon.
Staffing against realised demand, by hour and by day.
Arrival peaks and staffing stop lining up.
Documented work against raised charges, before day end.
Documented work is missing from the day’s invoices.
Promised wait against actual, re-queues and no-shows.
The promised-to-actual gap widens beyond the band.
Result available to patient informed, against policy.
A critical result outlives the declared inform window.
Reviews booked before discharge, and kept after it.
Discharges leave without their review booked.
Medication changes and instructions a patient can act on.
Summaries ship without a medication-change section.
Gateway settlements and counter cash against the posted ledger.
A settlement lands that the ledger cannot explain.
GST returns and TDS against the trial balance, every window.
A filing window opens and the export does not tie out.
Invoice aging, insurer claims and advances, by payer.
An aging bucket swells past its usual share.
Drug licenses, Schedule H entries and consent artefacts.
A license or register entry nears expiry unrenewed.
The same clinical loop you saw on the product page, with the agent that watches each step attached to it. Nothing new to operate: the fleet reads the work your teams already do.
Token issued, queue live.
The record arrives read.
Drafted while you talk.
Signed and routed.
FIFO, expiry, register.
One invoice, posted.
Insights route by role and permission, so each person sees the corner they can act on and nobody drowns in someone else’s alerts.
Nine theatre cases carry no procedure line, all from one list.
The sepsis bundle was skipped in 6 of 19 admissions, all overnight.
One in five summaries leaves without a medication-change section.
Four SKUs stock out before the next PO lands.
Patients told twenty minutes waited a median of forty-eight.
Two gateway settlements this week land without matching ledger entries.
A critical result reaches you in hours, because the delay is measured.
No agent writes an order, a prescription or a diagnosis. Every output is a proposal carrying its evidence set, the exact queries and counts behind the claim, routed to the person whose call it is. Declining asks why, and the answer becomes part of what the agent knows about your institution.
Neutral by design. 92health is not an insurance funnel and not a pharma-sponsored nudge engine, so an agent optimises toward the objective you declare and nobody else’s.
Schematic. Layouts are simplified, and every name, number and result shown is invented.
Forty minutes with your own workflows. We will show the model your institution would get and the first proposals it would raise.