The hospital that runs itself — so clinicians can run care.
MEDGRID is Boomlex's agent swarm for hospitals and health systems: specialist agents that move beds, notes, claims, and schedules under explicit clinician approval gates. The administration runs itself; the judgment stays human.
Physicians in ambulatory practice spend nearly two hours on EHR and desk work for every hour of direct patient care.
- HIPAA-aligned design
- Epic · Cerner · HL7 FHIR
- Clinician-in-command
- Audit-ready by default
Why hospitals & health systems need AI-native engineering
The modern health system has an allocation problem, and it is measurable. In direct observation of ambulatory physicians across four specialties, doctors spent 49.2% of the workday on EHR and desk work against 27% on direct clinical face time — nearly two hours of administration for every hour with patients (Sinsky et al., Annals of Internal Medicine, 2016). EHR event-log data tells the same story after hours: primary care physicians log roughly 86 minutes of nightly "pajama time" in the record (Arndt et al., Annals of Family Medicine, 2017). The clinical workforce is being spent on coordination the software was supposed to absorb.
Boomlex builds agent swarms that take over that coordination layer. Our agents read and write through the interfaces your hospital already runs on — HL7 v2 and FHIR feeds from the EHR, DICOM metadata from PACS, X12 837 and 835 transactions to and from your clearinghouse — and every consequential action passes an explicit human-approval gate. A bed is not reassigned, a claim is not filed, and a note is not signed without the named human owner saying yes. The swarm does the chasing; clinicians keep the judgment.
This page is written for the people who will be asked to defend the decision: the published evidence and where each figure comes from, the four-layer architecture and its guardrails agent by agent, a live console showing the swarm the way a house supervisor would see it, and the ten questions CMOs, CNOs, CFOs, and CIOs ask before signing. The short version — the first gated workflow goes live in six to ten weeks, with a full audit trail from day one.
Typical outcomes across programs of this scope — your baseline is measured in discovery, and results are reported against it.
What the research says
The case for taking administration off clinicians' hands does not rest on vendor claims. Two decades of time-and-motion studies, payer-survey series, and peer-reviewed waste analyses converge on one finding: the modern hospital spends an extraordinary share of its clinical and financial capacity on coordination work that software was supposed to absorb. The figures below are the ones we design against — each from an independent, published source, cited by name and year.
In direct observation of 57 ambulatory-practice physicians across four specialties, doctors spent 49.2% of the workday on EHR and desk work versus 27% on direct clinical face time — nearly two hours of administration for every hour with patients.
— Sinsky et al., Annals of Internal Medicine, 2016
Medscape's survey of more than 5,700 US physicians across 29+ specialties put burnout at 47% — down from 49% in 2024 and a 53% peak in 2023, but still nearly one in two practicing doctors.
— Medscape Physician Burnout & Depression Report, 2025
Premier's national survey of 280 hospitals across 23 states found providers spent $25.7 billion adjudicating denials in 2023 — up 23% year over year — with roughly $18 billion of it spent overturning denials that were ultimately paid.
— Premier Inc. national hospital survey, 2025
A 36-hospital time-and-motion study of 767 medical-surgical nurses found documentation consumed 35.3% of practice time — 147.5 minutes of a 10-hour shift, the single largest category of nurse time.
— Hendrich et al., The Permanente Journal, 2008
Across 41,256 ED admissions, in-hospital mortality rose from 2.5% for patients boarded under two hours to 4.5% for those boarded twelve hours or more. Boarding is a patient-safety problem, not just a throughput metric.
— Singer et al., Academic Emergency Medicine, 2011
The most rigorous peer-reviewed accounting of US healthcare waste identified administrative complexity as its single largest category — an estimated $265.6 billion per year, larger than pricing failure, fraud, or overtreatment.
— Shrank, Rogstad & Parekh, JAMA, 2019
Every figure above is independent, published research — not Boomlex marketing data. Methodologies differ across studies (direct observation, EHR event logs, provider surveys, transaction analyses), so we cite each source exactly and never combine figures across them.
References
- [1]Sinsky et al., Annals of Internal Medicine, 2016.
- [2]Medscape Physician Burnout & Depression Report, 2025.
- [3]Premier Inc. national hospital survey, 2025.
- [4]Hendrich et al., The Permanente Journal, 2008.
- [5]Singer et al., Academic Emergency Medicine, 2011.
- [6]Shrank, Rogstad & Parekh, JAMA, 2019.
- [7]Arndt et al., Annals of Family Medicine, 2017.
- [8]Experian Health, State of Claims, 2025.
- [9]Kodiak Solutions revenue-cycle benchmarking, 2025.
- [10]NEJM Catalyst, 2025.
What holds hospitals & health systems back today
Documentation is consuming the clinical day
Ambulatory physicians average nearly two hours of EHR and desk work for every hour of direct patient care (Annals of Internal Medicine, 2016), and event-log studies add roughly 86 minutes of after-hours charting nightly (Annals of Family Medicine, 2017). That time comes out of encounters, teaching, and sleep — and it resurfaces downstream as burnout, turnover, and agency staffing spend.
Denials are industrializing faster than appeals
41% of providers now say at least one in ten claims is initially denied (Experian Health, State of Claims, 2025), and initial denial rates averaged 11.8% in 2024 (Kodiak Solutions revenue-cycle benchmarking, 2025). Fighting back cost providers $25.7 billion in 2023 (Premier Inc., 2025), and most overturned denials clear only after an average of three payer review rounds of 45–60 days each. Payer automation is being answered with manual worklists.
ED boarding is a patient-safety problem
In a study of 41,256 ED admissions, in-hospital mortality rose from 2.5% for patients boarded under two hours to 4.5% at twelve hours or more (Academic Emergency Medicine, 2011). The beds usually exist — but discharge status, housekeeping, and transport live in different systems, and the path out of the ED is reconstructed by phone call.
The answer is spread across five systems
"Is this patient ready for discharge?" resolves across the EHR, the LIS, PACS, pharmacy, and a therapy note nobody has opened. The data exists — as HL7 v2 messages, FHIR resources, and DICOM headers — but no human sees the complete picture at the moment it matters, so decisions wait for rounds and length of stay stretches a day at a time.
Nursing time is going to paperwork, not patients
The canonical 36-hospital time-and-motion study found medical-surgical nurses spend 35.3% of practice time on documentation — 147.5 minutes of a 10-hour shift, the largest single category of nurse time (The Permanente Journal, 2008). Add census swings against schedules built weeks ahead, and charge nurses spend evenings texting for coverage while premium hours burn on quiet units.
AI systems built for hospitals & health systems
Patient-flow command center
Powered by: Hierarchical OrchestrationMEDGRID's flow orchestrator sits above specialist agents for beds, transport, housekeeping, and discharge planning. It consumes HL7 v2 ADT events — A01 admits, A02 transfers, A03 discharges — and unit census in real time, forecasts demand a shift ahead, and proposes moves ('clean 4-West beds now, stage two step-down transfers') that a house supervisor approves from one screen. Forty phone calls become a queue of ranked decisions, each with its reasoning attached.
Claims and denials swarm
Powered by: Sequential PipelineEvery claim runs a pipeline before it leaves the building: a Coder Agent checks documentation against CPT and ICD-10, an Eligibility Agent re-verifies coverage over 270/271 transactions, a Scrubber Agent applies payer-specific edit rules, and a submission agent files the 837 through your clearinghouse. When an 835 remit returns a denial, the Denial-Fighter Agent drafts the appeal with chart evidence attached and routes it to your revenue team for sign-off — drafted and filed in days, instead of sitting in a worklist for weeks.
Imaging triage and worklist prioritization
Powered by: Parallel Fan-OutWhen a study lands in PACS, fan-out agents examine its DICOM metadata simultaneously — one flags likely critical findings for radiologist attention, one checks protocol and image-quality compliance, one matches priors from earlier encounters. Results merge into a re-ranked reading worklist in seconds, so the suspected bleed surfaces above the routine follow-up. Radiologists read everything, in a smarter order; nothing is auto-diagnosed, and every ranking decision is logged for review.
Clinical documentation copilot
Powered by: Human-in-the-LoopAmbient documentation is the best-evidenced agentic workflow in medicine: 7,260 Kaiser Permanente physicians used AI scribes across 2.5 million-plus encounters, saving an estimated 15,791 hours of documentation time, with 84% reporting improved patient communication (NEJM Catalyst, 2025). MEDGRID's copilot follows the same discipline — the agent drafts from encounter context and structured EHR data, the clinician edits and signs, and the agent files with codes attached. The draft is never the record; the signature is.
Unified care-data brain
Powered by: Blackboard Shared MemoryA shared clinical blackboard aggregates what every agent can read and post: lab results from the LIS, imaging status from PACS, medications from pharmacy, vitals from monitoring, notes and orders as FHIR resources from the EHR. When bed 12's discharge criteria all read green, the blackboard says so once — and the flow, scheduling, and billing agents act on the same fact at the same moment, instead of discovering it hours apart.
MEDGRID — one brain for the whole hospital
Below is a live, self-running model of the coordination layer Boomlex deploys inside a hospital: one orchestrator brain at the center, specialist agents for emergency, imaging, labs, pharmacy, beds, staffing, revenue and supply — and a clinician approval gate in front of every consequential action. Nothing here is footage or a mockup; the simulation is rendering in your browser right now. Hover a domain to read its telemetry, click to open it.
3D icon = hospital domainOrb = patient journeyRing = clinician approval gateDot = agent-to-agent workFig. — Live simulation: an agent swarm coordinating a 340-bed hospital. Interactive — click any domain.
Inside MEDGRID: a four-layer architecture with guardrails
MEDGRID is organized as four layers with a strict contract between them: the layers that see and think are separated from the layer that acts, and the layer that acts is gated by named humans. Perception ingests the hospital's existing interfaces read-only. Reasoning & Planning turns live state into scored, explainable proposals. Gated Action executes only what an owner approves. Audit & Learning records everything and converts human corrections into better proposals.
The table below lists the specialist agents in a typical hospital deployment — what each one reads, what it writes, and the guardrail that bounds it. No agent writes clinical conclusions. Reads are scoped to the minimum fields the task requires; writes are drafts, flags, and queue entries until a human approves; and every row of this table is enforced in code and visible in the audit log, not documented in a policy binder.
Layer 1: Perception
See the hospital as it actually is, in real time.
Read-only listeners ingest HL7 v2 ADT feeds, FHIR resources, DICOM metadata, X12 835/837 streams, and bed-management events, normalizing them into a single live state model. Nothing is written at this layer — it only observes.
Layer 2: Reasoning & Planning
Turn state into ranked, explainable proposals.
Specialist agents evaluate live state against policies and forecasts — bed demand a shift ahead, denial risk before submission, discharge readiness — and produce proposals with evidence and reasoning attached. Every proposal is scored and prioritized before a human sees it.
Layer 3: Gated Action
Execute only what a named human has approved.
Consequential actions — a bed assignment, a claim submission, an appeal, a schedule change — pass approval gates owned by a named role. Approvals happen in one tap with reasoning attached; below-threshold confidence always escalates rather than guesses.
Layer 4: Audit & Learning
Record everything; improve from every correction.
Every observation, proposal, approval, edit, and rejection lands in an immutable log queryable by patient, agent, or hour. Clinician edits become training signal, so the swarm's proposals converge on what your teams actually approve.
Triage Agent
Emergency- Reads
- ADT feedtriage vitalsESI history
- Writes
- Ranked triage queuesurge flags
- Guardrail
- Never assigns acuity — recommends; the triage nurse confirms every level.
Bed-Flow Agent
Patient flow- Reads
- ADT eventshousekeeping statuspending discharge orders
- Writes
- Staged bed planshousekeeping dispatch requests
- Guardrail
- No bed is assigned without house-supervisor approval.
Discharge Agent
Care progression- Reads
- Blackboard discharge criteria — labstherapy notesmed-rec status
- Writes
- Readiness flagspending-task checklists
- Guardrail
- Flags readiness only; the discharge order requires physician co-sign.
Documentation Agent
Clinical documentation- Reads
- Encounter contextordersstructured EHR data
- Writes
- Redlined note drafts
- Guardrail
- Drafts never auto-file — the clinician edits and signs every note.
Coder Agent
Revenue cycle- Reads
- Signed notesCPT and ICD-10 rule sets
- Writes
- Proposed code setscoder queries
- Guardrail
- Codes are proposals; a certified coder approves before the claim is built.
Eligibility Agent
Revenue cycle- Reads
- Coverage on filepayer 270/271 responses
- Writes
- Verified eligibilityprior-auth requirement flags
- Guardrail
- Never contacts patients about coverage; discrepancies escalate to registration staff.
Denial-Fighter Agent
Revenue cycle- Reads
- 835 remitsdenial codeschart evidencepayer policy
- Writes
- Drafted appeals with evidence bundles
- Guardrail
- No appeal is submitted without revenue-cycle sign-off.
Imaging Triage Agent
Radiology- Reads
- DICOM metadatapriorsprotocol rules
- Writes
- Re-ranked reading worklistsQC flags
- Guardrail
- Never renders a finding — radiologists read every study.
Scheduler Agent
Access & capacity- Reads
- Clinic templateswaitlistscancellation events
- Writes
- Confirmationsbackfill offersreminders
- Guardrail
- Clinical questions escalate to staff instantly; no medical advice, ever.
Staffing Agent
Workforce- Reads
- Census forecastsschedulescredential and skill-mix data
- Writes
- Draft coverage plansfloat-pool requests
- Guardrail
- The charge nurse approves every schedule change before it posts.
PHI minimization by default
Agents receive only the fields their task requires — a scheduler sees appointment slots and contact preferences, never diagnoses. Access is mapped to your identity provider, scoped per agent, encrypted in transit and at rest, and every read is logged.
Private VPC or on-premises
The full stack — models included — runs inside your cloud tenancy or your own data center. PHI is never sent to consumer AI services. We work through your BAA, security questionnaire, and architecture review before any live data flows.
Immutable audit trails
Every observation, proposal, approval, edit, and rejection lands in an append-only log queryable by patient, agent, action, or hour. Designed for HIPAA-aligned deployments, the trail is built for your compliance team — and, when needed, your auditors.
Clinician approval gates
Human-in-the-loop is the architecture, not a setting. Consequential actions carry a named owner — house supervisor, charge nurse, revenue-cycle lead, attending — and below-confidence cases escalate rather than guess. Autonomy is earned per action class, with your sign-off.
- Weeks 1–8
Phase 1: Pilot ward
One unit, one or two workflows, full gates. Weeks 1–3 run read-only shadow mode against live feeds so owners see proposals without actions; weeks 4–8 go live gated, reported weekly against the pre-launch baseline your team helped set.
- 1–2 quarters
Phase 2: Service line
The architecture extends across a service line — more units, more payers, the full denial taxonomy. Action classes whose proposals are consistently approved unchanged become candidates for delegation; anything edited stays gated and feeds retraining.
- Thereafter
Phase 3: Enterprise
Multi-site rollout reuses the integrations, playbooks, and validated agents already built. Each site gets its own baseline, gates, and named owners, so scale never outruns governance — and system-level dashboards roll the sites up for the executive team.
Use cases we build for hospitals & health systems
The 2 a.m. bed crunch
Saturday, 2 a.m.: the ED holds eleven admitted patients and the house supervisor is calling units one by one. With MEDGRID live, the Bed-Flow Agent flagged the crunch building at 11 p.m. — three pending discharges never closed out, and housekeeping was never told. It stages the cleans, queues transport, and hands the supervisor a ranked plan: approve all, or adjust. Boarding that used to absorb a night shift resolves in under an hour, and the morning team inherits a unit, not a backlog.
Denials that fight back
A payer denies a three-day observation stay for 'insufficient documentation of medical necessity.' That denial used to sit in a worklist for three weeks. Now the Denial-Fighter Agent opens it in minutes: it pulls the attending's notes, vitals trends, and the payer's own published criteria, drafts a point-by-point appeal, and attaches the evidence. A revenue-cycle specialist tightens one paragraph and submits. The team always had the facts; what it lacked was the hours.
Notes that write themselves
A hospitalist finishes rounding on eighteen patients at 1 p.m. Under the old workflow she stays until 7 p.m. reconstructing notes from memory. With the documentation copilot, each encounter already has a structured draft waiting: history pulled forward, today's labs summarized, the hallway-dictated plan transcribed and coded. She reviews every note, corrects two, signs all eighteen, and leaves at 3:40. The notes are more complete than the ones she wrote exhausted — and the coders stop sending queries.
The no-show problem
A cardiology clinic runs 14% no-shows, and every empty slot is lost revenue and a patient getting sicker at home. The Scheduler Agent works the list around the clock: confirming by text in the patient's language, answering the questions that actually cause no-shows — parking, fasting, insurance — and backfilling cancellations from a waitlist it maintains itself. Anything clinical escalates to staff instantly. Patients get an answer at 9 p.m. on a Sunday; the clinic gets its slots back.
What changes for your leadership team
For the Chief Medical Officer
Boarding is a mortality problem — 2.5% rising to 4.5% as boarding passes twelve hours (Academic Emergency Medicine, 2011) — and documentation load is a quality problem. MEDGRID attacks both without touching clinical judgment: flow agents shorten the path out of the ED, and the documentation copilot returns hours to encounters, following the discipline of the largest published ambient-AI deployment (NEJM Catalyst, 2025). Every proposal carries its evidence, so your quality committee can inspect the reasoning, not just the outcome.
For the Chief Nursing Officer
Documentation is the single largest category of nursing time — 35.3% of med-surg nursing practice time, 147.5 minutes a shift (The Permanente Journal, 2008) — and status-chasing consumes much of the rest. MEDGRID's agents do the chasing: beds, transport, housekeeping, discharge tasks. Nothing is assigned over a nurse's head; charge nurses and house supervisors own the approval gates, and every proposal arrives with its reasoning attached. The goal is not fewer nurses — it is returning the nurses you have to patients.
For the Chief Financial Officer
Providers spent $25.7 billion adjudicating denials in 2023, roughly $18 billion of it overturning denials that were ultimately paid (Premier Inc., 2025). MEDGRID moves that fight upstream — eligibility verified before service, claims scrubbed against payer edits before submission, appeals drafted with evidence in days — and it is priced by workflow and site, not by token, so cost scales predictably. Where a workflow has a clean financial metric, part of the fee can sit against measured outcomes.
For the Chief Information Officer
No rip-and-replace. MEDGRID integrates through the interfaces you already operate — HL7 v2 feeds, FHIR APIs, DICOM/PACS metadata, X12 837/835 transactions — and runs in your private cloud tenancy or on-premises, with access mapped to your identity provider. Every agent's reads and writes are scoped, logged, and reviewable; model versions are pinned; and interface downtime puts the swarm into a safe hold. Your security review runs in parallel with the sandbox, not after it.
From first workshop to production AI
Step 1: Diagnose the flow
We shadow the workflows you want to fix — flow huddles, denial worklists, documentation time — and map every system they touch: EHR, LIS, PACS, clearinghouse, scheduling. You get a quantified baseline and a ranked list of automation candidates, scored by dollar impact and integration effort.
Step 2: Wire the sandbox
We connect read-only to your integration environment — HL7/FHIR interfaces, PACS metadata, claim files — and replay historical data through the first swarm. Your team sees real proposals against real cases before anything touches production, and your security review runs in parallel, not as an afterthought.
Step 3: Go live, gated
The first workflow ships with every action behind a human approval gate and full audit logging on. We deliberately start narrow — one unit, one payer, one clinic — so owners learn the swarm's behavior on familiar ground and every early decision is easy to inspect.
Step 4: Measure and earn autonomy
We report against the baseline weekly: boarding hours, denial overturn rate, documentation minutes, fill rate. Where the swarm's proposals are consistently approved unchanged, you may choose to delegate that action class; where they're edited, we retrain. Autonomy is earned per action, with your sign-off.
Step 5: Scale across the system
Once the first workflow holds its numbers for a full quarter, we extend the same architecture sideways — more units, more payers, more sites — reusing the integrations already built. Each addition gets its own baseline, gates, and owner, so growth never outruns governance.
Hospitals & Health Systems AI — common questions
How do you protect PHI and patient data?
Agents operate inside your network boundary or your cloud tenancy — PHI is not sent to consumer AI services. We apply PHI minimization (agents see only the fields their task requires), role-based access mapped to your existing identity provider, encryption in transit and at rest, and immutable audit logs of every data access. Deployments are designed for HIPAA-aligned operation, and we work through your BAA, security questionnaire, and compliance review before any live data flows.
Do you integrate with Epic, Cerner, and our existing systems?
Yes — via the interfaces those systems already expose: HL7 v2 feeds, FHIR APIs, CDS Hooks where available, DICOM/PACS metadata, and standard X12 transactions for claims. We do not require replacing or re-platforming your EHR. Where a vendor offers an app marketplace or sanctioned API program, we work within it; where only interface feeds exist, we build on those. Integration scope is confirmed in the discovery phase before we commit to a timeline.
Will AI agents make clinical decisions?
No. The swarm handles coordination and paperwork — moving information, drafting documents, ranking queues, chasing statuses. Anything that constitutes clinical judgment stays with clinicians: agents can surface a deteriorating trend or re-rank a radiology worklist, but they do not diagnose, order, or treat. Every consequential action passes a human approval gate, and the human-in-the-loop pattern is a structural part of the architecture, not a configuration option that can quietly drift off.
How do you prevent hallucinations from reaching a chart or a claim?
Three layers. First, grounding: every draft note, code proposal, and appeal must cite the specific source records it was built from, and statements without a source are blocked from the draft. Second, structured validation: codes are checked against CPT/ICD-10 rule sets and payer edits before a human ever sees them. Third, nothing consequential auto-files — a clinician signs every note, a certified coder approves every code set, a revenue-cycle specialist approves every appeal. A fabricated 'fact' has to defeat its evidence check and its human owner.
What happens when the EHR goes down?
MEDGRID treats interface silence as a first-class state. When HL7/FHIR feeds stop, agents enter a safe hold: no queued action fires against stale state, in-flight proposals are frozen and flagged, and owners are notified the swarm is paused. Your existing downtime procedures run exactly as they do today — the swarm never becomes a dependency for care. When interfaces return, agents replay the missed feed, reconcile deltas, and surface anything that changed during the gap for human review before resuming.
How are models validated and updated — and who carries oversight responsibility?
Models are validated against your historical data before deployment, and every model and prompt version is pinned and recorded in the audit trail. Updates ship through a change-control window with regression tests on your own cases, and material behavior changes are reviewed with your clinical informatics and governance leads before rollout. Accountability is structural: MEDGRID performs administrative coordination under named human owners, clinical judgment remains with credentialed clinicians, and we support your AI-governance committee with the documentation and monitoring it needs.
How fast do we see value?
Discovery and sandbox validation typically take four to six weeks; the first gated workflow goes live in six to ten weeks from kickoff. Because we start with a narrow, measurable slice — one denial category, one unit's flow, one clinic's scheduling — you see baseline-versus-actual numbers within the first month of live operation. Expansion decisions are then made on your data, not our projections.
How is this priced?
Two components: a fixed implementation engagement covering discovery, integration, and the first live workflow, then a platform subscription scaled to the number of active workflows and sites — not per message or per token, so costs stay predictable as volume grows. Where a workflow has a clean financial metric, such as denial recovery, we can structure part of the fee against measured outcomes. You get a full cost model in writing before signing.
Can this run on-premises, or does it require the cloud?
Both models are supported. Health systems with strict data-residency requirements run the full stack — models included — inside their own data center or private cloud tenancy. Others prefer a managed deployment in a dedicated, isolated cloud environment under a BAA. The agent architecture is identical in either case; what changes is where inference runs and who operates the infrastructure. We recommend the option your compliance posture and IT capacity actually support.
What happens when an agent is unsure?
It stops and asks. Confidence thresholds are set per action type: below threshold, the agent escalates to its human owner with the case, its reasoning, and what it would have done. Uncertain cases are never silently guessed — they are queued, answered by a person, and fed back as training signal so the same ambiguity is handled better next time. You can inspect the escalation log at any time to see exactly where the swarm's limits are.
Bring agent swarms to your operations
Tell us where the friction is across your hospitals & health systems and we'll map the AI Engineering Framework onto it — with a timeline, a cost estimate, and the agentic architecture we'd deploy.