Every hospital has the same three bottlenecks, and they feed each other: the phone nobody picks up, the schedule that never turns into visits, and the claim that comes back denied. One agent with memory and tools takes all three at once — because it is the same engine reading the same systems.
Patient service on Meta's official WhatsApp · booking written straight into the hospital's own schedule through the API · revenue cycle read every day by a Squad.
Each of them has a different owner on the org chart — the front desk, scheduling, and billing. But they share one origin: information that already exists in a system and does not reach the person who needed it in time.
The patient calls at the exact hour the front desk has a line at the counter. They wait, give up, try again tomorrow. When they write on WhatsApp, the message sits unanswered until someone catches a gap — and the gap usually opens after the visit was already lost.
The cost never shows up in any system: it is the appointment that was never booked.
The slot was published, someone booked it, nobody confirmed it, the patient did not show. The chair sat empty while a waiting list stood outside. At month-end the manager finds out that the specialty with the longest queue is the same one with the most holes in its schedule.
Idle capacity and a waiting list live in the same week, in the same hospital.
Care was delivered, the claim went out, and the health plan — the payer — returned its statement with lines denied. The appeal depends on pulling evidence out of three systems, and the contractual clock runs while somebody hunts for the right document.
When the deadline passes, the argument is over — and the loss stopped being recoverable.
The conversation that books the appointment is the conversation that confirms it the day before — and that confirmation is what prevents the no-show. The no-show is what leaves the schedule idle. The idle schedule is what squeezes the month's revenue. Splitting this into three projects with three vendors is exactly what makes each vendor solve its slice while the hospital keeps the whole problem.
At Hal-AI, the agent that answers the patient on WhatsApp, the Squad that sweeps the schedule before dawn and the Squad that reconciles the payer statement are the same piece of software holding different tools. Same memory, same audit trail, same dashboard.
This is not a chatbot at the front desk plus a robot in billing. It is one agent with declared access to the systems the hospital already runs.
From outside, it is the patient who cannot get an appointment. From inside, it is the manager who only finds the hole once the month has closed. Both stories happen on the same day, and nobody ever sees them side by side.
It is eleven at night. The patient just got off a shift and wants to book a cardiology follow-up. They write to the hospital's number, because a phone call at that hour gets nowhere. They want three things: whether their health plan is accepted, whether they need to bring any paperwork, and the earliest slot that exists.
None of those questions is clinical. All three already have an answer inside the hospital's systems at this exact minute. What is missing is not information — it is somebody awake to look it up.
Look at the last reply: a clinical question is not answered by the agent — it is handed to a person.
The manager does not have a data problem. The manager has a timing problem. The schedule-occupancy report lands after the month is closed. The denials spreadsheet is assembled by hand from statements that arrive in different formats, from different payers. The no-show report exists, and it explains the past.
What nobody delivers is the short list: these four specialties will end the week with idle slots, these batches hit the appeal deadline in five days, these twenty bookings look like the ones people skip. That is not a report — it is pending work, and it has to arrive earlier.
The same query that used to feed the monthly report now runs every day, before dawn, when nobody is using the system. The output is not a file on a server: it is a document with sections, tables and charts, delivered by email to a closed list of recipients.
It is the same piece that already publishes operations reports across the platform — here it reads the hospital's systems.
Patient service and booking on official WhatsApp. Schedule, capacity and no-shows. Revenue cycle, from billing to claims transmission. Each front has its own page, and all of them share the same memory and the same integrations.
Every page in this section describes the same reference hospital: a mid-to-large hospital group, with 28 outpatient specialties, roughly 60,000 outpatients a year, around 2,500 appointment slots published per week, an official WhatsApp number carrying about 23,000 conversations a month, and 14 contracted payers.
If your hospital is bigger or smaller than that, the order of magnitude changes — the shape of the operation does not.
Meta's official number answered by agents with memory: health plan coverage, exam prep, paperwork, hours and directions — plus booking, rescheduling and cancellation done inside the conversation, with no form to fill in.
From published slot to completed visit: where the schedule failed to convert, how far ahead a booking predicts a no-show, which specialty has a queue and which has a hole — with the day's funnel written before the shift begins.
The claim audited against the patient chart before the batch goes out: a billed line the documentation does not support, a missing record, a justification that does not match what was charged — the denial that never happens. On the other side, payer statements reconciled the day they land, denials classified by root cause and appeals assembled with the evidence.
The difference is not how good the text on screen looks. It is who decides what to do at each turn of the conversation — and what happens when the patient says something nobody drew into the design.
The agent has implicit access to nothing. Every endpoint it may call is registered in its configuration, with authentication, description and scope. What was not handed over simply does not exist in that agent's world.
Reading and writing are separate permissions. A patient-service agent can look up the schedule without being able to touch billing.
“I need to move next week's appointment to the afternoon, preferably with the same doctor” is not a menu option. For the agent it is a sequence of decisions: identify the patient, find the booking, check that provider's availability, propose, confirm and write it back.
When the request goes past what it knows, it does not improvise: it hands off to the front desk with the history attached, and the card carries an owner and a timestamp.
A flow is a map of the paths someone predicted. An agent is a professional with an objective, a manual and a keyring.
What changes from one role to the next is the personality, the Knowledge Vault and the keyring of tools. The memory and context architecture — the subject of the patent filed with the USPTO — is identical in all three.
Channel agents sit on official WhatsApp, webchat and voice, talking to patients. Squads have no channel of their own: they wake up on schedule, sweep the hospital's systems, decide what needs doing, and talk to the channel agents so those agents make the contact one by one. HAL Copilot never talks to a patient: it is the internal assistant for whoever runs the place.
None of this is file-based integration: all three read the same APIs, write to the same audit trail and show up in the same console. When the pre-dawn Squad finds that twenty of Thursday's bookings look like no-shows, what reaches the patient is not a dialer — it is the channel agent, inside the conversation the patient already had.
The conversation between the orchestrator and the channel agents is recorded too. It is not a dispatch queue: the Squad asks in plain language, the channel agent reports what it did and what it did not do — and the reason goes into the trail.
Name, personality, voice and real tools: look up the schedule, book, send media and interactive options, request a location, search the Knowledge Vault and hand off to a human. Several agents can share the same official number.
An orchestrator agent with no channel of its own. It wakes up on schedule, queries the hospital's APIs, decides, publishes the analytical document in HTML and PDF and emails it to a closed list. It has an auditable history, dry-run mode and a stop button.
The internal assistant for the board and for middle management: ask in plain language about your own operation, get a table and a chart on screen, export to PDF. The manager sets the mission, closes the page and gets notified when it is done.
The hospital's number is activated through Meta's own authorization flow. Hal-AI is a Meta Business Partner and runs on the official WhatsApp Business Cloud API — there is no handset taped to a desk and no unofficial number. Outside the 24-hour window, the only way to speak is an approved template, and that applies to the agent exactly as it applies to a person.
Below is a representation of the document the Squad publishes before dawn and delivers by email. The stylesheet is light and serifed on purpose: the deliverable is not supposed to look like a dashboard, it is supposed to look like a report.
Hospital Daily Dashboard · run #312
Document published by the Squad at the end of the 6:10 AM run, covering this week's outpatient activity and the open billing cycle.
The official number took roughly 23,000 conversations this month, 78% of which ended without anyone from the team entering the chat. The outpatient schedule published 2,500 slots this week and closed at 86% occupancy. In the open billing cycle, the denial rate sits at 4.1% of billed value, against 9.8% at the start of the cycle.
The front desk works from 7:00 AM to 7:00 PM. Adding up the bands outside that range — overnight, evening and late night — 41% of conversations arrive when there is nobody to answer: 8% + 20% + 13%.
Figure 1 — conversations on the official number by time of day. Within hours: 14% + 17% + 16% + 12% = 59%. Outside: 8% + 20% + 13% = 41%.
Of the 2,500 slots published this week, 2,150 were booked — that is the 86% occupancy. Against those bookings, the 24% no-show rate accounts for 516 absences, leaving 1,634 completed visits.
Figure 2 — the week's funnel. 2,500 published × 86% = 2,150 booked. 2,150 × 24% = 516 no-shows. 2,150 − 516 = 1,634 completed.
| Front | Indicator | Start of cycle | Current cycle | State |
|---|---|---|---|---|
| Patient service | Conversations closed without the team | — | 78% | on target |
| Patient service | Handoffs to the front desk | baseline | −76% | on target |
| Schedule | Outpatient schedule occupancy | 74% | 86% | on target |
| Schedule | No-show rate against bookings | 35% | 24% | above target |
| Revenue | Denial rate over billed value | 9.8% | 4.1% | above target |
| Revenue | Billing submission window | 12 days | 2 days | on target |
| Revenue | Batches rejected on technical inconsistency | 14% | 0.8% | on target |
| Indicators on the dashboard — 2 patient service + 2 schedule + 3 revenue | 7 | 2 above target | ||
The header names the routine and the run that produced the piece, and the monospaced line declares which API call each block came from. Every figure carries the arithmetic that supports it: the week's funnel shows the operands, not only the result.
The last table pulls the three fronts into a single read, and the block that closes the document separates what the Squad already handled from what depends on a person deciding.
Nobody exported a spreadsheet, cross-checked a list or wrote a summary. The document went out at 6:10 AM, before the day began, and reached a closed list of recipients by email — with the open item already named, instead of an attachment for someone to interpret.
Because the structure is identical on every run, today's dashboard can be compared with last week's without anyone redoing the math to line the two up.
Notice what the piece does not show: no clinical data and no identified patient. The aggregate dashboard works with counts, proportions and states — symptoms, diagnoses and test results never enter, because the configuration does not reach them.
And the arithmetic does not run through the model: tables and charts are declared by reference to the API call the agent already made in that same turn. The server reads the response and assembles the block. An empty document is refused, rather than shipping with a pretty axis and nothing inside.
The same document comes back into the run's chat as a clickable attachment — that is how the manager opens the piece without leaving the conversation with the orchestrator.
The published document as a bubble, the way it appears in the Squad's run chat.
Eight pages, each one with the mechanism open: which API is called, what happens when it fails, and who gets told. Start with whatever hurts today.
Patient questions, exam prep, health plan coverage, paperwork, directions and handoff — 24 hours a day, on Meta's official number.
AccessBook, reschedule, cancel and confirm in the official schedule through the HIS — the hospital information system — with a waiting list and canceled-slot recovery.
CapacityThe day's funnel, schedule occupancy, no-shows, lead time as a predictor of absence, and capacity redistribution by specialty.
RevenueStatement reconciliation, denied lines classified by root cause, appeals assembled with evidence, and an alert before the contractual deadline expires.
RevenueValidation before submission, one batch per payer, the protocol imported on its own, and the billing window back under control.
RevenueThe patient chart audited against the claim: a progress note that never records the billed procedure, a supply with no justification, a bed day with no matching note. With CBHPM, Simpro and Brasíndice — Brazil's reference price tables — kept current and edits simulated before the batch goes out.
ManagementThe Squad that wakes up, reads the systems, writes the daily dashboard in PDF and HTML and sends it — with the arithmetic shown behind every number.
FoundationHIS, ERP, TISS, HL7/FHIR, REST APIs and MCP; what the agent may do on its own; LGPD — Brazil's data protection law — audit trail and human takeover.
Smaller operationsIf this is a clinic, a private practice or a dental office, the path is the virtual receptionist on WhatsApp — quicker to turn on.
Patient service and booking on WhatsApp show up fastest: the patient feels it in the first week, the front desk feels it in the first month, and the schedule starts filling before any billing project finishes. The revenue cycle comes next, once the read integrations are standing and the audit trail is already being generated.
It is not a rule. A hospital with high denials and a healthy schedule starts from the other end — and the platform is the same in both cases.
A hospital needs clear limits, and it is better to say this before the contract than after. These limits are not settings: they are how the platform is designed.
The agent does not assess symptoms, suggest a diagnosis, recommend treatment, weigh in on medication or interpret test results. It handles access and the claim: it books, confirms, reschedules, repeats what the hospital has published and reads the administrative systems. Anything clinical goes to the care team with the context of the conversation attached.
Every booking, reschedule and cancellation lands in the hospital's official schedule, through the management system's API. There is no intermediate spreadsheet, no mirror database and no “the front desk will confirm later”. If the API does not answer, the agent says it could not do it and hands off — it never promises a slot it did not write.
The Squad audits the claim against the patient chart, points at the line the documentation does not support, classifies each denied line by root cause and assembles the appeal with the evidence attached — the work of finding it arrives done, with the line, the missing evidence, the likely cause and the deadline. What it does not do is accept a denial, write off a loss or change a billed amount on its own: the hospital signs the correction. Every action is recorded with what was queried, what was decided and what was written — and any action that moves money requires human approval declared in the configuration.
Once 24 hours have passed since the patient's last message, official WhatsApp only allows an approved template. The agent respects that: it does not improvise a free-form message, it uses the registered template or leaves the contact pending in the queue for the next day. That is Meta's rule, not a Hal-AI limitation — and it applies to the agent exactly as it applies to a person.
When a system does not answer, the run records the failure and the document goes out saying what was left out. An incomplete document is refused before publishing, instead of shipping with the gap disguised. While an integration is not connected, the screen itself stamps the data as a simulation, rather than pretending to be live.
The six that come up in every first conversation with clinical leadership, IT and billing.
The data stays in the hospital's systems. The agent queries it through the API at the moment of the conversation and uses the answer for that conversation — it does not copy your database out or build a parallel registry. What the platform keeps is the conversation history and the memory of that contact, with long-term memory switchable per agent.
Health data is sensitive personal data under LGPD. Any access to a patient chart or a clinical system depends on a specific project, with purpose, legal basis, scope and responsibilities set in the contract before anything is connected. On the aggregate dashboard, patient data does not appear at all: the manager sees counts, proportions and states.
No. Hal-AI does not replace the HIS, the ERP or your billing system: it talks to them. The REST APIs the hospital already has become tools for the agent — the endpoint is registered once, with authentication and a description, and from then on it exists on that agent's keyring. There is also support for MCP servers and for the industry standards, such as TISS and HL7/FHIR.
When a system exposes no API for some operation, that operation stays out of scope and is said plainly during project design, instead of becoming a promise that only surfaces at go-live.
The agent has no invented fallback. If the availability call fails, it does not offer a slot: it tells the patient it could not reach the schedule right now, records the request and hands the card to the front desk with the full history. If writing the booking fails, it confirms nothing — the rule is to never promise a slot that was not written.
On the routine side, the Squad's run records the failure and the published document states what was left out of the period. An incomplete piece is refused before publishing, instead of shipping with the gap disguised.
It depends on two things, and neither of them is training the agent: activating the official number through Meta's authorization flow, and access to the hospital's APIs. The agent's starting knowledge comes from what the hospital already publishes — specialties, accepted health plans, addresses, hours — and is completed in the Knowledge Vault with PDFs, text and images.
In practice, the design usually starts with a single front, read-only, in dry-run mode, before any write action is released. That is why the first conversation is about integration and governance, not about prompts.
The hospital, the same way it answers today for what the front desk does — and that is exactly why the platform is designed so errors are visible and reversible. Every run has an audit trail: what was queried, what was decided and what was written, with timestamps. Every conversation stays in the history. Every human takeover is recorded with an owner and a time.
Before any of that, there is the limit itself: the agent may only call the tools declared for it, and an action that changes the schedule or moves money can require human approval. A content error is fixed in the Knowledge Vault and holds for the next patient; a scope error is fixed by taking the tool off the keyring.
The call center needed to carry that volume gets considerably smaller. Repeating opening hours, listing accepted health plans, booking and calling people one by one to confirm attendance all get settled inside the conversation, with nobody from the team stepping in — and the patient is answered faster, in the middle of the night included. What is left for a person is the case that takes judgment: the angry patient, the contractual exception, the negotiation with the payer, the reschedule that takes a decision.
Human takeover exists for that: anyone at the front desk can take any conversation at any moment, and the agent goes quiet in that chat until the card is handed back. On the billing side the effect is the same: the team needed to check the same volume gets smaller, and doubling the number of claims stops meaning doubling the team that checks them.
Tell us which of the three fronts is stuck — the phone, the schedule or the denied claim. We'll design the agent that works on it with you, with the integrations and the governance your case demands.