For hospitals
The ward, the clinic, the pharmacy and the billing office are one patient
A hospital does not have a documentation problem, a pharmacy problem and a claims problem. It has one patient moving through departments that each hold a different version of them, and most of the day’s wasted effort is people reconciling those versions by hand. That is what a shared clinical record is for, and it is the reason the agents are built on top of one rather than beside it.
Six problems
What actually goes wrong between the ward and the bill
None of these are AI problems. They are coordination problems that a drafting model makes faster and worse if you put it on top without fixing them first.
The doctor writes it once for the chart and again for the discharge summary
Twice is optimistic — the referral letter and the claim narrative are usually a third and a fourth. Notes, summaries and discharge packets are drafted from the same encounter rather than retyped from it, and each one is a draft until the clinician responsible signs it. What is saved is the retyping, not the judgement.
Nobody knows who can go home until the consultant does the round
Discharge readiness is knowable earlier than it is known: the outstanding order, the pending result, the medication reconciliation nobody has signed. A flow agent watches those and flags the blocker by name, and a bed board reads the real ward hierarchy rather than a whiteboard someone updates when they remember.
The interaction warning is a thing everybody clicks through
A warning that fires on everything trains people to dismiss it. The interaction, allergy and dose-range engine here is deterministic and hard-blocks rather than warns, and it is not model output — which is what makes it defensible to block on. Its outputs are frozen by snapshot tests, so a rule change is a reviewed decision rather than a surprise on a ward.
Coding happens a week later, by somebody who was not in the room
The coding agent proposes in plain language at the time of the encounter, and the terminology adapter — never the model — resolves that into a code for a human to accept. The sequence is the safety property: proposals from an agent, codes from a code system, acceptance from a person, and the claim assembled from what that person confirmed.
The claim came back denied and nobody knows which charge caused it
Charges are computed by a deterministic rule engine from the encounter that produced them, pharmacy charges included, so a disputed line traces back to a clinical event rather than to a spreadsheet. Denials are categorised from the remittance and an appeal is drafted for a person to send, and charge-integrity flags are raised before the bill goes out rather than after it bounces.
Every department has its own system, and the integration is a person
The integration being a human being is the actual cost, and it is invisible on every budget line. Ambulatory, inpatient, orders, pharmacy and the revenue cycle sit on one clinical model, and because that model is stored in the exchange standard rather than translated into it, the interfaces to whatever you keep are a projection of the record instead of a nightly rebuild.
The floor, not a feature
Access control is architecture here, not configuration
A hospital is not one user, and a system where the ward clerk and the consultant see the same thing has either over-exposed one of them or under-served the other. Scoping sits in the data layer, underneath everything that could forget to ask for it.
- Facility scope on every clinical row, applied by a shared mixin rather than remembered per table.
- Role permissions per clinical role on top of that scope — a workspace is not merely hidden from the people who should not see it.
- A cross-facility read denied at the layer that builds the query, not in a handler somebody can forget to guard.
- Break-glass access that works when a patient needs it to, and that records who broke the glass and why.
- An immutable audit log, so “who opened this chart” is a question with an answer rather than an investigation.
An admission, end to end
Five steps, and the signature runs through all of them
- 01
Arrive
Registration and matching against the jurisdiction’s adapter, with any merge reversible and on the ledger.
- 02
Admit
A bed from the deterministic assignment rules, the movement written to the ledger, and the census derived from it.
- 03
Treat
Orders and results bound together, pharmacy checks that block rather than warn, and notes drafted alongside the work rather than after it.
- 04
Discharge
A packet assembled from the encounter, reconciliation presented for a pharmacist, and follow-up booked once a clinician has signed.
- 05
Settle
Charges computed from what happened, the claim assembled from confirmed codes, and the denial trail answered from the record.
Said plainly
The agents are the easy half
The drafting is not what makes this hard. Sharing one clinical record across departments, keeping the pharmacy rules deterministic, making a discharge summary reconcile to the encounter it came from and making a claim traceable to a clinical event — that is the work, and the agents are only safe to add once it is done. If your hospital does not look like the six problems above, tell us where this page is wrong about it; that is worth more to us than agreement.
Tell us where it is wrongRunning wards, clinics and a pharmacy at once?
This is built for the hospital holding one patient across a dozen departments — which also makes you the person best placed to tell us where it is wrong about your work.
Get in touch
Talk to the people building it
No chatbot and no ticket queue. Tell us what your hospital or practice actually looks like — how many beds, which departments, what the doctors are doing twice, what your billing team keeps chasing — and someone who works on the software will reply.
info@legosphere.com