AI agents for healthcare administration, with a hard line around clinical work

Q: How do you handle patient identifying data?

Minimum necessary per task, with redaction before anything leaves your infrastructure and several steps running entirely inside it. The specific boundary is written down before we build, not described afterwards.

Q: Our records system has no API. Is this possible?

Often, through a read replica or a scheduled export, and we would establish that during discovery. Sometimes it genuinely is not possible safely, and that is a real answer we are willing to give.

Q: Who approves the administrative writes?

Whoever is accountable for that record in your organisation. The gate routes to a role you define, and the approval is recorded against a named person. Practices with a single administrator usually route everything to one queue, and larger organisations split by record type, which is a configuration change rather than a rebuild.

There is a great deal of administrative work in healthcare that is repetitive, well defined and does not involve a clinical judgement. That is where an agent belongs. Everything on the other side of that line stays with the people qualified to make it, by design rather than by policy.

Book a technical call

Intakeforms and referrals
Verifycoverage and details
Prepareassemble the case
Approveaccountable staff
Updatewrite and log

Administrative steps run continuously. Anything touching a record or a patient stops for the person accountable for it.

The constraints that actually bind

The clinical boundary is not negotiable

An agent may assemble information for a clinician. It may not triage, advise or imply a clinical conclusion. We build this as an architectural boundary, not as an instruction in a prompt, because instructions are not a control.

Minimum necessary is a design constraint

Every step has to justify the patient data it touches. This shapes which model handles which task, what gets redacted before anything leaves your infrastructure, and which steps run entirely inside your own network.

Records systems are old and unforgiving

Many practice management and records systems have no meaningful API, and the integration surface is a scheduled export or a database replica. That is workable, but it changes what is possible and it has to be established before anyone promises a timeline.

What we build for healthcare teams

Administrative work that consumes clinical staff time without requiring clinical judgement.

Referral intake and completeness checking
The agent reads incoming referrals, checks whether the required information is present, and chases what is missing. Incomplete referrals get resolved before they reach a clinician rather than after.
Coverage and eligibility verification
Checking what a payer will cover, against the plan and the procedure, is exactly the kind of lookup that consumes hours and requires no judgement. The agent does it and reports what it found, with the source.
Appointment logistics and reminders
Scheduling, rescheduling and reminder sequences run automatically. Anything that changes a patient record or contacts a patient with something new passes a gate first.
Billing follow up and denial triage
Denials get read, categorised and matched against the likely cause, with the supporting documentation assembled. Your billing team decides how to respond rather than working out what happened.
Assembling the case for a person
When something needs a clinician or a manager, the agent gathers the history, the correspondence and the current state, and hands it over complete. It stops there, every time. The value is not the summary, it is that the person picking the case up does not have to reconstruct three weeks of context before they can make a decision.

Where the agent stops, and why it cannot continue

The clinical boundary is enforced the same way every other limit in our systems is enforced: the agent has no tool that crosses it. It cannot write to a clinical record, it cannot send a patient a clinical statement, and it cannot mark anything as triaged. Administrative writes stop at an approval gate for the person accountable for that record. This is deliberately conservative, and it is what makes the rest safe to run every day.

How approval gates work

What healthcare teams actually run

Integration is usually the hardest part of a healthcare build, and we scope it first.

Practice management systems
Records systems via replica or export
Payer portals
Scheduling systems
Secure messaging
SFTP and HL7 style feeds
PostgreSQL
Document stores

Where to go next

Other sectors

Questions about this

Will an agent ever give clinical advice?

No. It has no capability to do so, which is a stronger guarantee than a policy saying it should not. If a workflow would require crossing that line, we scope it out and say why.

How do you handle patient identifying data?

Our records system has no API. Is this possible?

Who approves the administrative writes?

Start with the admin nobody wants

Referral chasing, eligibility checks, denial triage. Pick one and we will scope it in two weeks.