Medical Transport AI Agent: Trip Requests and Approval Records
How to structure trip requests and approval records so an AI agent can route medical transport work while humans keep clinical and eligibility decisions.

An AI agent for a medical transport company can take in trip requests, check them against a structured record, propose pickup windows and draft confirmations, then route anything touching clinical need or patient eligibility to a named person. The design question is less about how much the agent can do and more about how each request and each approval is recorded, so the agent knows what it may touch and every decision leaves a trail. This article shows how to structure those records before any capable agent is given work to route.
Key takeaways
- Split every trip request into facts the agent may use, clinical and eligibility fields that only a named person may decide, and operational details the agent may propose.
- Give each approval its own record with an owner, a decision, a stated reason, the evidence referenced and a timestamp, so a later reader can see who decided what and why.
- Let the agent route, hold and draft, but keep medical necessity, eligibility and exceptions with trained staff.
- Scope each tool permission to the action a role actually needs, and review those permissions on a fixed schedule.
- Start with one trip type, check whether handoffs arrive complete, and widen the scope only after the records hold up.
Why trip requests need a boundary before automation
A medical transport request mixes logistics with clinical information. A discharge desk asking for a wheelchair van may include mobility details, a need for supplemental oxygen, a pickup window tied to an appointment, and a program or payer that decides whether the trip is covered. Experienced dispatchers handle this by reading the notes, calling the facility and checking the rules before they commit a vehicle.
An agent can do much of that reading and calling. The risk comes from treating the whole request as one block of text. When a free-text note says a patient can climb two steps, that sentence might be a clinician's assessment, a family member's guess or a copy error from an older form. If the agent cannot tell these apart, it will repeat the claim with the same confidence in every confirmation it sends. The boundary has to live in the record itself, not in a hope that the agent will be careful.
How to structure the trip request record
A trip request record should hold a small number of fields with clear sources. Each fact should name where it came from and who supplied it. A workable structure includes:
- Requester, organization and a verified callback path for changes
- Patient reference handled under the organization's privacy controls, kept out of free-text notes wherever possible
- Origin, destination and the appointment or discharge time that anchors the pickup window
- Service level requested, recorded as the requester stated it
- Mobility and equipment facts, each tagged with the clinician or facility that reported it and the date it was reported
- Eligibility status: pending, confirmed by staff, or declined, with the staff member's role and name
- Open questions the agent cannot answer, listed separately so they do not disappear into a summary
The important habit is separation. Facts go in one group, proposals in another, and decisions in a third. An agent reading this record knows it can build a schedule from the facts, suggest a window from the proposals and must stop at the decisions.
How to design the approval record
Approvals should be separate records, not a status word on the trip. An approval record links to the request, names the approval type (clinical necessity, eligibility, schedule exception, distance or vehicle exception), states the decision (approved, declined or returned for information), names the decider by role and name, cites the evidence it relied on and records a timestamp and any expiry date.
Treat approval records as append-only. If a decision was wrong or new information arrives, the correction should be a new record that references the old one. This matters for audits, for complaints and for the simple question of what the dispatcher knew at 7:40 in the morning. The agent can prepare an approval packet that summarizes the evidence, flags conflicts and recommends the next reviewer, but the decision field should be writable only by the human role that owns that approval type.
Duplicate confirmations are a common failure when agents retry actions. Each outbound action, such as sending a confirmation to a facility, should carry a unique key so that a retry does not create a second booking. The idea is the same one payment providers use for repeated API calls, described in their idempotency documentation.
Illustrative scenario: a hospital asks for a return trip after a surgery. The agent finds an approved window in the request record, proposes a pickup time and drafts the confirmation. It stops when the record shows that the mobility level was reported by a previous admission and not confirmed for this one, and it routes the approval packet to the clinical reviewer with that conflict marked.
Deciding what the agent may route alone
A decision guide keeps the boundary usable for staff who are not reading the architecture diagrams. Before you connect an agent to live requests, write down what it can do in each category:
- Propose a pickup window inside an already approved window: the agent may propose and confirm.
- Request a missing fact when the source document is already on file: the agent may draft the request for the requester.
- Change a mobility or service level: the agent routes to a clinical reviewer and does not decide.
- Determine eligibility or coverage: the agent assembles the evidence packet and routes to eligibility staff.
- Grant a distance, vehicle or cancellation exception: the agent routes to the dispatch supervisor.
- Encounter conflicting sources or unreadable notes: the agent holds the request and names the conflict.
The same discipline applies to scheduling tools the agent may use. If a staffing or timesheet process already requires supervisor sign-off for exceptions, the transport workflow should mirror it rather than invent a looser version. Teams that run this kind of approval chain elsewhere can compare their current rules with the approach in the staffing agency timesheet playbook, and the intake limits described for behavioral health clinic intake show a similar separation between scheduling and clinical judgment.
Permissions, audit trail and untrusted text
Most of the risk in an agent workflow sits in tools and inputs, not in the model's wording. The agent should hold credentials that allow only the actions in its decision guide. A confirmation tool should not also be able to edit an eligibility field. Access decisions should follow the principle of least privilege described in OWASP's authorization guidance, and permissions should be reviewed when routes, staff roles or facility contracts change.
Facility notes and emailed requests are untrusted text. A note may contain an instruction that the agent should follow, whether written by accident or by someone trying to manipulate the system. OWASP's prompt injection guidance recommends treating retrieved content as data and not as instructions, and keeping the agent's authority in the tool layer rather than in the text it reads. A request note that says "confirm eligibility" should produce a routed case, not a confirmation.
Log every agent action with the request identifier, the tool called, the input category, the outcome and the human who reviewed any escalation. These logs are what make a monthly permission review and an incident review possible. The NIST AI Risk Management Framework offers a useful structure for deciding who owns each risk, how it is measured and how it is responded to over time.
A first engagement, step by step
A sensible first engagement is narrow enough to check by hand. Work through these steps with the dispatch lead and the person who owns eligibility:
- List the trip request types you handle and choose the one with the clearest handoffs, often a recurring facility route.
- Write the record fields and mark each as a fact, a proposal or a decision.
- Assign the agent its permitted actions and name a human owner for every escalation category.
- Define one observable success criterion, such as every routed case arriving with all required fields filled and every conflict named.
- Run the agent in shadow mode against recent requests and compare its routing with what dispatchers actually did.
- Review incidents, tool permissions and record quality on a fixed schedule, and change one thing at a time.
Be Human's managed approach starts with one useful role, the memory and tools it needs, the approval boundaries and one success criterion that can be observed. After launch, managed operations can include monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, so the approval records stay accurate as staff, routes and payer rules change. The service is remote and available in English and Portuguese.
If you want to find the first task worth handing to an agent in your own transport operation, take the free Be Human AI diagnostic.
Frequently asked questions
Can an AI agent decide whether a medical transport trip is necessary? No, and it should not be configured to. The agent can gather the documents, summarize what they say, flag conflicts between sources and route the case to a clinical reviewer with a complete packet. The decision and its reason are recorded by that reviewer.
What fields matter most in a trip request record? Keep facts, proposals and decisions in separate groups, and tag every clinical or mobility fact with its source and date. Include the appointment or discharge time that anchors the pickup window, the service level as requested, and a list of open questions. Those fields let a second person understand the request without reading the whole thread.
How should approvals differ from trip status? A trip status describes where the booking is, while an approval record describes a decision someone made, who made it, on what evidence and until when. Approvals should be append-only, so corrections appear as new records rather than overwritten ones. This is what makes audits and complaint reviews possible.
How large should a first agent project be? Keep it to one trip type and one set of handoffs that staff already understand. Run the agent in shadow mode against recent requests, compare its routing with the dispatcher's, and widen the scope only when the records consistently arrive complete. Permissions and logs should be reviewed on a fixed schedule from the start.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


