AI in NEMT: what it does well today for dispatch, booking, and billing

Two kinds of AI with two different jobs

“AI” in NEMT software means two separate tools, and they fail in different ways. Keeping them apart makes it much easier to judge a vendor’s pitch.

Planning tools are the older and more predictable of the two. Their output can be checked line by line against hard rules. Assistants are newer and more flexible, and their main weakness has a name. NIST’s Generative AI Profile (NIST AI 600-1, July 2024) calls it confabulation: output that is stated with confidence and is simply false. Everything below follows from that difference.

Planning the day

Building tomorrow’s schedule is the clearest use of AI in NEMT. The rules are concrete, and the result is easy to check. A planner weighs every rule for every driver at once, which gets hard to do by hand after the first few vans.

A good planner works from the same constraints your dispatcher juggles:

• Pickup windows and appointment times. MTM’s Virginia Medicaid provider handbook (last updated July 30, 2026) gives the van a 15-minute grace period on each side of a trip’s first scheduled pickup. The rider has to be dropped off by the appointment time, or the trip counts as late.
• Vehicle fit. Wheelchair users ride only in lift or ramp vans. Stretcher trips go only to vehicles equipped to carry a stretcher. See level of service.
• Driver shifts. Start times, start locations, breaks, and end times.
• Ride-time limits. On a shared ride, the same handbook says no rider should spend over 45 minutes more on board than the average direct drive between their pickup and destination, unless that is specifically authorized.
• Empty miles. The unpaid driving after each drop-off until the following pickup, called deadhead miles.

Planning also means re-planning. Cancellations, add-ons, and returns change the day after it starts. MTM’s Virginia handbook gives the provider 45 minutes to reach a will-call rider after the ready call, and three hours when a hospital discharges a rider. A planner that can fit those calls into the running day saves the most frantic phone work. Our guide to route optimization explains how to judge the plans a tool produces.

Some decisions stay with you. Lock a regular rider to the driver they know. Accept a late arrival when the facility has agreed to it. And review each plan before it goes live, looking for late flags, long ride times, and a driver with a lopsided day.

Answering questions from live data

An assistant connected to your dispatch and billing records answers the questions an office asks all day, in seconds instead of three screens and a spreadsheet. Typical requests:

• Tomorrow’s trips that still have no driver
• Drivers running late right now, and which pickups that affects
• Facility invoices more than 30 days past due
• One rider’s trips last month, including no-shows and cancellations

Two things separate a useful assistant from a risky one. It should answer from your records rather than from general knowledge, and show or link the trips behind the answer. And it should respect the permissions of the person asking. A dispatcher who cannot open billing should not be able to get billing figures by asking the assistant. That matches the minimum necessary rule at 45 CFR 164.502(b), which asks you to limit health information to what each use requires.

Drafting messages

Drafting is low risk and saves real time, because a person reads every draft before it goes out. Good candidates:

• Reminder texts with the pickup time and place
• A note to a facility that a pickup is running late
• Payment reminders to facilities and private-pay riders
• A calm reply to a rider complaint
• A message to all drivers about a road closure

Keep medical details out of drafts. A reminder needs the time and the pickup spot, not why the rider is seeing the doctor. The minimum necessary rule applies to messages just as it does to screens.

AI voices on the phone carry an extra rule. In February 2024 the FCC ruled that AI-generated voices count as an artificial or prerecorded voice under the Telephone Consumer Protection Act. Automated calls to riders that use one need the rider’s prior express consent, unless an emergency purpose or an exemption applies.

Reading ride requests

Requests reach a NEMT office by fax, email, portal message, and voicemail. An assistant can read a discharge request and fill in the trip: rider name, pickup address, destination, time, mobility level, and how the rider gets back. That removes most of the retyping. The mistakes it makes are specific, so check for them:

• A handwritten “W/C” read as ambulatory, or the reverse
• The appointment time entered as the pickup time
• A missing return trip that nobody notices because the form looks complete
• Two riders merged because they share a last name

Set three rules. The assistant leaves a field blank and flags it instead of guessing. A person confirms the level of service, the times, and the return plan before the trip is booked. The original request stays attached to the trip. The details every booking needs are in our trip intake guide.

Billing: checking, not guessing

AI is useful in billing as a second pair of eyes. It is dangerous as a source of facts.

Where it helps:

1. Catching gaps before you invoice. Completed trips missing a signature, a drop-off time, or mileage.
2. Spotting trips that do not add up. MTM runs a similar check. MTM’s Virginia handbook scores providers on a “Miles vs. Minutes” measure, a fraud, waste, and abuse test that compares recorded trip time with mileage.
3. Following up. Drafting reminders for overdue facility and private-pay invoices.
4. Explaining denials. Translating a denial reason into plain words and pointing to the fields to check.

Where it must never act alone: filling in a missing pickup time, estimating mileage, or adding a signature. The invoice or claim has to match what happened on the road. The False Claims Act (31 U.S.C. 3729) defines “knowingly” to include acting in reckless disregard of whether a claim is true, with no proof of intent required. See NEMT claim denials and fraud prevention for the records that hold up.

A person approves every change

NIST’s profile names a second risk alongside confabulation: automation bias, meaning people defer too much to an automated system. The fix is a clear split between what AI may do alone and what waits for a person.

Every action should also be logged with the person who approved it. HIPAA’s Security Rule already requires audit controls that record activity in any system holding electronic health information, at 45 CFR 164.312(b), and an AI feature is part of that system.

HIPAA questions for any AI vendor

Rider names, addresses, destinations, and mobility needs are protected health information. Under HHS guidance on cloud services, a company that holds or processes that data for you qualifies as a business associate, and that holds even when the company has no way to view the data. Ask every vendor:

1. Does your business associate agreement cover the AI features? Get it in writing before any rider data goes in.
2. Which subcontractors process our data? Under 45 CFR 164.504(e), your vendor’s subcontractors must agree to the same restrictions it does.
3. Is our data used to train or improve models? The agreement sets the permitted uses, and the vendor may not use the data any other way.
4. Where is the data processed and stored? Broker contracts can be strict here. Modivcare’s 2025 provider attestation has transportation companies certify that they do no offshore work involving protected health information, including viewing, processing, or storing it outside the United States.
5. Does the assistant follow each user’s permissions? It should never see more than the person asking.
6. Is every AI action recorded? You want the prompt, the result, and the approver in the activity history.

For the full vendor checklist, see HIPAA-compliant NEMT software.

Measuring whether it pays off

Measure before you buy, then measure again during the trial. When the FTC announced its 2024 sweep against deceptive AI claims, it said there is no AI exemption from the laws already on the books. Treat a vendor’s savings figures as claims to test on your own trips.

Here is an example of the math. A six-van company that spends 90 minutes each evening building the next day, and gets that down to 20 minutes, saves 70 minutes a night. Over 20 working days that is about 23 hours of dispatcher time a month. Our NEMT KPIs guide explains the other numbers worth tracking.

AI in HealthRide

HealthRide includes both kinds of AI. Ryder Go plans your whole day in one click, pairing trips with drivers and vans by appointment and pickup times, driver locations, and which riders use a wheelchair, and by default you review the plan before the board changes. Ryder AI answers questions from your live trips and invoices, drafts messages, and books, assigns, or cancels trips only after you approve. It sees and does only what the signed-in person is allowed to, and every action is recorded.