Medical-malpractice intake: causation screening at the call, before the consult · NeuraVoice

A daughter calls a med-mal firm at 9:14 on a Tuesday morning. Her mother had a hernia repair eleven months ago. The mesh eroded. There were two follow-up surgeries, one of them emergent. The mother is home now, partial bowel resection, on a feeding regimen, can no longer drive. The daughter is the one who's been keeping the folder of discharge summaries.

She has roughly four minutes of patience for the intake call before she will hang up and dial the next firm on the Google results page. In those four minutes, the AI voice agent has to capture enough of the causation chain that the case-evaluation lawyer, when she opens the file at 11:00, can decide in under fifteen minutes whether to schedule a paid consult, route to the mass-tort queue, or send a courteous decline letter.

This is not a personal-injury intake with a different label. Medical-malpractice intake is its own discipline. The schema is different, the statute-of-limitations math is different, and the gating questions that determine whether the firm can even file are different. Nothing below is legal advice. It is a working register of the fields a med-mal AI intake has to capture, written from the operations side.

Causation screening has to happen at the call, not at the consult

A general PI intake can survive on three vectors: liability, damages, and statute. The lawyer evaluates causation at the consult, often after pulling a police report.

Med-mal does not have that luxury. Case development is expensive. Records requests run thirty to sixty days. An expert review can run two to ten thousand dollars before the firm files anything. Every hour spent evaluating a matter that turns out to be a bad outcome rather than a deviation from the standard of care is a hard sunk cost.

So the AI has to compress the causation analysis into the call itself. Four fields, in plain language:

What was the medical situation that brought the patient to the provider.
What did the provider do, or fail to do, that the caller believes was wrong.
What harm resulted, and is it ongoing.
What documents does the caller have, or know exists, that show the chain.

If any one of those four cannot be articulated by the caller in the first pass, the AI flags the matter as "causation incomplete" and routes it to a human intake specialist for a second call rather than to the case-evaluation queue. That single flag, applied honestly, saves the firm dozens of unbilled lawyer hours per month.

The SOL clock starts somewhere, and the AI has to ask which somewhere

Med-mal statute of limitations is, in many states, shorter than general PI. One to two years from discovery is a common range, with tolling for minors, for foreign-object cases, for continuing-treatment doctrine, and for the discovery rule itself. Some states cap the outer boundary with a statute of repose that runs from the date of the act regardless of when discovery happened.

The AI cannot apply the rule. It can capture the inputs the lawyer needs:

Date the procedure or treatment occurred.
Date the caller first became aware something was wrong.
Date of the most recent related treatment by the same provider.
Whether the patient is a minor.
Whether the patient is deceased and if so, the date of death.

Those five timestamps drive the SOL math for almost every fact pattern. If the AI captures them and the case-evaluation lawyer sees a discovery date eleven months ago in a one-year-from-discovery state, the file moves to the top of the queue that morning. The same case captured without the discovery date sits in the queue for two days while the firm calls the caller back to ask one question.

The treating-providers chain is rarely one provider

A general PI matter usually has one defendant driver. A med-mal matter routinely has four to seven medical actors involved in the same chain of care: a primary-care physician, a referring specialist, the surgeon, the hospital where the surgery happened, the anesthesiologist, the post-op nursing staff, and a follow-up provider who first noticed the problem.

The AI has to capture each of them as a separate field, not as a single "doctor's name" string. The reason is that vicarious liability is different for hospital employees than for credentialed physicians who happen to operate at that hospital, and different again for nursing-home staff. The case-evaluation lawyer has to know whether the surgeon was an employee of the hospital or an independent contractor with privileges before she can scope the defendant list.

A practical schema:

Primary care provider, name and clinic.
Referring specialist, name and clinic.
Procedure provider, name and facility.
Facility where the procedure occurred (hospital name, surgery center, clinic).
Post-procedure providers in the chain of care.
Provider who first identified the harm.

A good intake AI prompts for each slot in turn. A bad one asks "who was the doctor" and lets the caller pick whichever name comes to mind.

Records status determines which intake queue the case goes to

If the caller has already requested and received medical records, the file goes to the case-evaluation lawyer that week. If the caller has not, the firm's first task on the matter is a HIPAA-authorization request to each provider in the chain, which typically takes thirty to sixty days to complete and which an intake paralegal handles, not a lawyer.

These are two different work queues. Routing a no-records-yet case to the lawyer wastes the lawyer's time. Routing a records-in-hand case to the paralegal queue delays the evaluation by two weeks.

The AI has to capture, explicitly:

Whether the caller has any medical records in hand.
Whether the caller has submitted records requests and is waiting.
Whether the caller has been denied records (a separate problem).
Which providers in the chain the caller has not yet contacted.

That single field, "records status," determines the next-action assignment for ninety percent of viable matters.

Expert availability is a gate, not an afterthought

Most states require an expert affidavit, sometimes called a certificate of merit, before a medical-malpractice complaint can be filed. The exact name and procedural posture varies (Texas Chapter 74 expert reports, New York CPLR 3012-a certificates, Pennsylvania Rule 1042.3 certificates of merit, and so on), but the gate is the same: the firm needs a same-specialty expert willing to review the records and sign.

A firm that handles obstetrics cases has a network of OB experts. That same firm may have no neurology expert on call. If the caller's matter is a missed stroke diagnosis in an emergency department, the firm has to either decline the case or spend two weeks finding a neurologist who will review.

The AI cannot evaluate expert availability. It can capture the medical specialty involved, in plain English, and the case-evaluation lawyer can match against the firm's known expert network in the case-management system. The field to capture is something like "primary medical specialty implicated," and the AI should use the caller's words rather than try to taxonomize on the call. "The ER doctor missed a stroke" is more useful in the file than "neurology / emergency medicine."

Devices flip the case into a different schema entirely

If the matter involves a medical device, transvaginal mesh, hernia mesh, IVC filters, hip implants, certain surgical staplers, the analysis crosses into product liability and often into existing mass-tort consolidations. A firm that runs a med-mal practice and a mass-tort practice will route these to a different team and a different schema. A firm that does only med-mal will refer them out.

The AI has to detect device involvement and capture:

Was a device implanted, used, or prescribed.
Manufacturer and model if the caller knows.
Whether the caller has received any communication about a recall or MDL.
Whether the device was removed, and if so, was it preserved.

The last field matters. Preserved explanted devices are evidence. A device that was discarded by the hospital is a different posture for the case.

The hernia-mesh scenario at the top of this post is a likely device matter. A good intake schema picks that up and routes to the mass-tort queue rather than to the med-mal individual-case queue. A bad schema treats it as a surgical-malpractice case and the firm spends two weeks evaluating it at the wrong angle.

Hospital, physician, nursing home are three different liability postures

Vicarious liability runs differently for each:

Hospital employees (some staff RNs, some hospitalists) generally trigger respondeat superior against the hospital.
Privileged physicians who operate at the hospital are usually independent contractors, with apparent-agency arguments available in some fact patterns.
Nursing-home cases often involve corporate-defendant chains with separate state regulatory frameworks (federal nursing-home reform act, state survey records, and so on).

The AI has to capture, at minimum, the type of facility and the relationship the caller understood the provider to have with the facility. "She's been my mom's doctor for ten years and she has admitting privileges at the hospital" is a different posture from "the surgeon was assigned by the hospital, we'd never met him before."

Five vendor-evaluation questions for med-mal intake

If you are looking at AI voice agents for a med-mal practice, ask each vendor:

Show me the intake-schema configuration. Can the schema branch on "device involved" to capture manufacturer, model, recall awareness, and explant-preservation status, without engineering work.
Does the agent capture each provider in the treating-providers chain as a separate structured field, or does it concatenate to a single "doctor" string.
How does the agent handle the records-status question. Show me how the four states (in-hand, requested-pending, denied, not-yet-requested) are captured and surfaced to the case-evaluation queue in Clio, MyCase, or Filevine.
What happens when the caller cannot articulate the deviation from the standard of care. Does the agent flag "causation incomplete" and route to a human intake call, or does it push the matter into the lawyer queue regardless.
Show me the SOL-input schema. Are date-of-procedure, date-of-discovery, date-of-most-recent-treatment, minor-status, and decedent-status captured as separate fields the case-evaluation lawyer can sort on.

If a vendor cannot answer those five with a screen-share rather than a marketing line, they have built a general PI intake and put a med-mal sticker on it.

The contrarian close

The temptation is to push the AI to do more of the legal analysis: tell the caller whether they have a case, suggest the SOL is or isn't blown, name the specific state's certificate-of-merit requirement. This is the wrong direction.

The intake AI's job is to capture the inputs cleanly, in four minutes, in a structure the case-evaluation lawyer can read in fifteen. The case-evaluation lawyer's job is the analysis. A vendor pitching you on "AI that pre-screens med-mal cases" is pitching you on a liability surface, not a feature. The firms that have made AI intake work in med-mal are the ones that drew the line at capture, and held it.

Related reading on adjacent intake decisions:

If you are evaluating med-mal intake configurations for your firm, book a call with the team, start free trial, or see the pricing page.