AI Agents on a Personal Injury Case: Where the Attorney's Review Belongs
AI agents are changing where personal injury lawyers need to review the work, because an agent can move through several connected steps before anyone at the firm sees the output. Here's where attorney or supervised-staff review belongs on a PI file, and what professional responsibility already requires.
What changes when an AI agent works a personal injury file?
Single-task AI tools put a lawyer's eyes on every output, because nothing happened until someone asked for the next thing. An agent chains several steps on its own, so this article maps where review has to land on a PI case and what ABA Formal Opinion 512 already says about supervising that work.
When the Software Stops Waiting to Be Asked
Until recently, AI for personal injury litigation worked one request at a time. A paralegal uploaded records and asked for a summary, an attorney asked a research question, someone pasted facts into a demand template. Each output landed in front of a person before anything else happened, so review was built into the workflow whether the firm planned for it or not.
Agents change that sequence. An agent takes an instruction, breaks it into steps, carries the result of one step into the next and returns finished work product. The definitions and the general case for agentic AI are covered in our piece on agentic AI in law. This article asks something narrower: once the work arrives in one piece, where does the attorney's review still have to happen, and what do the ethics rules expect?
The question matters more in plaintiff practice than almost anywhere else. A PI file is long, records-heavy and sequential, so a mistake in the chronology flows into the damages summary and then into the demand.
Why a PI File Breaks One-Task Tools
A one-task tool sees the document in front of it. A personal injury matter rarely lives in one document. The ER visit, the orthopedic follow-ups, the physical therapy notes, the imaging and the billing ledgers come from different providers and arrive over months.
What a one-task tool loses between requests is the thread. Ask it to summarize the PT records in March and the surgical records in May, and it has no reason to connect the two. The person connecting them is the paralegal or the attorney, which is the work the firm hoped to shorten in the first place.
Anytime AI 3.0 addresses this with case memory: context from a matter carries across requests in that matter and stays inside it. That's useful on a personal injury file, and it's also why supervision gets harder rather than easier. When the system remembers more, more of its reasoning happens out of view.
What One Instruction Sets in Motion
Consider a hypothetical instruction a paralegal might give an agent: review this client's medical records and prepare a treatment chronology in our firm's format. ⚠ Verify with Hugo. A reasonable agent would work through something like this:
Read every uploaded record and identify each medical encounter.
Put the encounters in date order across providers.
Pull the diagnoses, procedures and referrals from each encounter.
Note changes in treatment, such as a move from conservative care to surgery.
Flag gaps between visits and entries that conflict with each other.
Assemble the chronology in the firm's template and return it.
That's six steps from one instruction, and nobody at the firm sees steps one through five unless they go looking. If step one misses a record because a provider's scan was illegible, every later step inherits the miss. The finished chronology looks just as complete either way.
Medical record review is where an AI medical chronology earns its place on a personal injury case, and we've covered how AI builds medical chronologies in detail. The supervision question is separate: which of those six steps can a lawyer actually check, and how?
What Does ABA Opinion 512 Require of Supervising Lawyers?
The American Bar Association's Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, Generative Artificial Intelligence Tools, on July 29, 2024. It creates no new rules. It applies existing Model Rules to AI use: competence under Rule 1.1, confidentiality under Rule 1.6, client communication under Rule 1.4, candor and meritorious claims under Rules 3.1, 3.3 and 8.4(c), supervision under Rules 5.1 and 5.3, and reasonable fees under Rule 1.5.
For AI ethics, attorneys usually focus on three of those duties. Competence means a reasonable understanding of what a tool can and can't do, not expertise in how it's built. Supervision means the lawyer can't hand professional judgment to the tool and has to verify its output, with the depth of review matched to the task. Confidentiality means weighing whether client information could leave the matter, and the opinion specifically flags tools that learn from what's put into them.
The opinion was written about generative AI tools and doesn't use the word agent. Nothing in it suggests that a tool chaining several steps gets a lighter standard than one answering a single prompt. Read against an agent, Rule 5.3 asks roughly what it would ask of a new paralegal: know what they can do, give clear instructions, check the work in proportion to what's at stake, and keep a firm policy on who signs off.
The cautionary case most lawyers know is itself a personal injury case. In Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y.), the plaintiff alleged a metal serving cart struck his knee on an international flight. His lawyers filed a brief citing decisions that ChatGPT had invented. On June 22, 2023, Judge P. Kevin Castel sanctioned them $5,000 under Rule 11, and in a separate opinion the same day dismissed the claim as time-barred under the Montreal Convention's two-year limit.
Those citations came from one prompt, and they still reached the court. An agent's output is several steps deep, so the review has to be designed in rather than assumed.
Where Should an Attorney Review an AI Agent's Work?
The table below maps common stages of a PI file to what an agent hands back and what a lawyer or supervised paralegal should check before relying on it.
Stage of the file | What the agent hands back | What the attorney checks |
|---|---|---|
Records intake | A list of encounters across providers | Every provider you requested records from appears, and no received record is missing |
Medical chronology | A dated timeline with diagnoses, procedures and flagged gaps | Spot-check entries against the source pages, and read every flagged gap and conflict yourself |
Damages summary | Treatment and billing totals | Recompute from the ledgers, and confirm which amounts your jurisdiction lets you claim |
Demand letter | A draft in the firm's template | Every factual statement traces to a record; the number and the tone are your call |
Legal research | A memo with cited authority | Pull and read every cited case in a primary source before it goes near a filing |
Questions to Put to Any Vendor
Whether a product calls itself an AI legal assistant for personal injury work or an agent platform, the supervision questions are the same. These are worth asking before any agent touches a client file:
Does the vendor train models on your client files, and does the contract say so?
Does context from one matter stay inside that matter? Opinion 512 names cross-matter disclosure as a confidentiality risk.
Does the system keep the context of a case across workflows, or does each task start from scratch?
Can you trace each statement in the output back to a page in the source records?
Does the output come in your firm's templates or the vendor's?
Can your firm build agents around its own workflow, and if not, who builds them for you?
What preferences carry across your matters, and can you see and change them?
For AI for plaintiff law firms, the first two answers carry the most weight. Anytime AI doesn't train on client files, encrypts data with AES-256 at rest and TLS 1.2 or higher in transit, and can't see client data, and case memory stays within its matter. The security page covers the rest. Anytime AI 3.0 also runs built-in agents alongside custom agents, which Anytime AI can build around a firm's own process.
What AI Agents for Personal Injury Lawyers Can't Do
An agent can flag a gap in treatment. It can't tell you whether the gap is the client's choice, a lost record or a provider who never sent the file, and it can't tell you a record is missing if nobody requested it.
It can't judge credibility. The client's account, the treating physician's narrative and the defense medical examiner's report are weighed by a lawyer who has met the people involved.
It can assemble the numbers that go into a demand. The figure you put in the letter, and when you're willing to move off it, is a judgment about venue, the carrier and the client.
Its output is only as good as its input. Illegible scans, handwritten notes and partial productions limit what any system can read, and the gaps don't always announce themselves.
And no vendor can honestly promise an agent will never state something the records don't support. The defensible position is a system whose output you can trace to the source, used by a lawyer who does the tracing.
Final Thoughts
An agent changes how much work arrives at once. It doesn't change who answers for it. For each workflow your firm hands to an agent, write down the step where a person checks the output against the records, and who that person is. That's the policy Opinion 512 points toward, and it's the step that catches a missing record before the defense does.
FAQs
Does ABA Formal Opinion 512 apply to AI agents?
The opinion addresses generative AI tools generally and doesn't carve out agents. Its duties of competence, confidentiality, supervision and candor apply to any tool whose output a lawyer relies on.
Do I need client consent to use AI on a personal injury case?
Opinion 512 says informed consent is needed before putting client information into a tool that learns from inputs, and that boilerplate in an engagement letter isn't enough. Where a tool doesn't train on client data, the answer depends on the circumstances and your state bar's guidance.
Can an AI agent review medical records for a personal injury case?
Agents can read records across providers, order encounters by date and flag gaps or conflicts in treatment. A lawyer or supervised paralegal still has to confirm the record set is complete and spot-check entries against the source pages.
Who is responsible if an AI agent gets a demand letter wrong?
The lawyer who sends it. Rules 5.1 and 5.3 place supervisory responsibility on lawyers, and Opinion 512 applies those rules to AI output the same way it applies them to work from staff.
Does case memory share information between my clients' matters?
In Anytime AI, case memory stays within the matter it came from. Ask the same of any vendor, because Opinion 512 identifies information surfacing across matters as a confidentiality risk.
Get Started
Ready to go deeper — and safer?
See how Anytime AI gives plaintiff firms the strategic edge
and the security their clients deserve.