Agents and Access to Justice

A person is sued for a debt they believe they already paid. They have a summons, a few bank statements and emails on their phone. They need someone to work out what happened, identify the defenses worth raising and respond before the deadline. The amount in dispute may be too small to support the cost of conventional representation. It is still enough to matter to the person being sued.

This is where I want to start thinking about AI agents in litigation.

An agent could help collect the records, build a chronology, connect each allegation to the available evidence and prepare work for a lawyer to review. It could keep that work current as new documents arrive. The lawyer would spend less time reconstructing the file and more time deciding what to do with it.

That is a plausible route toward democratizing access to justice. It depends on whether the work becomes cheaper without making the client bear the cost of mistakes the system cannot recognize.

Make representation possible

The Legal Services Corporation's 2022 Justice Gap Study found that low-income Americans received no or insufficient legal help for 92 percent of the civil legal problems that substantially affected their lives. That figure measures unmet need, not the share of lawsuits an agent could handle. It is nevertheless a reason to ask whether we can change the economics of providing help.^[1]

Litigation requires sustained attention to ordinary work. Someone has to obtain the missing statement, compare the complaint with the client's account, organize exhibits, check citations and notice when a new order changes what comes next. Each step consumes time. Together, they can make a modest case expensive to handle.

An agent's useful contribution would be to carry these tasks through under defined instructions. A document arrives; the chronology is updated; a discrepancy is flagged; the lawyer receives the source and the question that needs an answer. The benefit would come from reducing the work between decisions, while keeping those decisions with someone responsible for the representation.

For a legal aid office, that could create capacity for additional clients. For a small practice, it could make a lower fixed fee workable. Neither result follows automatically from faster software. Savings have to reach the people seeking help, and the time needed for supervision has to be counted honestly.

Give the agent a bounded job

Consider the debt case as a fictional example. A lawyer has interviewed the client and decided to investigate whether the claimed balance includes a payment the creditor failed to credit.

The agent is instructed to organize the statements, identify possible matching payments and prepare a comparison with the complaint. Every entry must link to its source. Missing records must remain marked as missing. The agent may prepare questions for the lawyer, but it may not turn an uncertain match into an established fact.

Now a statement arrives showing that the payment went to a different account. That might be an administrative detail. It might undermine the defense. The agent should identify the discrepancy and bring it back to the lawyer rather than quietly adjust the story to preserve its original assignment.

The useful unit of delegation is a decision with conditions attached. The lawyer has authorized an investigation on a particular factual premise. When the premise changes, the system needs to stop treating the instruction as complete.

This applies throughout litigation. A draft discovery response may expose a privilege question. A new filing may raise a defense nobody has assessed. A client may explain that the outcome they want has changed. The system needs a reliable way to return those questions to a person who can decide them.

The lawyer remains responsible

Existing professional obligations matter before anyone designs a more ambitious workflow. ABA Formal Opinion 512 says lawyers cannot leave generative AI tools alone to provide legal advice, negotiate clients' claims or perform functions requiring a lawyer's personal judgment or participation. It also calls for review of AI-generated court submissions, including their analysis and citations. The opinion interprets the Model Rules; it is not binding law in every jurisdiction.^[2]

Court rules impose their own responsibilities. In federal civil litigation, Rule 11 requires a reasonable inquiry supporting representations in pleadings, motions and other covered papers. Using an agent does not remove that obligation. State proceedings have their own applicable rules.^[3]

The client's authority also remains distinct from the lawyer's supervision. In New York, Rule 1.2(a) reserves the decision whether to settle to the client, and Rule 1.4 governs communication. An agent cannot substitute a predicted preference for the client's decision.^[4]

For the workflow I have in mind, filings, substantive advice and decisions affecting the client's rights would remain behind meaningful human review. The lawyer would need the evidence and enough time to assess it. A button labeled “approve” is not a substitute for that work.

Test the difficult cases

The strongest objection is that recognizing an exception can itself require legal judgment. An agent might correctly summarize ninety documents and miss the sentence in the ninety-first that changes the case.

That objection should shape the testing. A system built to assist debt defense should be tested against mismatched account numbers, incomplete payment histories, contradictory client accounts and documents that look relevant but belong to another person. Success on clean examples would tell us little about whether it can help safely with a real file.

The workflow should also limit what a mistake can do. Preparing a chronology and filing an admission should not be available through the same unrestricted permission. Uncertainty should hold consequential work for review. Escalation needs a named lawyer, a deadline and a backup, so that stopping the agent does not mean abandoning the client.

The record should show the instruction, supporting documents, changes and actual human decisions. It should let the lawyer reconstruct what happened without relying on the model's later explanation. That record also contains client information. Confidentiality obligations, including New York Rule 1.6, remain relevant to how it is stored and who can access it.^[4]

Measure access, not output

A service that produces more documents has not necessarily improved access to justice. I would want to know whether more people received timely advice, whether lawyers could take matters they previously had to decline, what clients paid and how often errors required correction.

There are limits software will not resolve. People may need interpreters, accessible services, help gathering evidence or someone to appear with them in court. They may struggle to use a digital intake form. An affordable service needs a way to reach a person when the software is not enough.

The promise is still substantial. If agents can reliably carry forward the work around a lawyer's decisions, competent representation could become practical in more cases. The lawyer sets the limits so that a client with fewer resources gets the benefit of the technology without being left alone with its failures. That is the version of agentic litigation worth building.