AI-Assisted Discovery in NJ Small Firms: When Does It Cross Into the Unauthorized Practice of Law?

Discovery is where AI tools tend to prove themselves fastest. Upload a few thousand documents, ask the platform to flag potentially privileged items, sort by custodian, or generate a first draft of interrogatory responses, and you've saved a day or two of paralegal time. For a solo practitioner in New Jersey handling commercial litigation or family law matters, that kind of efficiency is genuinely meaningful.

But there's a problem that doesn't get discussed enough in the AI-for-lawyers conversation: some of what these tools produce, depending on how they're deployed and who's supervising the output, can brush against the unauthorized practice of law under RPC 5.5. And when it does, the attorney isn't off the hook just because a machine did the work.

What RPC 5.5 Actually Prohibits

New Jersey's RPC 5.5 prohibits a lawyer from assisting "a person who is not a member of the bar in the performance of activity that constitutes the unauthorized practice of law." Most attorneys read this as a rule about staffing, not about software. That framing is outdated.

The New Jersey Supreme Court's Advisory Committee on Professional Ethics (ACPE) has consistently held that the practice of law includes applying legal principles to specific facts to advise clients or prepare legal documents. That definition doesn't change when the entity doing the application is an AI model rather than a paralegal. What matters is whether a lawyer is meaningfully supervising the output before it reaches a client, opposing counsel, or the court.

When an AI platform drafts interrogatory responses directly from intake documents and a non-lawyer staff member reviews them before sending, with only a cursory attorney sign-off, you have a supervision gap that RPC 5.5 and RPC 5.3 are both designed to address. The fact that the "person" in RPC 5.5 is software doesn't eliminate the risk. It just shifts how the analysis works.

The Discovery Workflow Scenarios That Create Exposure

Not every AI discovery task carries the same risk. Here's a realistic breakdown:

Lower risk: Using AI to OCR and organize documents, run keyword searches, generate a chronological timeline from metadata, or identify duplicate files. These are clerical functions. A competent attorney reviewing the results is doing the legal work.

Moderate risk: AI-generated privilege logs. The tool is making a call, implicitly or explicitly, about whether attorney-client privilege or work product protection applies to a specific document. That's a legal judgment. If the attorney is reviewing every entry against the underlying document before the log goes out, the workflow is defensible. If the attorney is spot-checking ten percent of a five-hundred-item log, it probably isn't.

Higher risk: AI-drafted responses to interrogatories or requests for admission, where the model is characterizing facts, selecting what to admit versus deny, and framing objections. This is squarely legal work. A solo attorney who lets a non-lawyer staff member finalize those responses, even lightly edited by the AI, has a real problem under both RPC 5.3 and RPC 5.5.

The Vendor Agreement Angle Most Attorneys Miss

Here's a practical issue that almost never comes up in ethics discussions: several AI discovery platforms include language in their terms of service stating that the platform's outputs are "not legal advice" and that the attorney bears sole responsibility for all work product. That's standard, and it's fine. But some platforms go further and explicitly market certain features, including "automated privilege review" or "AI-generated responses," as attorney work product generators.

When you sign a vendor agreement that positions AI output as a finished work product rather than a drafting aid, you're building a workflow assumption into your contract that conflicts with how New Jersey ethics rules require you to treat that output. The agreement doesn't create an ethics violation by itself, but it shapes how your staff will use the tool, which creates the actual risk.

Before you activate any AI discovery feature for active NJ litigation, read the vendor's feature documentation. If the platform describes a feature as producing a finished product rather than a starting draft, treat that as a flag to build a mandatory attorney review step into your intake procedure for that output type.

A Practical Checkpoint Before Your Next Discovery Deadline

The question to ask about any AI-generated discovery output isn't "Did a lawyer sign off on this?" It's "Did a lawyer apply independent legal judgment to this specific document before it left the firm?"

One way to operationalize that: create a short internal checklist that travels with any AI-generated discovery item before it goes out. It doesn't need to be elaborate. It needs to confirm that the supervising attorney read the underlying source material, not just the AI summary, that the legal standard being applied (privilege, relevance, admissibility) was the attorney's determination, and that the final document reflects the attorney's own characterization of the facts. A date, initials, and three checkboxes on a shared matter document will do it.

That kind of paper trail isn't just good practice under NJ ethics rules. If a dispute ever arises about the adequacy of your discovery responses or privilege assertions, it's the difference between a defensible record and a blank one.

AI-assisted discovery is worth using. The efficiency gains are real, and the tools are getting better. But in New Jersey, "AI-assisted" has to mean exactly that: the attorney is doing the legal reasoning, and the AI is handling the mechanical work underneath it.