When AI Writes the Discovery Responses: A Practical Walkthrough for NJ Small Firms

Discovery responses are one of the most time-consuming, margin-draining tasks in a small litigation practice. Interrogatory answers, document request responses, objection boilerplate, it's the kind of work that eats a Tuesday afternoon and generates very little client goodwill. So it's no surprise that solo and small-firm attorneys in New Jersey are increasingly feeding discovery requests into AI tools and asking for a first draft.

That's a defensible choice. What happens after the draft comes out of the tool is where things get complicated.

This post walks through what a responsible AI-assisted discovery workflow actually looks like, where New Jersey's ethics rules create specific pressure points, and what you should have in place before you sign your name to an AI-generated response.

What AI Gets Right (and Where It Immediately Fails You)

General-purpose AI tools like ChatGPT or Claude are genuinely useful for structuring discovery responses. They can produce organized objection frameworks, match interrogatory numbering, and draft consistent language across dozens of requests without fatigue. That's real value.

The failure point is factual specificity. If you paste in interrogatories and ask for substantive answers without giving the tool your client's actual file, you'll get plausible-sounding answers that are either fabricated or too vague to serve. Some attorneys have made the mistake of treating the AI's output as a starting point for editing when it should have been treated as a structural scaffold to fill in manually. That distinction matters more in discovery than almost anywhere else, because discovery responses go out under a verification, signed by your client and, implicitly, certified by you.

Legal-specific tools like Clio Draft, Harvey, or CoCounsel can integrate with your matter files to ground responses in actual client documents. That integration narrows the hallucination risk, but it doesn't eliminate the attorney's duty to read what goes out the door.

The RPC 3.3 Problem Nobody Talks About in This Context

Most AI-and-ethics conversations in NJ small firm circles focus on confidentiality under RPC 1.6 or competence under RPC 1.1. RPC 3.3 (Candor Toward the Tribunal) gets less attention in the discovery context because discovery responses aren't filed with the court in the same way a brief is. But this is a mistake.

In New Jersey state court practice, discovery responses are governed by the Rules of Court, and discovery abuses can reach the court in a hurry: through motions to compel, sanctions applications, or certifications of discovery compliance. If an AI-generated response contains a false statement of fact because no human checked the underlying documents, you've potentially got a candor problem on your hands, not just a client communication problem.

The verification your client signs isn't a formality. It's a representation that the answers are true to the best of their knowledge. If you didn't review the client file carefully before finalizing AI-generated substantive answers, you may have put a false certification in front of a judge without knowing it.

A Workflow That Actually Holds Up

Here's what a defensible AI-assisted discovery process looks like in practice for a NJ small firm:

Phase 1: Input curation. Before you touch an AI tool, assemble the relevant documents and client notes for the matter. The AI is only as accurate as what you give it. For interrogatories requiring factual responses, paste in the actual client data alongside the question. Don't ask the AI to guess.

Phase 2: Structure first, substance second. Use AI to generate the objection framework, the format, and the general legal boilerplate. Let it handle "Plaintiff objects to this interrogatory as overbroad, unduly burdensome, and seeking information protected by the attorney-client privilege, and without waiving those objections, answers as follows..." That part the tool does well.

Phase 3: Manual substantive review. Every substantive answer gets read against the actual client file. This isn't optional and it isn't quick. Budget the time.

Phase 4: Client review before verification. The client should read the substantive answers before signing the verification. AI can make confident-sounding errors. The client is the check. Document that you sent it to them and that they confirmed accuracy.

Phase 5: Final attorney sign-off with a checklist. Before the responses go out, confirm: no hallucinated case citations in any objection legal arguments, no factual claims you haven't traced to a source document, and no contradictions with prior disclosures.

One Practical Tip on Objections

AI tools tend to pile on objections. Vague, ambiguous, overbroad, unduly burdensome, seeks information not reasonably calculated to lead to the discovery of admissible evidence, all in one response, for every request. That pattern is recognizable, it irritates judges, and in NJ practice, boilerplate objections without specificity can be waived or stricken. Edit the objections your tool generates down to the ones you can actually defend if opposing counsel calls you about it. The AI draft is a menu, not a final order.

What to Do Before Your Next Discovery Deadline

If you haven't thought about this yet, start with one concrete step: pull up your last AI-assisted discovery response and trace three of the substantive answers back to a source document in the client file. If you can't do it in under ten minutes, your review process has a gap worth closing before the next set of responses goes out.

That's the check that keeps a useful tool from becoming a liability.