AI Tools for NJ Discovery: 5 Things to Verify Before You Certify That Production Is Complete

Somewhere between the promise of AI-assisted document review and the moment you sign a certification under New Jersey Court Rules, a gap opens up. It's not obvious. The tool looks confident, the output looks clean, and you're 40 hours into a case that was never supposed to be this document-heavy. So you sign.

That's where the problems start.

NJ solo and small-firm practitioners are adopting AI for discovery work faster than they're building the verification habits to match. The tools are genuinely useful: they can categorize documents by relevance, flag privilege candidates, summarize custodian files, and draft interrogatory responses from a client's intake notes. But every one of those outputs carries a failure mode that lands on the certifying attorney, not the vendor.

Here are five things worth checking before you put your name on that production.

---

1. Confirm the tool actually processed every document in the collection

AI review platforms ingest files, but they don't always tell you clearly when something didn't load. Password-protected PDFs, corrupted attachments, unusual file formats, and embedded objects (a spreadsheet inside a Word document, for instance) can drop out of the processing queue silently. The tool's "complete" dashboard may not mean what you think it means.

Before certifying, pull the ingestion log or processing report. Compare the document count the platform shows against the raw file count from your collection. If the platform doesn't surface a processing report at all, that's worth knowing before you rely on it for anything high-stakes.

2. Spot-check the privilege calls, not just the privilege log

Most AI tools flag privilege candidates; some generate a draft privilege log automatically. The problem is that the model is pattern-matching on phrases and metadata, not applying New Jersey's specific work-product doctrine or the attorney-client privilege contours the courts actually use.

Pull ten documents the tool marked as privileged and ten it marked as non-privileged. Read them. You're not looking for perfection, you're looking for the tool's systematic blind spots. If it's consistently missing a category (say, communications where in-house counsel is cc'd but isn't the primary author), you need to know that before opposing counsel does.

3. Check whether responsive documents were excluded from production because of a relevance threshold you didn't set intentionally

Most AI review tools let you configure a relevance score cutoff. Documents below the threshold get coded non-responsive and don't get produced. The default setting in many platforms is not conservative, it's whatever the vendor calibrated for general commercial litigation, which may not match your case.

Find out what threshold is actually running. If you didn't set it, find out what the default is. Then look at a sample of documents just below the cutoff line. Under NJ's discovery rules, the certifying attorney is affirming the production is complete. "The algorithm decided" is not a defense your client will find satisfying.

4. Run the interrogatory responses against the actual documents, not just the AI's summary of them

AI tools that draft interrogatory responses from a document set are particularly prone to a subtle error: they summarize what the documents seem to say rather than what they actually say. Dates get rounded. Names get approximated. Amounts from one document get attributed to a different transaction.

Pick five interrogatories. Find the underlying documents the tool used to generate each response. Read both. If the response is accurate, great. If there's a paraphrase that softens a number or conflates two events, that's something you want to catch in your office, not in a deposition.

5. Verify the production format meets NJ's technical requirements before you hand it over

This one has nothing to do with AI's substantive accuracy, it's about output format. NJ state court discovery has specific expectations around Bates numbering, load files, and metadata fields. Federal matters in the District of New Jersey often have ESI protocols negotiated at the Rule 26(f) conference.

Some AI platforms produce output in a format that needs to be reformatted before it's actually court-compliant. Others strip certain metadata fields. Check the ESI protocol or standing order in your case, then check the production package format against it before delivery. A technically deficient production can trigger a motion to compel even when the substantive review was sound.

---

None of these checks require hours. Most of them take 20-30 minutes if you build them into the workflow before you get to the certification stage. The more useful framing: treat the AI's output as a first-pass associate review, and apply the same oversight you'd apply to any associate's work before it goes out the door. NJ RPC 5.1 puts supervisory responsibility on the attorney of record, and that responsibility doesn't transfer to the software subscription.

If you're using a platform that makes any of these five checks difficult or impossible, that's a data point worth factoring into your next vendor decision.