If your AI review can’t be audited, it isn’t ready for court

HaystackID says generative AI has moved from experimentation to standard use in eDiscovery. Its webcast says hundreds of customers are using it for review and privilege, and that more than 100 million documents have been analyzed.

That should change how litigators think about the work. The question is no longer whether AI can shave time off review. The question is whether the workflow can be explained, certified, and audited when the other side or the court asks how the documents got from the collection set to the production set.

If a workflow cannot be explained, certified, and audited, it is not ready for production.

Courts are already asking for proof

The immediate peg is Schulte v. LinkedIn Corp., where the panel pointed to a ruling allowing search-term culling before AI-assisted review. That matters because it shows courts are starting to fit these workflows into doctrine instead of treating them as novelty.

The same pressure is showing up in the sanction record. Our own database tracks 28 AI-related sanctions, with New York leading at five, California at four, and Oregon at three. Mata v. Avianca ended with a $5,000 fine after Schwartz filed a brief with six ChatGPT-fabricated cases. In Whiting v. City of Athens, Tennessee, the Sixth Circuit sanctioned two Tennessee lawyers for appellate briefs with more than two dozen fake or misrepresented citations and facts. The message is plain. Courts do not care that the tool was fast if the result cannot be trusted.

The real risk is not speed. It is unreadable process.

A lot of litigators still talk about generative AI in eDiscovery as an efficiency experiment. That is the wrong frame now. Efficiency is cheap until the workflow breaks under scrutiny.

A judge does not need to know your vendor’s pitch. A judge needs to know whether the matter team can say what the system did, who checked it, what was excluded, and where the record of that decision lives. If your answer is a shrug, you do not have a production process. You have a hope.

What a lawyer should do Monday morning

Start with the review protocol. If AI is being used for search-term culling, review, or privilege, write down the sequence in plain English. Keep the human checkpoints visible. Capture the logic for exclusions and escalation. Make sure someone can reconstruct the path from input set to production set without relying on memory.

Then test the paper trail. Ask whether the team can certify the workflow, not just the output. Ask whether the audit log exists and whether anyone would actually be able to use it after a challenge. Ask whether the client would be comfortable seeing the process described line by line in a declaration. If the answer is no, the workflow is not ready for production, no matter how well the demo went.

The counterargument is real. It still loses

The strongest objection is that AI-assisted review is already in use at scale and the market is moving fast, so insisting on more documentation will slow matters down. HaystackID’s numbers support the first half of that point. Hundreds of customers and more than 100 million documents are not pilot territory.

But scale cuts the other way. The more a workflow is used, the more likely it is to be challenged. The more courts see AI in filings and discovery, the less patience they will have for vague assurances. That is why the buying decision now belongs with defensibility, not novelty. If a workflow cannot be explained, certified, and audited, it is not ready for production.