Everlaw CTO: AI Beats Human Review on Defined Ediscovery Tasks, but Misuse Brings Risk

From waterfall review to early insight

Christoff described ediscovery as having "operated like a waterfall": negotiate a protocol, hand the material to a review team, and wait weeks or months for a readout. He said that linear model is breaking down. Attorneys are showing "real curiosity much earlier in the lifecycle" and want an early read on risks to shape strategy, including whether a matter is heading to trial or settlement.

He separated this from traditional early case assessment, where the main goal is to cull material. What he describes is understanding the risks in the content so lawyers can inform strategy. According to Christoff, this lets attorneys "show up as a more engaged partner throughout the process" rather than telling clients they will know more in a few months.

The performance figures Everlaw cites

Everlaw points to first-pass document review as the clearest case, because performance can be measured with established metrics. According to the company, across four live-litigation datasets its Coding Suggestions feature achieved an average recall of 89 percent. In one dataset that allowed a direct comparison with first-pass human review, Coding Suggestions reached 82 percent recall versus 60 percent for human reviewers.

• An Am Law 100 firm used Coding Suggestions on more than 126,000 documents in a government investigation, achieved over 90 percent accuracy and coded the full set in approximately one day. Total review time fell by 50 percent to 67 percent, with a quarter of the personnel needed for a comparable managed review.
• Orrick, a global law firm, used the tool in an intellectual property case and cut document review costs by more than half.
• Dinsmore & Shohl combined strategic search, Coding Suggestions, Predictive Coding, statistical validation and human review across 26,000 contracts, and reached a recall rate of 98.2 percent.

Everlaw says the recipe for success is to define the task clearly, iterate on prompts, validate against a controlled sample, and only then scale up.

Strategy-first discovery and grounded answers

Everlaw argues the larger opportunity lies beyond faster document checking. Its Deep Dive tool lets teams ask natural-language questions across a whole document set and get citation-backed answers. Core teams can identify key actors, themes and timelines at the outset instead of waiting for a large review team to work through hundreds of thousands of files. Everlaw has reported that Deep Dive has been used in cases with millions of documents.

Christoff stressed that usefulness depends on engineering choices. Legal teams need to understand "what knowledge is embedded within an LLM’s training data versus what knowledge is actually in their documents." His team tests this by asking Deep Dive why the sky is blue with a matter loaded. The system says it cannot find any reason, because answers are designed to be grounded in the document set. Reliable tools, he said, should support every statement with citations and "freely state when they don’t have enough information to answer your question."

Where AI exposes teams

The risk, Christoff said, arises when AI is used as disconnected, ad hoc shortcuts. Exporting sensitive evidence into standalone consumer tools fragments the record, breaks data governance and leaves risk partners in the dark about client confidentiality. He criticised the way "the word AI gets used as a fairly broad brush," noting "a world of difference between consumer-grade tools, where dropping data into a public interface risks the loss of attorney-client privilege, and enterprise-grade systems built from the ground up with zero data retention and zero model training on client data."

He also described a "blinking cursor problem": a blank prompt that works well for ten documents but not for a hundred, a thousand or ten thousand, where teams need a statistical, explainable approach, work division and an audit trail. The setup must be proportionate to scale, he said, since a single contract and 1.5 million documents from a large company call for very different processes. "Same tool, different process," Christoff said.

The concern grows as tools become more agentic, orchestrating workflows, making preliminary determinations and organizing evidence. According to the interview, that makes strict internal guardrails more important.

What it means for legal teams

For law firms and legal departments, the message from Everlaw is that AI policies should distinguish between tool types rather than simply allowing or banning AI. The company's figures give teams measurable results to evaluate and defend, but its CTO frames them as valid for bounded tasks that are validated on samples first. Enterprise-grade systems with auditability and grounded answers, in his view, are the safer route as matters scale.