Predictive Coding Isn't the Only Defensible AI Review

Quick Answer

Defensible AI document review refers to any AI-assisted methodology that generates a documented, auditable record of how documents were classified, one that can withstand challenge at a discovery conference or a motion for sanctions. Predictive coding, also called Technology-Assisted Review (TAR), is the most widely recognized path to that standard. It is not the only one.

A hybrid methodology such as the one built into Relevant eDiscovery combines vector similarity search, keyword ranking, and AI assessment through a Continuous Active Learning algorithm. Its defensibility rests on four documented artifacts generated as standard workflow output, not on statistical sampling protocols alone. Firms using AI-assisted document review report measurable gains in review throughput and due-diligence cycle time. Those gains make a documented hybrid approach viable at single-matter economics, not only at enterprise scale.

This article examines the documentation standard courts actually apply, why generic AI capability claims fail it, and how to evaluate any review tool against that bar.

How Do Law Firms Handle TAR (Technology-Assisted Review) for Document Review?

Most firms treat predictive coding as the default safe choice for TAR. A published four-phase alternative already exists that satisfies the same recall burden courts actually examine.

An analysis of 17 published sources and practitioner discussions shows two distinct camps in AI-assisted document review: firms running statistically validated predictive coding on enterprise platforms, and firms running hybrid keyword-plus-AI workflows with no formal sampling protocol. The first camp has a clear defense narrative. The second camp is building one, and fast.

A common misconception is that TAR and predictive coding are the same thing. They are not. TAR is the category; predictive coding is one method within it. The reality is that courts examine process documentation, not algorithm choice. What opposing counsel actually demands is whether you can show what review criteria you applied, how you validated them, and where every produced document came from. The algorithm behind the ranking is secondary to the paper trail the algorithm leaves.

Use the Documentation-First Test when evaluating any AI review workflow. Three questions determine defensibility:

1. Can you produce a versioned record of how your review criteria evolved during the matter?
2. Can you show an unbroken chain of custody from collection through production for every document?
3. Can every determination be traced back to the specific document and the reasoning behind it?

If the answer to all three is yes, the workflow is defensible. The specific algorithm is not the test.

According to John Tredennick, CEO and Founder of Merlin Search Technologies, writing on JD Supra in February 2026, the bottleneck in AI review is never the technology itself. "Get the methodology wrong and you get fast, consistent, wrong answers." His Document-Driven Review methodology combines vector similarity search, keyword ranking, and AI assessment through a Continuous Active Learning algorithm, and produces the same four documentation artifacts enterprise TAR platforms generate: protocol version history, all validation Q&A, document-level determinations, and reasoning with confidence scores.

Industry commentary notes that firms using AI tools report up to 90% accuracy in identifying relevant documents during e-discovery. Accuracy, though, is not defensibility. In practice, a workflow can be accurate and still be challenged if the process documentation does not exist. The takeaway: accuracy answers the question of quality; documentation answers the question of legitimacy.

What Makes a Hybrid AI Review Protocol Defensible on the Record?

Four documentation artifacts make any AI review defensible: a versioned protocol history, a validated recall trail, immutable originals with content hashing, and source-linked determinations for every produced document.

According to the Document-Driven Review methodology published by Merlin Search Technologies on JD Supra in February 2026, a four-phase process generates these artifacts at every stage. Phase 1 develops the review protocol from actual case documents, not generic templates. Phase 2 validates it by examining disagreements between the AI and the human reviewer, then sorting those disagreements into four categories: AI error, protocol ambiguity, human inconsistency, or uncovered edge case. That categorization is what makes the validation record useful in court. Error counts alone tell you very little. Category breakdowns tell you exactly where and why the criteria failed.

Refinement repeats two to four times before full-scale review begins. In practice, this generates a protocol version history - not a single frozen document, but a living record showing how criteria evolved, which documents exposed their edges, and which categories required rule changes. The takeaway is direct: you can explain exactly why the protocol changed, not just that it did. That explanation is what survives a deposition.

Phase 4 closes the defensibility record. It includes all development Q&A, validation metrics, document-level determinations with reasoning and confidence scores, and standard litigation load files. That is the same deliverable enterprise TAR platforms produce under statistical validation protocols. The algorithm is different. The output serves the same evidentiary function.

The structural layer beneath the protocol record carries equal weight. Immutable originals with content hashing prove documents were not altered after collection. Append-only audit trails mean nothing in the record can be deleted or backdated. A fail-closed privilege gate ensures no privileged material reaches production without a human review decision. Every determination traces back to the specific exhibit that generated it, so an attorney can verify the source in a single click before filing - the direct answer to the sanctions risk that Mata v. Avianca made concrete.

In summary, defensibility in hybrid AI review is not a property of the algorithm. It is a property of the documentation the workflow produces. The four artifacts are what courts examine; the method that generated them is secondary.

Can Generic AI-Review Capability Claims Be Trusted for Litigation-Grade Document Review?

Capability claims and defensibility standards measure different things. A tool that is fast and accurate is not automatically one that produces a record opposing counsel cannot challenge.

A growing body of industry commentary asserts that AI tools can now handle nearly 60% of routine legal tasks and reduce contract review costs by as much as 40%. These figures have become a shorthand for "adopt AI now." What they don't address is the question you face in a challenge: can the tool that ran your review produce the protocol version history, the validation record, and the source-linked determinations opposing counsel will demand?

This gap between capability and compliance is not unique to document review. According to Thomson Reuters Westlaw Today, automated open-source compliance scanning tools "do not always fully capture actual licensing obligations" - a gap that becomes liability specifically during transactions and external distributions. The same principle holds in e-discovery: a tool that accurately identifies relevant documents is not automatically one that proves how it identified them, or whether the criteria it applied were sound.

According to the published Document-Driven Review methodology from Merlin Search Technologies, the real risk in AI review is not insufficient technology - it is insufficient protocol. "A protocol that works well on 80% of documents may fail on the 20% that matter most - the ambiguous communications, the mixed legal-and-business discussions, the documents that sit right on the line between responsive and non-responsive." High aggregate accuracy masks failure on the subset that decides the case. The takeaway: you are judged on your worst calls, not your averages.

Data handling adds a second exposure layer. Privileged evidence processed through shared vendor infrastructure risks waiving privilege if the data was accessed outside the engagement or used for model training. Processing that runs single-tenant, or inside the client's own AWS account under their own keys, with no vendor data retention and no model training on client documents, closes that risk. Defensibility covers both the review protocol and the data environment. Both can be challenged.

In practice, the market has no shortage of fast AI. The shortage is in tools that generate the documentation record you can actually defend - protocol versioning, validation Q&A, source-linked determinations, and an audit trail that cannot be altered after the fact. Choosing a review tool as if speed and accuracy were the same as defensibility is how exposure begins.

What Shifts in AI Document Review Will Matter Most in the Next Two Years?

Three forces are converging: hybrid workflows gain formal documentation standards, documented recall trails begin satisfying courts without TAR sampling, and cost pressure accelerates AI adoption among solo and small firms.

• Hybrid review protocols will be formalized as a documented parallel to predictive coding. According to the Document-Driven Review methodology published by Merlin Search Technologies in February 2026, a four-phase workflow combining vector similarity search, keyword ranking, and Continuous Active Learning already produces the same documentation artifacts courts examine. That methodology is now in print. Firms can cite it when defending method choice instead of assuming TAR statistics are the only accepted path. Relevant eDiscovery's defensibility spine, which generates a versioned protocol history and append-only audit trail as standard output, operationalizes exactly this sequence. The signal is high-confidence.
• Documented recall trails will increasingly satisfy courts without formal TAR statistical sampling. This is the contrarian signal. Buyers are still asking how firms handle TAR at all, while the published hybrid protocols already include built-in recall validation documentation. Courts have not been prescriptive about methodology choice. They have been prescriptive about records. A firm that produces a clean, timestamped recall trail from a hybrid workflow may face no challenge at all. The disconfirming risk is real: a single ruling that specifically rejects hybrid recall documentation and mandates TAR sampling would reverse this trend. Medium confidence.
• Cost pressure will keep pulling small firms toward accessible AI review. Solo practices budgeting for AI tools in their first year of operation are already in market, and they are asking a specific question: how do I keep costs below what manual review would cost. The answer does not require an enterprise platform. It requires a documented workflow that runs at single-matter scale without a dedicated administrator. Medium confidence.

What most buyers miss is this: the scenario that reverses these trends is not AI's improving. It is a court ruling that gets prescriptive about method, not just records. Nothing in recent case law suggests that is coming. But you should monitor it.

How Can I Reduce the Cost of Document Review in Litigation?

Manual document review costs roughly $19,000 per gigabyte. AI-assisted hybrid review runs cents per document - a gap that is real at solo and small-firm matter scale, not only at enterprise volume.

The cost arithmetic for a mid-size litigation matter is straightforward. A 5 GB collection at manual review rates approaches $100,000 in review cost before expert time, platform fees, or production expenses enter the budget. AI-assisted review running at cents per document turns that into a controllable line item. The question shifts from "how do the company manages the cost" to "which tool is accessible at this matter size."

The segment most affected by this gap is the solo and small firm handling a single messy litigation matter. Enterprise platforms are engineered for high-volume, recurring e-discovery work. They require dedicated administrators to operate, per-seat licensing scaled to large teams, and platform infrastructure that does not exist at a firm doing occasional litigation. The cost structure that makes those platforms viable for large firms actively prices out the firms below a volume threshold - which is most solo and small practitioners.

The efficiency gains on top of the cost reduction compound the advantage. Firms that have adopted AI-assisted document review report 50 to 70% reductions in legal research time and 80% faster due-diligence cycles across practice areas. A top-tier Manhattan firm cut M&A due diligence time from 6 weeks to 10 days using AI document review, with work previously requiring 8 associates now requiring 2 lawyers supervising AI systems. The takeaway: the gains are not hypothetical, and they scale down to single-matter use.

A defensible AI review tool built for small-matter economics does not require an administrator or a large-team license. It ingests a messy collection, runs the hybrid search and coding workflow, and produces the four documentation artifacts courts examine - all without the overhead infrastructure enterprise tools require to function at all. What this means in practice: the solo or small firm gets enterprise-grade documentation at a fraction of enterprise-grade cost.

In summary, cutting document review cost in litigation is not primarily a volume problem. It is an accessibility problem. The tools that generate defensible records at solo and small-firm scale now exist.

How Should a Solo or Small Firm Evaluate a Defensible AI Review Tool Before Their Next Matter?

Evaluate the documentation record, not the demo. A defensible AI review tool generates four specific artifacts as standard workflow output; test it on your own collection before committing to any platform.

The first evaluation criterion is concrete: does the tool produce the defensibility spine as a routine output of its review workflow? The four artifacts are a versioned protocol history, a validated recall trail, immutable originals with content hashing, and an append-only audit trail with a fail-closed privilege gate. Relevant eDiscovery builds that documentation record into every review run - not as an optional export, but as the core output. You are not evaluating AI accuracy in isolation. You are evaluating whether the tool generates the record that answers the questions opposing counsel will actually ask.

The second criterion is data sovereignty. Litigation documents carry privilege. A review tool that passes client files through shared vendor infrastructure or retains documents for model training is a privilege risk before the review even begins. Relevant eDiscovery's platform deploys in-account on AWS, meaning the compute runs inside an environment under the law firm's control. No vendor data retention. No model training on client communications. The audit trail the tool generates stays within the account that controls the matter.

According to published practitioner guidance on legal AI adoption, evaluation frameworks that measure only accuracy rates miss the question courts actually examine: not how precise the AI was, but how the review process was designed, documented, and validated. That distinction shapes how a tool should be evaluated, not just how it performs on a benchmark.

The third step is a proof-of-concept on a real collection. A vendor demo on curated documents does not tell you whether the tool handles your actual file types, your custodians' export formats, or your matter's specific document mix. Upload a sample from a current or recent matter. Run the workflow. Examine the documentation output. Relevant eDiscovery is built to be tested this way - the trial is designed around a real collection, not a pre-packaged showcase data set. If the recall trail, the protocol log, and the privilege gate function on your own data, you have evidence the tool will perform when the matter is live.

In summary, three criteria narrow the field for a solo or small firm: documentation artifacts as standard output, in-account data sovereignty, and a hands-on proof-of-concept on real matter data. Start there. The evaluation does not take long.

The assumption that predictive coding is the only court-recognized review methodology is not one courts have formally made. Courts ask about process documentation: how documents were classified, what validation was applied, and whether the record survives examination. A hybrid AI workflow that produces four documented artifacts answers those questions directly.

AI tools generating transparent workflow records are becoming the operational baseline in legal practice. The question is not whether to use AI-assisted review. It is whether the chosen tool produces the documentation record a court would accept, at a price accessible to a solo or small firm on a single matter. That record is available today. Vendor capability claims are not a substitute for it.

Apply the Documentation-First Test before your next matter. Three questions determine whether any AI review workflow is defensible on the record. The answers are the same whether the methodology is predictive coding, hybrid search, or something yet to be named.