AI Privilege Review: Costs, Workflow, and Where It Fails

Privilege review is the line item that turns a manageable discovery budget into a painful one, and it is also the part of review where a mistake has the worst consequences. AI can cut the cost of a privilege pass by half or more on the right matters, and unlike two years ago there is now real data on what it saves, what courts have said, and where it breaks. This post walks through the numbers, the workflow, and the failure modes, so you can make the call with evidence instead of a vendor demo.

Why privilege review costs so much

Privilege review costs more than ordinary responsiveness review for a structural reason: it needs a lawyer, and it produces a second deliverable. A responsiveness call is largely binary and can be delegated to contract reviewers at $0.50 to $3.00 per document. A privilege call turns on the role of every person on the thread, the dominant purpose of the communication, and whether onward circulation waived the protection. That judgment bills at attorney rates, one document at a time.

Then comes the log. Under Rule 26(b)(5)(A) of the Federal Rules of Civil Procedure, a party that withholds material on privilege grounds must expressly make the claim and describe the nature of the withheld documents in enough detail for other parties to assess the claim, without revealing the privileged content itself. Every withheld document becomes a small drafting exercise: author, recipients, date, a careful description, the basis asserted. That is billed separately from the review.

Vendor-cited benchmarks for 2026 put the components at roughly $4 to $8 per document for a privilege read by senior contract attorneys, plus $5 to $15 per privilege log entry. On a 100 GB matter, privilege review and log generation together can run $30,000 or more, stacked on top of a first-pass responsiveness review that often costs several times that.

Two features of the rules work in your favor here. Rule 26(b)(1) makes proportionality an explicit limit on discovery scope, so a documented, cheaper, reliable method is legally relevant, not just commercially attractive. And Rule 26(f)(3)(D) requires the discovery plan to address privilege issues, including the timing and method of complying with 26(b)(5)(A). That conference is where an AI-assisted protocol gets agreed, and agreeing it early is most of the defensibility battle.

What AI privilege review costs today

Pricing for generative AI review is still unsettled, which is itself useful information. The ComplexDiscovery Summer 2025 pricing survey found per-document billing is the most common model (37% of respondents), followed by hybrid structures (30%), with per-token (6%), per-gigabyte (4%), and outcome-based pricing (1%) far behind. Nearly half of respondents could not state a per-document cost at all.

Among those who could, the largest group reported $0.26 to $0.50 per document, with the next group at $0.11 to $0.25. Compare that to $4 to $8 per document for a human privilege read. Even after you add attorney QC, the arithmetic is dramatic, provided the accuracy holds up. Per-gigabyte billing, the old hosting staple, barely exists for GenAI review: 81% of survey respondents found the question inapplicable, because tokens, not gigabytes, drive model cost, and a gigabyte of spreadsheets generates far more tokens than a gigabyte of PDFs.

One more number worth knowing: 17% of providers said they revert to manual review at standard rates for complex or unusual documents. Ask any vendor what their tool does with spreadsheets, chat exports, and redacted PDFs before you sign.

A worked example with real numbers

The most concrete public example comes from Trustpoint.One, a Relativity partner, describing a real matter handled with the Relativity aiR suite. The matter involved 20,000 documents for relevance review, 5,000 of those for deeper case-strategy analysis, and 1,000 for privilege screening and log preparation.

Trustpoint estimated traditional manual review across all three workflows at $38,214. The AI-assisted approach came in at $14,875, a reduction of roughly 61%. The workflow ran inside RelativityOne under a zero-data-retention security agreement with Microsoft, which matters for client confidentiality: the documents are processed without being retained for model training.

Two details from that engagement generalize well. First, an AI-specialist attorney wrote the review criteria in plain English up front, and the system returned a written rationale with citations for every document, which is what makes the output auditable. Second, for the privilege pass specifically, aiR combines content analysis with social network mapping to identify attorneys and their interaction patterns, then drafts the log entries. The attorney's job shifts from reading everything to checking the reasoning on the documents that matter.

You can run the same arithmetic on your own matters. Take 10,000 documents that need a privilege screen: at $4 to $8 per document, the human read alone is $40,000 to $80,000, before the log. An AI pass at $0.26 to $0.50 per document costs $2,600 to $5,000, plus attorney time to QC the flagged and uncertain set, plus log editing. Even with generous QC assumptions, you are looking at 50 to 70 percent savings on that line item, which is consistent with the Trustpoint figure. The same review-economics logic applies to AI-assisted due diligence in private equity. Vendor claims of 80 to 98 percent reductions usually measure the model pass in isolation, not the total workflow cost. Treat them as ceiling numbers.

How the workflow actually runs

A defensible AI privilege review has six steps, and the order matters.

1. Agree the protocol at the 26(f) conference. Disclose the tool, the criteria, the validation plan, and the sampling method. Surprise is what generates motions, not the use of AI itself.
2. Write the privilege criteria in plain language. Identify the law firms, attorney names and domains, in-house counsel roles, and the patterns that indicate legal advice, litigation preparation, waiver, and common-interest arrangements. This is attorney work; it cannot be delegated to the vendor.
3. Map the legal actors. Use entity and social network analysis to find attorneys and paralegals across the collection, including personal addresses and outside counsel forwarded internally. This catches the privileged documents that keyword lists miss.
4. Run the AI pass. Every document gets a determination plus a written rationale with citations to the text. Keep the rationales; they are your audit trail.
5. QC with targeted sampling. Attorneys review the documents the model flagged as privileged, the ones it was uncertain about, and a random sample of the ones it cleared. Check the reasoning, not just the tag: a fluent wrong rationale is worse than no rationale.
6. Generate and edit the log. Let the system draft entries from its structured output, then have an attorney edit descriptions to the 26(b)(5)(A) standard: specific enough to allow assessment, vague enough to protect the content.

Notice what is not in this list: handing the vendor a hard drive and accepting what comes back. The human review design mirrors what works in other document-heavy AI workflows, the same pattern we described in our post on LLM claims triage for insurers, where the model triages and humans decide on the flagged slice.

What the courts have and have not said

The honest summary is short, but it changed in mid-2026. No court has yet issued a Da Silva Moore-style opinion blessing generative AI review in the abstract; the 2012 Da Silva Moore decision did for TAR what no case has done for GenAI, declaring technology-assisted review "better than the available alternatives, and thus should be used in appropriate cases." But the judiciary has now weighed in on a live GenAI workflow. In Schulte v. LinkedIn (N.D. Cal., July 1, 2026), Magistrate Judge Laurel Beeler resolved a challenge to LinkedIn's use of Relativity aiR to make final responsiveness calls on roughly 204,000 documents, and she denied every request the plaintiffs made.

The cases that do exist cut both ways. In Ohio ex rel. Yost v. Jones (S.D. Ohio, April 2025), the court rejected a party's claim that running an AI "agent" over his records constituted a diligent search and ordered him to review and produce all responsive documents. In Sheets v. Charlotte County (M.D. Florida, April 2025), the court suggested a party could use AI to prepare discovery responses provided he actually evaluated the factual and legal references the AI generated. The common thread: courts care about whether a responsible human verified the output, not about whether AI was involved.

Schulte is worth reading in full, because it tells you exactly what a court expects to see. LinkedIn had disclosed three things: no seed or training set was used, aiR was making the final responsiveness calls, and quality control was human review of samples from each responsiveness type. The court held that this disclosure more than satisfied the ESI order, and refused to compel production of elusion estimates, error rates, and reviewer counts, applying the settled rule that discovery on discovery is disfavored absent a specific deficiency in the production. It also upheld running search strings before the AI pass, rejecting the argument that LinkedIn had to feed every custodial file to the model: two custodians alone held about 800 GB, and proportionality is still the law. As a Law.com analysis put it, the court evaluated GenAI review under established discovery principles, which is itself the headline: the tool is new, the rules are not.

Meanwhile, most of the practice still moves through negotiated protocols rather than judicial opinions. In one 2025 matter, Tesla's ESI order recorded that it "may use TAR and/or GEN AI tools to further analyze documents for relevance after search terms are used to narrow the starting document universe." That is the current mechanism: disclose the method in the discovery plan, get agreement or a ruling on the protocol, and document your validation. If you wait for a landmark appellate decision before touching AI review, you will wait years while paying manual rates.

Where it fails: the independent evidence

Almost everything published about GenAI review accuracy comes from vendors or their partners. The important exception is an independent reanalysis of the Redgrave case study, one of the most-cited validations of Relativity's aiR for Review, by TAR researchers Gordon Cormack and Maura Grossman. Their findings deserve slow reading.

They found that aiR was "comparable to, but not superior to" review with continuous active learning, the standard TAR 2.0 workflow, whether the final responsiveness call was made by the tool or by a first-level reviewer. When the human reviewer made the final call, active learning reached the same recall with less review effort. Most strikingly, a simple S-CAL protocol trained on just 2,000 reviewer determinations matched aiR's effectiveness, offered a calibrated trade-off between recall and precision, and "runs quickly and cheaply on commodity hardware," while aiR offered only two viable recall-precision operating points with no calibrated way to choose between them, on slow, specialized infrastructure.

Three caveats apply, and they cut in both directions. First, that study examined responsiveness review, not privilege review; privilege adds harder questions about roles, purpose, and waiver on top of the same engine. Second, these results concern one tool's version at one point in time. Third, the underlying measurement problem is genuinely hard: on the TREC 2009 legal benchmark collection of 569,034 emails, relevant-document richness ran from 0.3 to 1.5 percent. When the thing you are hunting is that rare, small validation samples produce noisy recall estimates, for AI and human review alike. The same skepticism we applied to vendor accuracy claims in AI contract review applies here: ask what the denominator was, who graded the answer key, and whether anyone independent has reproduced the number.

When AI privilege review is the wrong choice

A checklist for saying no, or not yet:

• The collection is small. Under a few thousand documents, protocol negotiation and validation overhead can exceed what you save. Just have associates read it.
• The privilege questions are novel. Common-interest disputes, dual-purpose communications, or regulator-specific doctrines that cannot be written down as plain-language criteria will produce confident garbage.
• The other side is looking for a fight. If opposing counsel will attack any methodology, a methodology battle can cost more than the review. Sometimes manual review is the cheap option once motion practice is priced in.
• The collection is hostile to text models. Scanned handwriting, heavy redaction, voicemail, or fragmented chat exports degrade performance, and 17% of providers quietly bill these back at manual rates.
• You cannot get the data terms you need. The Trustpoint example ran under a zero-data-retention agreement. If a vendor will not commit to equivalent terms, or your client is a bank or public body with data-residency rules, the correct move is a private deployment on infrastructure you control, which is the model we use in our enterprise AI work for regulated clients, not a consumer API.

Common mistakes

• Buying the demo percentage. Ask whether the claimed savings include criteria drafting, QC sampling, log editing, and protocol negotiation. The 61% Trustpoint figure is credible precisely because it describes the whole matter workflow.
• Skipping the 26(f) disclosure. An undisclosed AI methodology discovered mid-case invites motions to compel and, worse, re-review.
• Treating the log as free. Auto-drafted entries still need attorney editing. A description that reveals too much can waive the privilege; one that says too little draws a challenge. Budget for the edit.
• Validating on easy documents. A random sample of obvious invoices tells you nothing. Your QC sample must be enriched with borderline documents, the ones near the decision boundary.
• Assuming one tool wins everything. The evidence supports a hybrid: keyword screens for obvious sets, active learning where it is cheaper and now demonstrably competitive, generative AI where written rationales and plain-language criteria add value. Thirty percent of providers already price hybrid models for exactly this reason.

Frequently asked questions

Is AI privilege review defensible in court? There is still no landmark ruling blessing it in the abstract, but in Schulte v. LinkedIn (July 2026) a federal court accepted a GenAI review workflow where the use was disclosed in advance and human QC sampling was documented, and it refused to order discovery into the tool's error metrics. Defensibility comes from process: a disclosed protocol, documented criteria, validation sampling, and attorney sign-off on the log.

How much does AI privilege review cost per document? Survey data clusters around $0.11 to $0.50 per document for the AI pass, versus $4 to $8 per document for a human privilege read, plus $5 to $15 per log entry. Total matter savings of 50 to 70 percent are realistic once QC and log editing are included.

Can AI draft the privilege log? Yes, tools like aiR for Privilege draft log entries from their structured output, and this is where much of the saving lives. An attorney must still edit every entry to the Rule 26(b)(5)(A) standard: enough detail to allow assessment, not enough to disclose the content.

Does generative AI replace TAR? No. Independent testing found active learning matched aiR's recall with less review effort, and a simple S-CAL protocol trained on 2,000 documents matched it outright. The two approaches answer different needs; serious shops use both.

What about client confidentiality when documents go through an LLM? Insist on zero-data-retention terms in writing, as in the Relativity-Microsoft arrangement Trustpoint describes. For clients with strict residency or secrecy requirements, run open-weight models on infrastructure you control instead of sending data to a shared API.

Where to go from here

If you are scoping AI-assisted privilege review for a live matter or building a repeatable workflow for your firm, the hard parts are the protocol, the criteria, and the validation design, not the model. Our Claude for Legal practice implements Anthropic's legal tooling for law firms and in-house teams, including review workflows with the audit trail courts will expect. Talk to us about your next matter.