The Redgrave LLP case study “Generative AI for Complex Document Review” provides a unique opportunity to explore potential workflow options to project the comparative results from a cost, time, risk and effort perspective. Clients and peers are wrestling with ‘Is AI really worth it?’. That question has led me to adapting my validation and ROI models to test aiR for Review against closed client traditional linear or Relativity Active Learning (RAL) reviews. The well documented Redgrave study metrics allow me to use general market billable and AI tool rates* with different scenarios. Enjoy my projections and download my worksheet if you will take an anonymous quiz.

The Case for 2^nd Pass Review (+2^nd Pass)

The paper’s pure aiR for Review approach saves 98% of billable and overall project time over traditional or RAL approach IF you assume a single RAL review and QC pass. If a defensible production was the only review goal, that might be the end of the story. In my view, the goal of eDiscovery is matter resolution. Resolution requires counsel to have key evidence, derive facts and apply the law. AI generated document summaries and fact chronologies must be checked against the actual documents by the case team to realize their value, not just contract review attorneys. This has traditionally been done by firm associates who surface the key ‘hot docs’ to lead counsel. The case study omits this step, where I believe much of the analysis value is generated. While contract reviewers can certainly flag potential hot docs, we do not have any study metrics on RAL pass that would support a more complex tiered model. Therefore, I created a couple of potential workflow models (+2^nd Pass models) to go beyond hours to actual costs. This is overly simplified and I would expect a more efficient tiered approach in a real world review.

The Case for a Hybrid aiR/RAL Review

Because the Redgrave team kindly provided the aiR For Review scoring metrics for the Validation review, I built a hybrid tiered model that directed Very Responsive (4) documents directly to counsel, Responsive (3) to associate team and Borderline (2) documents to the RAL team. This seems like a balanced approach that immediately converts the highest value evidence into Facts. There are multiple other approaches including diversion to aiR for Case Strategy, analysis of the aiR summary/topics for prioritization or additional validation sampling. The paper did not contemplate those, so I will save them for my client cases.

*Billable rates and AI costs vary wildly. For the purposes of these models, I assumed a high-risk matter that justified AmLaw 100-200 retained counsel team and a global service provider for AI and contract reviewer rates. I used Claude to validate the rate estimates I came up with from recent RFP and invoice analysis engagements. I took the average publicly available per document AI review tool rates rather than use any client negotiated rates. My AI homework is in the appendix.

The Review Workflow Overviews:

1. Traditional linear review – Single 1^st Pass review with contract team. The baseline of minimum effort. I did not bother with QC or management cost on this model. Recall and precision rates from Herbert L. Roitblat, Anne Kershaw & Patrick Oot, “Document Categorization in Legal Electronic Discovery: Computer Classification vs. Manual Review,” 61 J. Am. Soc’y for Info. Sci. & Tech. 70 (2010).
2. Relativity Active Learning(RAL) – Single pass prioritized review with contract team per paper. Used exact hours from paper.
3. RAL + 2^nd – Adding a 2^nd pass associate team review of all flagged relevant by RAL so that the entire production set has “eyes on” by those more directly involved in the legal analysis. There are many scenarios where long term or in-house contract review teams are fully sufficient for 1^st pass with validation. I struggled to stick to the study parameters and do an apple-apple in the model. If you want to do comparisons that more closely resemble your workflows download the workbook and have at it. Send it in and I will publish it.
4. aiR for Review – Prompt developed by partner and model applied to full document set. . I have recently done the aiR prompt development and validation on recommended richness collections for closed cases for overall accuracy and ROI analysis.  I believe that prompt development could be done in less time with a high-quality case summary, review protocol and other ‘Prompt Kickstarter’ materials. I appreciate Redgrave’s effort creating an appropriate academic scenario for a public data set. This is just easier with real cases.
5. aiR + 2^nd – Takes the rather bloated AI predictions and does a 2^nd pass associate review of all potential production documents. Again, this is the “eyes on” by counsel for everything going out the door approach. It does not address the elusion issue. I assumed a 4 associate team for ‘Time to Completion’. That review team size and rate assumptions can be changed in the model.
6. Hybrid – My home-baked scenario leveraging the aiR scores to redirect documents per above paragraph. It is my attempt at one possible practical, balance review workflow in light of the very low elusion rate found in the study.

Time-Cost Model Comparison

Notes:

• Team size and review rates can be modified to adapt the days to completion.
• Prompt and Validation review time is only added to aiR, not RAL (even though validation should be).
• “Time to Completion” is a team-adjusted, rolling-production estimate (Contract team = 24, Associate team = 4, Counsel = 1, all at 8 hrs/day), not a simple hours ÷ 8 calculation — that’s why RAL’s 1,123 hours only takes 5.8 days calculated instead of 140. The study reported a little over 7 days for RAL. Reviews are complicated and always take longer than expected.
• aiR’s 4.8 days is prompt-development and validation labor only — the actual population-wide scoring run is a separate ~1-hour automated pass, priced per-document rather than in hours.
• aiR + 2nd Pass takes longer than Hybrid despite similar total hours (302.7 vs. 317.5) because aiR + 2nd is a strict sequential chain (aiR must fully finish before the eyes-on pass can start), while Hybrid’s three review tiers run concurrently once aiR’s base work is done.
• Richness factor Redgrave team has made the case (correctly) that RAL and aiR performance would improve dramatically with recommended 20-30% richness collection. That would bring RAL cost down, but not base aiR cost. The improved precision would bring the 2^nd pass review cost down as well.
• aiR vs. RAL review rates SME’s at Relativity believe that aiR for Review coding rates are significantly higher than RAL/traditional rates. Without published case studies or first person experience of those performance gains, I kept the model rates based on the study.

It is interesting that RAL+2^nd is faster and cheaper than aiR+2^nd. Those savings come with an increased risk of elusion. RAL also needs to solve its human reviewer error issue highlighted by Redgrave. No matter how you slice it, the pure aiR first pass review is the fastest and cheapest model if counsel is willing to sign that certificate of completion without having looked at most of the documents. I have not found that attorney yet, but I have been told that it happens with tech-savvy top counsel and excellent teams. There are MANY scenarios such as non-adversarial productions, very low risk collection content and court protections where I could see a pure aiR workflow with some safety steps being used.

Accuracy Metrics Comparison

Notes:

• Because we now have eyes on every document flagged potentially relevant by RAL or aiR in 2^nd pass models their precision becomes “100%”. We all know that is not reality, but I did not want to vary from the study metrics. The Redgrave study has an excellent section on this issue. 2^nd pass fixed precision, but it cannot fix Recall.
• The True Positive numbers are ALL below the 3,150 paper’s 3,150 true relevant estimate. That is the reality of subjective review. The actual production volume is True Positive + False Positive.
• NONE of these models address what George Socha and Donald Rumsfeld once called the ‘unknown unknowns’ hidden in the False Negatives. We all have our preferred QC methods such as sampling, similarity, conceptual clustering and others to find these. This is a risk and reasonable effort decision best made with counsel and outside the bounds of this exercise.

My Takeaways

• Properly applied AI is just better in many scenarios. Elevate’s Jeremy Pickens has done an excellent analysis of those potential scenarios and when he thinks RAL is a better choice.
• If you are going to do eyes on review, use RAL to move the more relevant docs to the front of the line.
• Think outside the box with these tools. Use them in combination and on selective subsets to meet your goals with the minimum of time and cost.
• Assess collection richness frequently and use traditional criteria, filters and such to eliminate the easy non-relevant before launching review.
• Test your workflows on your own historical reviews. You do not need to run aiR on the entire set to test prompt techniques and identify unique factors in your collections.
• Document your testing for defensibility and to get counsel approval.
• Know the strengths and limitations of your chosen AI tool. Most have size volume limits, work best within a set extracted text size range and may require specific email or document fields.

You can download my worksheet from HERE. I have two quick, optional questions for those that choose to share their experiences with AI review. No pay wall and you do not have to register or give me your email unless you want to discuss eDiscovery. I hope that you enjoyed going down this rabbit hole with me. Now that I have returned to coordinating cases for clients and doing expert work, I have to fight for research time (unpaid). You can next catch me at David Horrigan’s Legal Data Intelligence State of the Union panel at RelFest 2026 or just send me a note.

Greg Buckles wants your feedback, questions or project inquiries at Greg@eDJGroupInc.com.  Reach out for a free 15 minute ‘Good Karma’ call if he has availability. He solves problems and creates eDiscovery solutions for enterprise and law firm clients.

Greg’s blog perspectives are personal opinions and should not be interpreted as a professional judgment or advice. Greg is no longer an investigative journalist and all perspectives are based on best public information. Blog content is neither approved nor reviewed by any providers prior to being published. Do you want to share your own perspective? Greg is looking for practical, professional informative perspectives free of marketing fluff, hidden agendas or personal/product bias. Outside blogs will clearly indicate the author, company and any relevant affiliations. 

Greg’s latest nature, art and diving photographs on Instagram.

Append 1: Market Rate Research (Claude generated):

Here’s what’s publicly available, organized by role, with links and the interpretive calls I had to make along the way.

AmLaw 200 attorney rates (Partner / Counsel / Associate)

Links:

• Wolters Kluwer Real Rate Report / LegalVIEW Insights: https://www.wolterskluwer.com/en/solutions/enterprise-legal-management/legalview-analytics/real-rate-report and https://www.wolterskluwer.com/en/news/wolters-kluwer-study-sets-the-stage-for-2026-with-sharp-contrasts-in-regional-law-firm-rates
• Legal Dive summary of the 2022 Real Rate Report (national partner/associate/paralegal averages): https://www.legaldive.com/news/153-cost-difference-between-nyc-phoenix-lawyers-Wolters-Kluwer-ELM-Solutions/634079/
• Major Lindsey & Africa 2024 Partner Compensation Survey (via Global Legal Post): https://www.globallegalpost.com/news/amlaw-200-partner-pay-hits-record-14m-1042659731
• Valeo 2026 Early Indicators AmLaw 200 Hourly Rate Report (the report itself is paid; this is the public press description): https://www.valeopartners.com/product/valeo-2026-early-indicators-amlaw-200-hourly-rate-report/
• LegalBillReview.com 2026 benchmarks summary: https://www.legalbillreview.com/blog/legal-billing-rates-2026-benchmarks-trends

Interpretation call: these are standard/billed rates before any negotiated discount — real client-paid rates run lower, and Wolters Kluwer’s own commentary notes actual realization is typically well under standard rate. Your invoice-based numbers, net of discounts, are almost certainly more accurate for a specific engagement than these market-wide averages.

Contract attorney / document review rates per Claude

This is the one where public sources genuinely diverge, because “contract attorney rate” means three different things depending on who’s being asked — what the reviewer is paid, what the staffing agency bills the firm, and what the firm bills the client. I found data on all three:

Interpretation calls I made:

• The $85/hr HDR figure and the 50-docs/hr throughput assumption directly back up your model’s review-speed and associate-rate assumptions — this is a genuinely useful anchor since it’s a rate a federal court found reasonable, not just a vendor quote.
• The Citigroup case shows the ceiling courts will tolerate ($200/hr, post-reduction) for pure first-pass document review by contract attorneys, even in a case with a $590M common fund — useful if you want a defensible upper bound for your “eyes-on associate” rate rather than a floor.
• None of these sources cleanly separate “contract attorney” from “associate” the way your model does — court opinions and vendor surveys use “contract attorney,” “reviewer,” and sometimes “associate” loosely, so treat these as directional, not a precise match to your rate tiers.

eDiscovery vendor / GenAI per-document rates

This last one is the best news for your workbook: your $0.25/doc aiR assumption sits almost exactly in the middle of the actual current-market GenAI per-document range ($0.11–$0.50) reported by 34 real respondents in a Q4 2025–Q1 2026 field survey — that’s about as close to a public backstop as you’re going to get for a number Relativity doesn’t publish. I’d cite the ComplexDiscovery/EDRM survey specifically for that rate line in your workbook notes, since it’s free, dated, methodologically disclosed, and directly on point.

One honest caveat across all of this: none of these public sources are a substitute for your own RFP and invoice data — they’re broad market surveys and litigated-fee opinions, not benchmarks calibrated to opioid-MDL-scale document review with a compliance-heavy responsiveness standard. They’re best used the way court opinions use them: as corroboration that your numbers sit within a defensible range, not as the primary source for the number itself.