A done-for-you, attorney-supervised litigation review service that ingests a matter's document population and returns a defensible produced set — responsiveness coding, privilege review, a court-ready privilege log, redactions, and a documented validation record supporting counsel's Rule 26(g) certification. AI does the extraction, coding, drafting and QA; licensed review attorneys own the judgment chokepoints. Priced per gigabyte reviewed and per privilege-log entry — never hourly.
Run timestamp: 2026-07-02 22:05 UTC · Slug: ai-native-managed-document-review-engine · Decision: BLUEPRINT
2. Final decision
DECISION: BLUEPRINT — Proceed to a privilege-review pilot.
Document review is the single largest, most-outsourced, most budgeted line item in litigation — 60–75% of all eDiscovery spend and, on ABA figures, roughly 80% of total litigation cost. A mature managed-review vendor market already proves the budget exists. Generative-AI review is going mainstream in 2025–2026 exactly as data volumes explode. The work is mostly structured with a small number of genuine judgment moments (privilege edge cases, protocol design, validation sign-off), and unauthorized-practice-of-law rules require attorney supervision — which both protects margin and blocks casual entrants. It clears the evidence threshold and triggers no fatal disqualifier. The one real threat — platforms bundling GenAI review for free — is addressed head-on in §29.
3. Executive summary
~$15–20B
global eDiscovery spend, 2025 (estimates range $15.4B–$19.6B) V
60–75%
of eDiscovery cost is document review — the largest driver V
~$9.8→13.6B
review spend growing through 2028 V
73%
of discovery costs are review (RAND) V
47%
of US legal departments outsource litigation support / eDiscovery V
$15–30/GB
prevailing review & production pricing (2026) V
When a company or law firm is sued or subpoenaed, someone has to read every potentially responsive document, decide what is relevant, and — most dangerously — decide what is privileged and must be withheld or redacted before it is produced to the other side. Get privilege wrong and you can waive attorney-client protection, hand your opponent a smoking gun, and face sanctions or a malpractice claim. This work is traditionally done by armies of contract attorneys billed per hour through Alternative Legal Service Providers (ALSPs) — slow, expensive, and inconsistent.
This business sells the outcome, not a tool: "hand us the documents and the protocol; we return a defensible, privilege-protected, production-ready set with the validation record your counsel needs to certify it." AI performs first-pass responsiveness and privilege scoring, entity and issue tagging, redaction proposals, and privilege-log drafting; licensed review attorneys design the protocol, resolve edge cases, sign off on privilege, and own defensibility. Pricing is per gigabyte reviewed plus per privilege-log entry — aligned to volume and to the catastrophic risk it removes, and structured so that as models improve, human minutes per gigabyte fall and margin expands.
The wedge is the scariest, least-commoditizable slice — privilege review and logging — for the under-served mid-market: corporate legal departments and mid-size litigation firms that lack an in-house review team and are priced or paced out by the largest ALSPs.
4. Thesis
Litigation document review is a high-volume, rules-heavy, documents-in / coded-set-out workflow with a small number of true judgment moments. That is precisely the shape of work that frontier models plus a narrow attorney chokepoint can industrialize. The incumbents (Consilio, Epiq, Cimplifi, TCDI) proved the budget and the defensibility playbook, but their cost structure is still contract-attorney hours; the platforms (Relativity, Everlaw, DISCO) are commoditizing the tool but deliberately not selling the supervised, certified outcome. The gap is a service that (1) absorbs the whole task, (2) prices per outcome unit rather than per hour, (3) owns the defensibility documentation an opponent and a judge will scrutinize, and (4) keeps a licensed attorney at the privilege chokepoint so it never crosses the unauthorized-practice-of-law line. As models get better at first-pass responsiveness and privilege detection, the human QC minutes per gigabyte fall — the service gets cheaper, faster, and more defensible, not obsolete.
5. Discovery rationale
This run started with no preselected idea. Search terrain spanned legal operations, insurance back-office, financial compliance, and healthcare administration. Document review surfaced on the deciding signals: (1) it is the largest single cost in its category (review = 60–80% of litigation/eDiscovery spend), so even a small share is a real business; (2) it is already massively outsourced to a mature ALSP market, so budget is proven, not hoped-for; (3) it is at a technology inflection — GenAI review moved into operational workflows in 2025–2026; (4) it has a regulatory moat (UPL supervision rules, FRCP 26(g) certification, privilege-waiver stakes); and (5) it has a clean whitespace — the mid-market is under-served relative to big-pharma-grade ALSP engagements, and only ~32% of legal teams use ALSPs at all. Against the ~90 prior blueprints in this workspace, none covers eDiscovery or document review — the closest legal entries (PI demand packages, patent prior-art search, contract/lease abstraction) are different buyers, workflows and outcomes.
6. Candidate comparison
Five candidates were generated and scored 1–5 (higher is better). Winner selected on evidence, not preference.
Candidate
Demand evidence
Budget / competitor proof
MVP narrowness
Margin path
Licensing safety
Novelty vs prior ~90
Total
Managed document review & privilege engine ★
5
5
5
5
4
5
29
Early case assessment (ECA) triage & scoping service
The winner leads on the two hardest-to-fake dimensions — active demand and existing budget — and on margin path (GenAI collapses the dominant labor cost). Translation was penalized on anti-commoditization (self-certified translation is often accepted); legal-hold was penalized on weaker willingness-to-pay and thin per-unit pricing.
7. CODE validation
C — Consumer / buyer trend
Discoverable data is exploding (chat, Slack/Teams, mobile, collaboration tools), while GenAI review moved from pilot to production in 2025–2026 and corporate legal budgets are rising — 61% of GCs expected larger 2025 budgets, with a stated preference for outcome-based external engagements. Simultaneously, the December 2025 FRCP amendments and a federal split on AI-and-privilege are pushing defensibility and validation to the center of every review. V
O — Opportunity
Legacy managed review still prices on contract-attorney hours — expensive, slow, variable. The mid-market (corporate legal departments and mid-size firms) is under-served: the biggest ALSP engagements are built for AmLaw-100 and Fortune-500 matters, and only ~32% of legal teams use ALSPs at all. Platforms hand buyers a GenAI tool but not the supervised, certified, done-for-you outcome — leaving the actual job (a defensible produced set) unowned. V / I
D — Demand
47% of US legal departments already outsource litigation support and eDiscovery; document review is ~34% of the outsourced legal workload; a mature vendor market (Consilio, Epiq, Cimplifi, TCDI, Percipient) and a recurring industry pricing survey (ComplexDiscovery, Winter 2026) prove buyers are actively spending and actively price-shopping this exact service. V
E — Economic sizing
eDiscovery spend ~$15.4B–$19.6B (2025); review is 60–75% of that (~$9.8B–$13.6B and growing through 2028). If an AI-native provider captured even 0.5% of the review pool at 55–65% gross margin, that is a $50M+ revenue business. The realistic serviceable beachhead — mid-market commercial/employment matters, privilege-first — is a low-single-digit-billion slice. V / I
8. Rubric scorecard
Gate
Score
Reasoning
1. Low trust burden / already outsourced
5
Managed review is one of the most-outsourced tasks in law; buyers routinely hand documents to third-party ALSPs and contract attorneys.
2. Low task-level judgment
4
Most documents auto-code on responsiveness and clear privilege; genuine judgment concentrates in privilege edge cases, protocol design, and validation sign-off.
3. High intelligence threshold
5
Requires synthesis across documents, privilege doctrine, work-product rules, responsiveness criteria, custodian context, and edge cases — a frontier-model strength paired with expert review.
4. Regulation as a moat
5
UPL rules require attorney supervision of selection/privilege; FRCP 26(g) certification, privilege-waiver stakes, sanctions exposure, and protective/502(d) orders raise willingness to pay and deter casual entrants.
5. No physical / on-site labor
5
Entirely data- and document-based; delivered remotely inside a secure review environment.
6. Sam Altman test
5
Better models raise first-pass recall/precision on responsiveness and privilege, cut human QC minutes per GB, and improve the validation record. The service strengthens as models improve.
Composite: 29/30. The only sub-5 is task-level judgment, by design — the privilege chokepoint is where a licensed human must stay.
9. Target buyer
Economic buyer: the person who owns litigation cost and risk and signs the engagement — a corporate Associate General Counsel / Head of Litigation / Director of Legal Operations at a mid-market company, or the litigation partner / practice-group leader at a small-to-mid law firm. Champion / day-to-day contact: the eDiscovery / litigation-support manager, senior litigation paralegal, or the associate running the matter.
Segment (beachhead first)
Why they buy
Trigger
Mid-market corporate legal departments (200–5,000 employees) with recurring employment/commercial litigation but no in-house review team
They get sued often enough to feel the cost, but not enough to staff review; they want the risk off their desk
New complaint served; litigation hold issued; subpoena or agency document request received
Small-to-mid litigation firms (5–75 attorneys) that lack a dedicated review/eDiscovery team
They must review to meet a court deadline but can't defensibly staff a large review overnight
Discovery deadline set; rolling-production schedule ordered; motion to compel looming
Insurance-defense & employment-defense firms with high matter volume
Per-matter review is a repeatable, margin-sensitive cost center
New assigned defense matter with document production obligations
Overflow channel: existing ALSPs / eDiscovery service bureaus
They need surge capacity and a lower-cost first pass to protect their own margins
Review peak they can't staff; RFP they'd otherwise decline
10. Jobs-to-be-Done
Functional: "Review this document population, code it for responsiveness, catch and withhold everything privileged, redact what needs redacting, and give me a production-ready set and privilege log by the court deadline."
Risk: "Make sure I never inadvertently produce a privileged document, and give me a validation record my outside counsel can certify under Rule 26(g) and defend at a meet-and-confer or a motion to compel."
Economic: "Do it for meaningfully less than a room full of contract attorneys, with predictable per-unit pricing I can pass through or budget."
Emotional: "Let me stop lying awake about a privilege waiver and a sanctions motion, and stop begging for staffing I can't find."
Social: "Let me look buttoned-up and defensible in front of the judge, the client, and opposing counsel."
11. Painful problem
Document review is the "elephant in the eDiscovery room": it consumes 60–75% of eDiscovery cost and, on ABA estimates, roughly 80% of total litigation spend (~$42B/yr). The pain is not just cost — it's risk concentrated in privilege. A single inadvertently produced privileged email can waive privilege over an entire subject matter, hand the opponent decisive evidence, and trigger a malpractice claim against counsel. Traditional relief — spinning up contract attorneys through an ALSP — is slow to staff, inconsistent between reviewers, and billed per hour, so cost scales linearly with data volume in an era when data volume is exploding. Mid-market buyers feel this most acutely: enough litigation to be painful, not enough to justify permanent review infrastructure, and too small to be a priority client for the largest ALSPs.
"The cost of review still accounts for roughly 75% of the total cost of an eDiscovery project… document review sits at the commercial center of most eDiscovery engagements and is the largest cost driver in complex litigation." — 2026 pricing analyses V
12. The outcome we sell
We do not sell a review platform, an AI copilot, or seats. We sell a defensible produced set and privilege log, delivered on deadline, with the validation record counsel needs to certify it. The customer's experience is a service relationship with a named review-attorney lead, not a login. Concretely, per matter the client receives:
A responsiveness-coded document set (relevant / not relevant, with issue/hot-doc tags).
A privilege call on every document (privileged / not / partially — withhold or redact), reviewed and signed off by a licensed attorney.
A court-ready, formatted privilege log keyed to the applicable rules and the governing protective/502(d) order.
Proposed redactions applied and QC'd.
A defensibility packet: the review protocol, model/validation metrics (recall, precision, elusion, sampling), audit trail, and a certification-support memo for outside counsel.
A production-ready deliverable in the client's platform/format (or ours), plus a claw-back register.
13. First one-feature MVP wedge
ICP: Mid-market corporate legal department OR small-to-mid litigation firm with a live commercial or employment matter and no in-house review team.
Trigger event: A discovery deadline with a document population of ~10–200 GB that must be reviewed for privilege before production.
Pain: They cannot defensibly staff a privilege review fast enough, and a privilege waiver is a catastrophic, career-level risk.
One-feature MVP: Done-for-you privilege review + privilege log (not full responsiveness review) — the scariest, highest-value, least-commoditizable slice.
Input: Processed document set (we process if needed) + a privilege protocol (privilege criteria, key custodians, attorney/firm names, date ranges, the 502(d) order).
Output: Withhold/redact set + formatted privilege log + validation record + attorney certification-support memo.
Human chokepoint: A licensed review attorney reviews all model-flagged privilege calls and edge cases and signs off on the log.
Success metric: Zero inadvertent privileged productions (claw-back-free), privilege log accepted at meet-and-confer, and turnaround inside the deadline at a per-GB price below legacy managed review.
What users ask for next: full responsiveness review, redaction-at-scale, ECA/early scoping, second-request/investigation surge, and a standing per-matter arrangement.
14. Evidence summary
The strongest signals are the ones hardest to fake: a mature, budgeted vendor market; a recurring industry pricing survey; review's dominant share of a $15–20B category; and 47% of legal departments already outsourcing this work. The technology-inflection signals (GenAI review at 90%+ vendor-reported accuracy; 80–85% recall as the defensibility benchmark; platforms bundling GenAI in 2026) confirm both the opportunity and the central threat. Regulatory signals (UPL supervision; FRCP 26(g); December 2025 FRCP amendments on AI privilege) confirm the moat and the licensing boundary.
15. Claim table
Claim
Label
Note
Global eDiscovery spend was ~$15.4B–$19.6B in 2025 (estimates vary by firm/scope)
The market is large, growing, and — critically — dominated by the exact line item this business owns. Review's share of spend (60–75%) means the addressable pool inside eDiscovery is ~$10–14B and rising. Demand is not speculative: nearly half of legal departments already outsource this work, a recurring pricing survey tracks it, and a top-100 provider list exists. The AI-enabled subsegment is forecast to compound at ~27% — the tailwind and the threat both live here.
18. Active buyer conversations
Industry pricing surveys (ComplexDiscovery Winter 2026) exist precisely because buyers actively compare per-GB and per-hour review prices — proof of an active, price-shopping market.
Law-firm client alerts from Akin, K&L Gates, Greenberg Traurig, BakerHostetler on AI-and-privilege in 2025–2026 show GCs and litigators anxious about defensibility right now.
Vendor "free GenAI" announcements (Relativity, Everlaw) and the ensuing "price reset for small firms" commentary show the mid-market actively re-evaluating how review gets bought.
ALSP marketing pages (Consilio, Epiq, Cimplifi, TCDI, Percipient) selling "managed document review" and "predictable pricing" confirm buyers ask for a managed outcome, not a tool.
ACEDS / Sedona / eDiscovery communities and LinkedIn litigation-support groups are full of practitioners debating GenAI review defensibility, validation, and staffing — a warm, reachable audience.
19. Competitive landscape
Category
Examples
What they do
Gap we exploit
Large ALSPs / managed review
Consilio, Epiq, Cimplifi, TCDI, Percipient
Contract-attorney-led managed review at scale; predictable pricing packages
Cost structure is human hours; built for large matters; mid-market is a lower priority; GenAI adoption uneven
Review platforms (GenAI-bundled)
Relativity (aiR), Everlaw (EverlawAI), DISCO (Cecilia)
Sell the software/tool; bundling GenAI review into subscriptions in 2026
They sell a tool, not a supervised, certified, done-for-you outcome; buyer still needs attorneys + defensibility ownership
AI-review point tools
eDiscovery AI, Harvey, GC AI
GenAI review add-ons / assistants
Copilots the buyer operates; not a managed service with attorney sign-off and a certification packet
Contract-attorney staffing
Staffing agencies, LOD/ElevateFlex
Supply reviewers by the hour
Linear cost, variable quality, slow to staff — the incumbent budget we redirect
The whitespace is explicit: nobody in the mid-market sells the defensible outcome at a per-unit price with a licensed attorney owning privilege. Platforms deliberately avoid the service; ALSPs price on hours; staffing supplies bodies.
20. Competitor and budget validation
Existing budget source: litigation budgets already fund managed review — 47% of legal departments outsource litigation support/eDiscovery, and review is ~34% of the outsourced legal workload. Companies pay ALSPs and staffing agencies today; we redirect that spend.
Why current alternatives are insufficient: contract-attorney review is expensive and slow to staff; platform tools shift operating burden and defensibility risk back onto the buyer; the largest ALSPs under-prioritize mid-market matters. None combine AI-native cost structure + attorney-supervised defensibility + per-unit outcome pricing.
Why we can win: a GenAI-first first pass collapses the dominant labor cost, letting us undercut contract-attorney review on price while improving consistency and defensibility, and we sell the outcome (not seats) with an attorney owning privilege — the one thing platforms won't do and staffing can't guarantee.
Why we are not a clone: we are not another ALSP billing hours, and not another platform selling a login. We are a per-unit, outcome-priced, attorney-supervised AI-native service whose margins expand as models improve.
21. Pricing evidence and proposed pricing
Evidence: prevailing review & production runs ~$15–30/GB; onsite contract-attorney review commonly >$40/hr; hosting $5–15/GB/mo; processing $3–10/GB. Buyers already think in per-GB and per-document units — never our own hours.
Proposed pricing (per-unit / outcome — never hourly)
Unit
Price (illustrative)
Rationale
Privilege review, per GB reviewed (MVP)
$18–28 / GB
Below blended legacy managed-review cost; anchored to $15–30/GB norm; margin comes from low AI COGS
Privilege-log entry, per logged document
$3–8 / entry
Logging is high-value, judgment-adjacent, and discretely countable
No hourly billing. Volume risk is bounded by scoping (ECA) before commitment. Contingency/success-fee pricing is explicitly avoided in litigation review (fee-splitting and ethics concerns) — see §22.
22. Regulatory and compliance considerations
Unauthorized practice of law (UPL): ABA Model Rules 5.3/5.5 and bar ethics opinions limit a non-lawyer vendor's role to administrative, technical, and logistical tasks; selecting responsive documents, making privilege calls, and preparing privilege logs are legal judgments that require attorney supervision. Our privilege calls are made or reviewed and signed off by licensed attorneys.
FRCP 26(g): the producing party's counsel certifies discovery responses; our deliverable must supply the validation record (recall, precision, elusion, sampling) that supports that certification. We support — we do not replace — outside counsel's certification.
Privilege / waiver & FRE 502(d): reviews run under the matter's 502(d) claw-back order; we maintain a claw-back register and a documented privilege protocol.
Defensibility & AI disclosure: some judges have standing orders on AI use; the December 2025 FRCP amendments and a federal split on AI-and-privilege raise documentation requirements. Defensibility is a documentation/process problem — our packet is the answer.
Data security & confidentiality: matters contain privileged and PII/PHI data; SOC 2 Type II, encryption, access controls, protective-order compliance, and no-model-training-on-client-data commitments are table stakes.
Cross-border: GDPR / data-residency and, for some matters, foreign blocking statutes require regional handling.
Administrative, technical, logistical work: processing, QC of formatting, sampling execution, log assembly, production mechanics
Licensed review attorneys (ours)
Barred attorneys under supervision structure
Design/approve the protocol; make/review privilege calls; resolve responsiveness and privilege edge cases; sign off on the log
Client's outside counsel
Client-side barred attorney
Owns the matter, the strategy, and the Rule 26(g) certification; retains final authority over production
Must not claim: to "practice law," give the client legal advice on the merits, or replace outside counsel's certification. Required: engagement terms defining attorney supervision; documented protocol and audit trail; conflicts screening; no fee-splitting with non-lawyers; explicit statement that final certification remains with the client's counsel. Pricing legality: per-unit and flat pricing are safe; contingency/recovered-dollar pricing is avoided in litigation review due to fee-splitting and champerty concerns.
24. AI-native advantage
AI does not merely assist here — it changes the economics. In legacy review, cost is dominated by human reading time that scales linearly with data volume. In the AI-native model, a GenAI first pass (with per-document reasoning) scores responsiveness and privilege, drafts log entries, and proposes redactions, so humans spend their minutes only on flagged calls and edge cases. As models improve, that flagged fraction shrinks — the same deliverable costs fewer human minutes per GB.
Task type
Owner
OCR, extraction, de-dup, email threading, language ID
Automation
Responsiveness & privilege scoring with rationale; issue tagging; PII/PHI detection
Retrieval & knowledge layer: matter-specific index; counsel/law-firm name lists; privilege doctrine references; prior calls from this matter for consistency.
AI workbench layer: LLM scoring of responsiveness and privilege with rationale; redaction proposals; log-entry drafting; hot-doc and issue tagging.
Deterministic rules layer: hard rules (e.g., any doc to/from named outside counsel domains → privilege candidate; date-range cutoffs; auto-redact rules for PII patterns).
Human chokepoint layer: attorney review queue for all privilege candidates, low-confidence calls, and conflicts between AI and rules.
QA layer: statistical sampling of AI-only calls; recall/precision/elusion; second-attorney QC on privilege withholds; discrepancy resolution.
Learning loop: every attorney correction updates matter rules, prompts, exemplars, and QC checks; recurring patterns become SOPs.
Model-portability layer: model-agnostic prompt/eval harness so we can swap or route across frontier models as they improve, and validate any model against a gold set before use.
26. AI-vs-human operations pipeline
AI Automation Deterministic rule Human chokepoint (licensed attorney)
Automation
Ingest & process; OCR; de-dup; thread families
Rule
Flag docs touching counsel/firm domains & date ranges
AI
Score responsiveness & privilege with rationale; tag issues/PII
AI
Draft privilege-log entries; propose redactions
Attorney
Review privilege candidates & edge cases; sign off
Render redactions; format log; assemble defensibility packet
Attorney
Final QC & release for counsel's certification
27. Dynasty translation layer
1. Buyer translation
Who pays: AGC/Head of Litigation or litigation partner. Urgent problem: a discovery deadline and a privilege-waiver risk they can't defensibly staff. Outcome wanted: a produced set that is on time, defensible, and claw-back-free.
2. Service translation
Done-for-you managed review. Client receives a coded, privilege-reviewed, production-ready set + log + defensibility packet. Automation handles processing/scoring/drafting; licensed attorneys own privilege and sign-off.
3. Workflow translation
Intake & protocol → process/normalize → AI score & draft → attorney review → QA/validation → deliver → claw-back & renewal for the next production wave.
4. Tooling translation
Start on an established review platform (RelativityOne / Everlaw) + an LLM eval/prompt harness + a privilege-log/validation generator. Build proprietary orchestration and QA later; buy processing/hosting first.
5. Sales translation
"You have documents and a deadline. Hand us the population and the protocol; we return a defensible, privilege-protected production and the record your counsel needs to certify it — at a per-GB price below a contract-attorney room, and faster."
6. Delivery translation
MVP delivered manually/semi-manually: attorneys reviewing an AI-scored queue on a hosted platform, log generated from a template. Automate orchestration, QA sampling, and packet assembly over time.
7. Expansion translation
Privilege review → full responsiveness → redaction-at-scale → ECA/scoping → second-request/investigation surge → standing per-matter managed service and white-label for smaller ALSPs.
Why not a copy: we are neither an hours-billed ALSP nor a login-selling platform; we are a per-unit, outcome-priced, attorney-supervised AI-native service that owns defensibility.
Narrow wedge that differs: privilege review + logging for the mid-market — the highest-risk, least-commoditizable slice, not full-scope enterprise review.
Under-served segment: mid-market corporate legal departments and small-to-mid firms without in-house review teams; only ~32% of legal teams use ALSPs at all.
Unsolved pain: platforms give tools but leave defensibility, supervision, and the certification packet to the buyer; staffing gives bodies but not consistency or an AI cost structure.
Our differentiation: the defensibility packet (protocol + validation + audit trail + certification memo), attorney-owned privilege chokepoint, per-unit pricing, and a learning loop that hardens each matter's calls into reusable rules.
Vs. prior blueprints in this workspace: no eDiscovery/document-review entry exists; distinct buyer, workflow, and outcome from legal entries like PI demand packages, patent prior-art search, and lease/contract abstraction.
29. Anti-commoditization analysis
The sharpest threat is real and must be named: in 2026 Relativity (aiR) and Everlaw are bundling GenAI review into their platforms at little or no incremental charge, and commentators note this lets smaller teams "keep more matters in-house." If GenAI review is nearly free inside the tool, why pay for a managed service?
Why the service still wins:
Buyers want the outcome, not the tool. A bundled AI feature is an input; a defensible, privilege-protected, certified production is the job. Mid-market legal teams do not want to operate a review platform or own the malpractice risk of an unsupervised AI privilege call.
UPL and Rule 26(g) require a supervising attorney. A tool cannot supply supervision, sign-off, or a certification-support record. The regulatory floor keeps a human service in the loop by law.
Defensibility is a documentation/process problem, not a technology problem. Courts evaluate the protocol, validation record, and meet-and-confer — exactly what we own and productize.
Cheaper models deflate our COGS, not our price. As GenAI gets better and cheaper, our human QC minutes per GB fall and our margin expands, while our per-GB price stays well under legacy managed review. Tool commoditization is a tailwind to our unit economics.
Anti-commoditization moat: proprietary QA/validation harness, a growing library of matter-hardened privilege rules and exemplars, defensibility templates mapped to jurisdictions and protective orders, and a reputation for claw-back-free productions.
Sam Altman test — answered honestly: if a future general model makes first-pass privilege scoring near-perfect, the human minutes shrink toward the irreducible core — protocol design, edge-case adjudication, and certification. Those are the highest-value, hardest-to-automate, regulation-protected steps. The service gets cheaper and more defensible; it does not disappear, because someone accountable and licensed must still own the call.
30. Service delivery workflow
Scoping/ECA: receive collection stats; sample; estimate volume, privilege rate, timeline, and a fixed per-unit quote.
Protocol build: attorney designs the privilege protocol with the client's counsel (criteria, custodians, counsel/firm lists, 502(d) order).
Process & normalize: OCR, de-dup, thread, index.
AI first pass: score responsiveness/privilege with rationale; draft log entries; propose redactions.
Attorney review: adjudicate all privilege candidates, edge cases, and low-confidence calls; resolve rule/AI conflicts.
QA & validation: statistical sampling of AI-only calls; recall/precision/elusion; second-attorney QC on withholds.
Assemble & deliver: render redactions, format log, build defensibility packet, produce set, register claw-backs.
Follow-up: support meet-and-confer questions; handle rolling productions; renew for next wave.
31. Operations as product
The operation is the product: variance is the enemy in a review, because inconsistency between reviewers is what opponents attack. We eliminate it with: SOPs per review type; structured protocol-intake checklists; required-evidence lists (counsel/firm name lists, 502(d) order, custodian map) before work starts; automated completeness checks; exception queues; reviewer-assignment logic by expertise; confidence scoring; audit trails and version control; gold-standard exemplar sets per matter; red-team "would this survive a motion to compel?" checks; client-ready log and packet templates; and root-cause + postmortem analysis on any missed privileged doc or overturned call.
32. No-holes quality engine
Two-pass privilege: AI call + attorney sign-off; second-attorney QC on all withholds and a sample of "not privileged."
Validation to defensibility benchmarks: target recall in the 80–90%+ band with documented sampling and elusion testing.
Consistency control: matter-level exemplar set and prior-call retrieval so the 10,000th document is coded like the 10th.
Conflict detection: flag every disagreement between deterministic rules and AI for human resolution.
Claw-back safety net: 502(d) register; immediate escalation protocol for any suspected inadvertent production.
Audit trail: every call, model version, prompt, and reviewer logged for challenge-readiness.
Gold-set regression: any model/prompt change re-validated against a held-out gold set before deployment.
33. What the human expert actually does
Task
License
Min/unit at launch
Min/unit at day 90
Automation path
Quality risk
Cannot automate
Audit trail
Protocol design with counsel
Attorney
Per matter (2–5 hrs)
1.5–3 hrs (templates)
Protocol templates + prior-matter reuse
High
Legal judgment on privilege criteria
Signed protocol doc
Privilege sign-off on AI candidates
Attorney
~2–4 min/flagged doc
~0.7–1.5 min/flagged doc
Higher model precision shrinks flagged set
Critical
Final privilege call
Per-doc decision log
Edge-case adjudication (partial privilege, work product)
Attorney
~5–8 min/case
~3–5 min/case
Exemplar retrieval + suggested rationale
High
Doctrinal judgment
Rationale record
Privilege-log QC
Attorney
~1–2 min/entry
~0.5–1 min/entry
AI-drafted entries + format rules
Med
Accuracy/consistency sign-off
Log version history
Validation approval
Attorney
~30–60 min/matter
~20–40 min/matter
Auto-computed metrics + templated memo
High
Certification-support judgment
Validation report
Processing / log assembly / sampling execution
None (operator)
Ongoing
Ongoing (more automated)
Orchestration automation
Med
—
System logs
34. Minimum viable offer
"Defensible Privilege Review, done for you — per gigabyte, on your deadline." Send us your processed document set and your privilege protocol. Within your deadline we return: every privilege call made and attorney-signed, a court-ready privilege log, applied redactions, a validation record (recall/precision/sampling), and a certification-support memo for your counsel — claw-back-free, at a per-GB price below a contract-attorney room. First matter includes a fixed-fee scoping/ECA quote so you know the number before you commit.
35. Fulfillment process (first 3 customers, manually)
Stand up a hosted review workspace (RelativityOne or Everlaw) under a signed engagement + protective order.
Founder/lead review attorney designs the protocol with the client's counsel.
Run AI scoring via the platform's GenAI review (or an external LLM harness), export flagged queue.
Attorney reviews flagged privilege candidates and edge cases; second attorney QCs withholds.
Generate the privilege log from a template; apply redactions; compute validation metrics; assemble the packet.
Deliver production set + log + defensibility packet; debrief; capture every correction into the rules/exemplars.
Day-one tools are bought, not built: a review platform, an LLM harness, a log/validation generator, secure transfer, and e-signature. Proprietary orchestration and QA come after the first handful of matters.
36. Tools and systems
Review/hosting: RelativityOne or Everlaw (buy first; their bundled GenAI is an input we resell inside a service).
AI harness: model-agnostic prompt/eval layer (frontier LLMs) for scoring, rationale, log drafting, redaction proposals.
Humans are concentrated exactly where judgment and liability live: protocol design, privilege sign-off, edge-case adjudication, and validation approval. Everything upstream (processing, scoring, drafting) and downstream (metrics, formatting, assembly) is automated or operator-run. The attorney reviews an AI-prioritized queue rather than reading everything, and a second attorney QCs all withholds — so quality rises while human minutes per GB fall as models improve.
38. Nonlinear scaling and unit economics
55–65%
target gross margin (path to 65%+ as automation rises) I
>$500k
target revenue per FTE at scale I
↓ min/GB
human review minutes per GB fall as models improve I
Illustrative COGS per GB reviewed (privilege MVP)
Cost line
Launch
Day 90
Year 1
Model inference (scoring, drafting)
$0.50–2 / GB
$0.40–1.5
$0.30–1.2
Processing + hosting
$8–20 / GB
$6–15
$5–12
Attorney review minutes / GB
Largest line
−30–40%
−50–60%
QA / second-attorney QC
Moderate
Lower (sampling automated)
Lower
Support, packet assembly, PM
Moderate
Lower
Lower
Automation share: ~50–60% of documents auto-cleared at launch → ~70% by day 90 → ~80%+ by year 1 (tracking industry GenAI automation claims, applied conservatively). Throughput: one review attorney supervising an AI queue can clear multiples of a manual reviewer's daily volume. Cycle time: days, not weeks, for mid-size matters. Rework target: <2% overturned calls; quality-failure (inadvertent privileged production): ~0; escalation rate: tracked and driven down. Margin expansion comes from the shrinking human-minutes line as models improve.
Acquisition metrics (targets): CAC payback < first 1–2 matters; lead-magnet (privilege-risk scan) → consult conversion ~15–25%; waitlist/consult → pilot ~30–40%; pilot → paid recurring ~50%+; repeat-matter retention high because litigation recurs and switching mid-matter is costly.
"White-glove privilege review your firm can certify"
High per intro
Firms outsource review
Partner-sourced matters
Co-branded pilot
ALSP overflow / white-label
Bureaus need surge + cheaper first pass
"AI-native first pass under your supervision"
Volume
ALSP staffing peaks
Overflow volume
Capacity agreement
Targeted outbound to AGC/Head of Litigation & litigation-support managers
Reachable via title + recent-litigation triggers
Diagnosis memo on their privilege-review exposure
Personalized
Litigation is public record
Reply/meeting rate
Fixed-fee scoping offer
Webinars/CLE with litigators on GenAI defensibility
CLE credit draws the buyer
"Is your AI review defensible? What courts expect"
Mid
Active AI-privilege anxiety
Registrations → consults
Recorded asset + offer
40. Sales and outreach plan
Lead with a diagnosis, not a demo: a short memo estimating the prospect's privilege-review exposure and what a defensible workflow would cost per GB versus a contract-attorney room. Sell the outcome and the risk removed. Anchor on per-GB and per-log-entry pricing (their mental model), offer a fixed-fee scoping/ECA so the first number is known, and make the first matter a low-risk pilot with a defensibility guarantee (claw-back-free target + validation record). Close on the deadline urgency that triggered the search.
41. Founder-led / expert-led content plan
A credible review attorney publishes to teach the buyer how to think about defensible AI review: the anatomy of a privilege waiver; what a judge actually scrutinizes; recall/precision/elusion in plain English; how to read a vendor's validation record; the December 2025 FRCP amendments; the federal split on AI-and-privilege. Turn the best-performing organic pieces into paid-ad and CLE material. Position the founder as the person who makes GenAI review defensible.
42. First 30 days of content
10 educational posts: (1) Why review is 70%+ of your litigation bill; (2) The anatomy of a privilege waiver; (3) What a judge scrutinizes in an AI review; (4) Recall vs precision vs elusion, explained; (5) How to build a defensible privilege protocol; (6) 502(d) orders and claw-backs; (7) The Dec 2025 FRCP AI-privilege amendments; (8) The federal split on AI & privilege; (9) Contract-attorney rooms vs AI-native review — real cost math; (10) Reading a vendor's validation record.
3 diagnostic teardowns: (a) "We reviewed this (redacted) privilege log — here's what an opponent attacks"; (b) "This validation record wouldn't survive a motion to compel — here's the fix"; (c) "Per-GB cost teardown: room-of-attorneys vs AI-native."
2 lead-magnet angles: "Defensible AI Privilege Review Checklist"; "Privilege-Waiver Risk Scan (free, per matter)."
1 webinar/live review: "Is your AI document review defensible? A live protocol + validation walkthrough (CLE)."
1 outbound diagnosis template: a one-page memo estimating a named prospect's privilege-review exposure and a per-GB alternative.
43. Lead magnet and waitlist plan
Lead magnet / diagnostic: a free Privilege-Waiver Risk Scan — the prospect shares matter parameters (data volume, custodians, deadline, whether counsel domains are known) and receives a short report: estimated privilege rate, the top waiver risks in their setup, a defensible-workflow outline, and a per-GB cost estimate. This captures the exact pain signal (a live matter with a deadline) and demonstrates competence.
Waitlist CTA: "Get on the pilot list for done-for-you defensible privilege review." Conversion path: scan → scoping call → fixed-fee ECA → pilot matter → recurring per-matter engagement. What they get before paying: the scan + a scoping quote. Qualification: a live matter, a deadline, and a data population in the 10–200 GB range.
44. Warm GTM plan
Work the founder's legal network first: litigators, in-house counsel, and eDiscovery/litigation-support managers. Offer free privilege-risk scans and scoped pilots to warm contacts and to everyone who downloads the checklist or requests a scan. Prioritize anyone showing a live-matter signal (they asked about a deadline, staffing, or a specific production).
45. Targeted outbound plan
Build a list of mid-market companies with recent litigation (public dockets, PACER, news) and small-to-mid litigation firms without in-house review teams. Personalize around the actual matter: lead with a one-page diagnosis of their likely privilege-review exposure and a per-GB alternative, not a generic demo ask. Sequence: diagnosis memo → scoping-call offer → fixed-fee ECA. Add insurance-defense and employment-defense firms (high matter volume) as a repeatable segment.
46. Answer-engine / search visibility plan
Buyers increasingly research "defensible AI document review," "AI privilege review," and "privilege log service" in Google and in ChatGPT/Perplexity/Claude. Publish authoritative, well-structured explainers on defensibility, validation, and FRCP so the company is cited as the expert answer. Maintain a clear services page and FAQ schema targeting per-GB pricing, defensibility, and UPL-safe supervision — the exact questions the answer engines get asked.
47. Pilot design and early-demand-trap mitigation
First cohort: 3–5 mid-market matters (privilege-review only), capped. Early-access incentive: discounted first matter in exchange for a reference and detailed feedback. Feedback cadence: per-matter debrief plus weekly review of overturned calls. Product feedback vs custom work: anything that recurs across matters (a rule, a log format, a jurisdiction quirk) is product; one-off client-specific asks are scoped and priced separately, not absorbed. Trap mitigation: resist the urge to take full-scope enterprise reviews early; hold the privilege-review wedge until the workflow is hardened.
48. Early-access feedback flywheel
Every attorney correction is a training signal: overturned AI calls update prompts, exemplars, and deterministic rules; recurring edge cases become SOPs and QC checks; each matter's counsel/firm name list and privilege patterns enrich the retrieval layer. The result compounds — matter N+1 is coded more accurately and with fewer human minutes than matter N. What must be fixed before expanding pilots: consistent recall at benchmark, near-zero inadvertent productions, and stable per-GB COGS.
49. Build-before-scale checkpoints
After 5 matters: harden protocol intake, required-evidence lists (counsel/firm lists, 502(d) order), and QA sampling.
After 10 matters: harden SOPs, exception queues, reviewer checklists, and log/packet templates.
After 20 matters: pause new pilots until COGS/GB, rework rate, escalation rate, and cycle time are measured and stable.
Workarounds that signal non-scalability: attorneys re-reading full populations (not just flagged queues), or per-matter bespoke protocols that never converge into reusable rules.
50. 7-day launch plan
Stand up a hosted review workspace and an LLM scoring harness; draft engagement + protective-order templates.
Build the privilege-log and defensibility-packet templates and a validation-metrics script.
Write the "Defensible AI Privilege Review Checklist" lead magnet and the Privilege-Waiver Risk Scan intake.
Line up 1–2 barred review attorneys (supervision structure) and confirm conflicts/screening process.
Publish the services page + 3 founding posts; open the pilot waitlist; email 20 warm contacts.
51. 30-day launch plan
Run 10–20 privilege-risk scans; convert 3–5 into fixed-fee scoping/ECA quotes.
Land the first 1–2 pilot matters; deliver claw-back-free with a full defensibility packet.
Capture every correction into rules/exemplars; write the first postmortem.
Ship 8–10 educational posts and 1 diagnostic teardown; begin outbound to mid-market litigation and boutique firms.
Stand up SOC 2 program scoping and no-training-on-client-data commitments.
Harden protocol intake, QA sampling, and templates; hit stable per-GB margin.
Run a CLE webinar; secure 2–3 referral-partner firms; sign one ALSP overflow/white-label pilot.
Launch the full responsiveness-review and redaction expansion offers to existing pilots.
Publish a case study (redacted) demonstrating defensibility + cost delta vs contract-attorney review.
53. Metrics and KPIs
Metric
Target
Inadvertent privileged productions
~0 (claw-back-free)
Validation recall
80–90%+ with documented sampling
Overturned-call (rework) rate
<2%
Attorney minutes per GB
Down 50–60% by year 1
Automation share (auto-cleared docs)
50–60% → 80%+ by year 1
Gross margin
55–65% → 65%+
Cycle time (mid-size matter)
Days, inside deadline
Scan→consult / pilot→paid
~15–25% / ~50%+
Repeat-matter retention
High (litigation recurs)
54. Risks and mitigations (summary)
The dominant risks are platform commoditization, a single inadvertent privileged production destroying trust, UPL missteps, defensibility challenges, and data-security incidents. Each is addressed in the register below with a concrete mitigation. The through-line: keep a licensed attorney at the chokepoint, own the defensibility documentation, and price per-unit so model deflation expands margin.
55. Exhaustive risk register
1. Platforms bundle GenAI review for free and buyers self-serve — Likelihood: High · Impact: High
Mitigation: sell the supervised, certified outcome (not the tool); anchor on UPL/26(g) requiring an attorney; own defensibility documentation; use tool deflation to widen our price advantage. Target the mid-market that won't operate a platform or own privilege risk.
2. A single inadvertent privileged production — Likelihood: Low · Impact: Critical
Mitigation: two-pass privilege (AI + attorney sign-off) + second-attorney QC on all withholds; 502(d) register and claw-back protocol; conservative privilege-candidate thresholds; validation before delivery; insurance (E&O). One incident can end a brand — QC is non-negotiable.
3. Unauthorized practice of law — Likelihood: Med · Impact: High
Mitigation: licensed attorneys make/review privilege calls under a documented supervision structure; non-attorneys limited to administrative/technical tasks; final certification stays with client counsel; no fee-splitting; engagement terms codify roles.
4. Defensibility challenge (motion to compel / TAR attack) — Likelihood: Med · Impact: High
Mitigation: documented protocol, validation record (recall/precision/elusion/sampling), audit trail, and meet-and-confer support packet; align to Sedona/CAL norms and jurisdictional standing orders; expert-declaration-ready records.
Mitigation: SOC 2 Type II, encryption in transit/at rest, least-privilege access, no model training on client data, protective-order compliance, vendor DPAs, breach-response plan and cyber insurance.
6. Model hallucination / wrong privilege rationale — Likelihood: Med · Impact: High
Mitigation: humans adjudicate all flagged calls; AI provides rationale but never the final call; gold-set regression before any model/prompt change; confidence thresholds route uncertainty to attorneys.
7. Margin compression from platform/hosting costs — Likelihood: Med · Impact: Med
Mitigation: multi-platform strategy; negotiate volume hosting; migrate commodity processing to lower-cost infra; per-GB pricing floors; automation reduces the human line that dominates COGS.
8. Incumbent ALSPs go AI-native and undercut — Likelihood: Med · Impact: High
Mitigation: move faster in the mid-market they under-serve; win on per-unit pricing + defensibility packet + speed; build a matter-hardened rules/exemplar moat; pursue white-label to convert competitors into channel.
9. Long sales cycles / procurement & security review — Likelihood: Med · Impact: Med
Mitigation: fixed-fee scoping/ECA to create a fast first "yes"; pre-built security package (SOC 2, DPA, protective-order templates); pilot-sized first matter; referral partners to shortcut trust.
10. Attorney talent supply / cost — Likelihood: Med · Impact: Med
Mitigation: automation minimizes attorney minutes per GB; leverage remote barred attorneys; concentrate senior judgment on edge cases; build reusable rules so each matter needs less senior time.
11. Client counsel resists outsourcing privilege — Likelihood: Med · Impact: Med
Mitigation: position as counsel's supervised tool that supports (not replaces) their certification; give them control of the protocol and final release; start with lower-stakes matters to build trust.
12. Regulatory shift on AI in discovery (disclosure mandates, bans) — Likelihood: Low-Med · Impact: Med
Mitigation: track standing orders and FRCP amendments; make workflows disclosure-ready; keep human-in-the-loop so AI-use is defensible; adapt packet to new disclosure requirements.
13. Concentration risk (few large matters) — Likelihood: Med · Impact: Med
Mitigation: diversify across many mid-market matters and referral partners; avoid dependence on any single client or firm; standing per-matter arrangements for recurring litigants.
14. Pricing race-to-the-bottom as AI review commoditizes — Likelihood: Med · Impact: Med
Mitigation: compete on defensibility and outcome guarantee, not lowest price; bundle the packet; move up-market into higher-stakes privilege and investigations where price sensitivity is lower.
56. What could kill this
The two lethal scenarios: (1) a single inadvertent privileged production in an early flagship matter that destroys the defensibility reputation before it's established; and (2) platforms so fully commoditizing GenAI review that mid-market buyers decide "good enough" self-serve is acceptable and stop valuing a supervised, certified outcome. The first is mitigated by uncompromising two-pass QC and never delivering without validation; the second by staying anchored to the regulatory floor (a licensed attorney must own the call) and by using tool deflation to widen our price-and-margin advantage rather than being replaced by it.
57. Go/no-go reasoning
Go: proven, budgeted, mature market; review is the dominant cost line; clear mid-market whitespace; per-unit outcome pricing; regulatory moat (UPL + 26(g)); a service that gets cheaper and more defensible as models improve; and a narrow, high-value MVP wedge (privilege review) with a clean licensing boundary. No fatal disqualifier is triggered. The central threat (platform commoditization) is real but has a credible, evidence-based counter-thesis rooted in law and buyer behavior.
58. Final recommendation
Build it, wedge-first. Launch a done-for-you, attorney-supervised privilege review + logging service for mid-market corporate legal departments and small-to-mid litigation firms, priced per GB reviewed and per log entry, with a defensibility packet as the differentiator. Prove claw-back-free delivery and stable per-GB margin across the first 5–10 matters, then expand into full responsiveness review, redaction-at-scale, ECA, and investigation surge — and into ALSP white-label. Treat tool commoditization as a tailwind to unit economics, and keep a licensed attorney immovably at the privilege chokepoint.
Market-size figures vary by research firm and scope; ranges are shown rather than a single number. Vendor-reported accuracy/automation metrics are labeled Inferred. Legal/ethics/UPL and FRCP references are primary-source-oriented practice authorities.