Mapping AI across the legal value chain, from intake to appeal

A mid-size litigation firm in Chicago recently ran the numbers on a single breach-of-contract matter: 14 distinct work stages from the first client call to the appellate brief. AI tools touched eleven of those stages in ways that changed the actual cost or speed of the work. At three other stages, vendors sold "AI-powered" products that were, on inspection, the same keyword search engines from 2015 with a chatbot wrapper. Telling the difference is the single most useful skill in evaluating legal AI spend today.

This lesson walks that full matter lifecycle and flags exactly where AI creates real economic change, and where it doesn't.

Why the value chain framing matters

Law firms bill mostly by the hour (a fee model, not a topic we'll expand on here beyond noting it as context). AI's economic impact shows up as: hours removed, error rates reduced, or new services made viable at a price clients will pay. If a tool doesn't move one of those three levers, it's not changing economics, no matter how it's marketed.

We'll track the matter from intake to appeal in rough chronological order.

Stage 1-3: Intake, conflicts, and matter scoping

1. Client intake and triage. Large firms and legal aid organizations now use AI chat intake tools (e.g., systems built on models like OpenAI's GPT-4 class or Anthropic's Claude) to pre-screen client inquiries, summarize facts, and route matters. This genuinely changes economics: a paralegal reviewing a 20-minute AI-generated intake summary versus conducting the full intake call themselves. Real time saved, measurable in minutes per matter.

2. Conflicts checking. AI-enhanced conflicts databases can parse messy entity names (subsidiaries, DBAs, name variants) far better than exact-match legacy systems. This reduces false negatives that create malpractice exposure. Genuine change.

3. Matter scoping and budgeting. Some AI tools now predict matter cost and duration based on historical firm data (case type, jurisdiction, opposing counsel). Accuracy varies widely and this is an emerging category, not yet proven at scale, but the economic logic (better budgets, fewer write-offs) is real when it works.

Stage 4-7: Discovery and investigation

4. Document review and e-discovery. This is the oldest and most proven legal AI use case. Technology-assisted review (TAR, sometimes called "predictive coding") has been court-recognized since *Da Silva Moore v. Publicis Groupe* (2012, S.D.N.Y.). Modern large language modellarge language modelEin Large Language Model ist ein KI-System, das auf riesigen Textmengen trainiert wurde, um Sprache vorherzusagen und zu erzeugen. Damit sind Aufgaben wie Schreiben, Zusammenfassen und Beantworten von Fragen möglich.Vollständige Definition ansehen → (LLMLLMEin Large Language Model ist ein KI-System, das auf riesigen Textmengen trainiert wurde, um Sprache vorherzusagen und zu erzeugen. Damit sind Aufgaben wie Schreiben, Zusammenfassen und Beantworten von Fragen möglich.Vollständige Definition ansehen →) based review tools add issue-tagging and privilege-flagging on top of relevance ranking. Real, large economic change: review costs that ran $2-4 per document with contract attorneys can drop substantially with AI-assisted first-pass review, though final numbers depend heavily on document volume and complexity (treat any specific per-document rate as an estimate, not a benchmark).

5. Deposition and transcript analysis. AI tools now summarize hours of deposition transcript, flag contradictions across witnesses, and surface key admissions. Genuine time savings for litigators who previously read transcripts line by line.

6. Legal research. This is where the real-versus-relabeled distinction gets sharpest. Tools like Westlaw's CoCounsel (Thomson Reuters) and Lexis+ AI (LexisNexis) now generate research memos with citations, not just ranked search results. That's a genuine shift from "find documents" to "answer questions with sourced reasoning." But some competing tools marketed as "AI legal research" are still boolean or vector search with a summary paragraph slapped on top, no real reasoning chain. Test any tool by asking it a question with a trick: a jurisdiction split or an overruled case. Tools that just relabel search will often miss it.

7. Fact investigation and public records search. Genuinely improved by AI's ability to parse unstructured sources (news archives, court dockets, social media) at scale. Real change for background investigation work.

Stage 8-11: Drafting, negotiation, and case strategy

8. Contract drafting and review. AI contract review tools (Kira, Luminance, Harvey) flag non-standard clauses against a playbook. This is one of the clearest ROIROIReturn on Investment: das Verhältnis von Nettogewinn zu den Kosten einer Investition. Ein ROI von 300 % bedeutet, dass jeder investierte Dollar 3 Dollar zurückbringt.Vollständige Definition ansehen → (return on investmentreturn on investmentReturn on Investment: das Verhältnis von Nettogewinn zu den Kosten einer Investition. Ein ROI von 300 % bedeutet, dass jeder investierte Dollar 3 Dollar zurückbringt.Vollständige Definition ansehen →) cases in legal AI: a first-pass review of a 50-page commercial lease that took an associate two hours can be flagged in minutes, with the associate then validating rather than starting from a blank read. Genuine economic change, well documented across corporate and real estate practice groups.

9. Brief and memo drafting. LLMLLMEin Large Language Model ist ein KI-System, das auf riesigen Textmengen trainiert wurde, um Sprache vorherzusagen und zu erzeugen. Damit sind Aufgaben wie Schreiben, Zusammenfassen und Beantworten von Fragen möglich.Vollständige Definition ansehen →-assisted drafting of first drafts (not final filings) is now common. The economics change when it shifts associate time from "generate text" to "edit and verify text," provided verification discipline holds (see the caution below on hallucinationhallucinationEine Hallucination liegt vor, wenn ein KI-Modell Output erzeugt, der flüssig und selbstbewusst klingt, aber faktisch falsch, erfunden oder nicht durch die Quelldaten gedeckt ist.Vollständige Definition ansehen →).

10. Litigation strategy and outcome prediction. Tools analyzing judge- and venue-specific historical rulings (e.g., Lex Machina, Premonition) help set settlement strategy and predict motion outcomes. Genuine change to case strategy economics, especially at the settlement decision point, though predictive accuracy should always be treated as directional, not deterministic.

11. Appeal drafting and citation checking. AI citation-checking tools verify that quoted case language matches the source and that citation formatting meets court rules (Bluebook or local equivalents). After several public sanctions cases involving fabricated citations (the *Mata v. Avianca* 2023 sanctions order in the Southern District of New York is the most cited example), this stage has become a mandatory checkpoint, not optional. Genuine, high-value use because the cost of *not* doing it is a sanctions order.

The three where vendors just relabel search

Buyers should watch for "AI" branding applied to:

• Static clause libraries marketed as "AI playbooks" that are really just template repositories with no learning or matching intelligence.
• Docketing and deadline calendaring tools rebranded as "AI-powered" when the underlying logic is simple rules-based date calculation that predates any machine learning.
• Basic full-text search across document management systems (iManage, NetDocuments) relabeled "AI search" when it's the same inverted-index search from a decade ago with a chat interface bolted on.

The test: ask the vendor "what would this tool do differently if I removed the AI model?" If the honest answer is "nothing," you've found a relabeling case.

Evaluating tools: a simple framework

For any tool at any of the eleven genuine stages, ask three questions:

1. What specific hours or error rate does this remove? Get a number, even a rough one, from the vendor and from your own pilot.
2. What's the hallucination or error risk, and what's the verification cost? A tool that saves four hours of drafting but requires two hours of citation-checking has net-saved two hours, not four. Track the net, not the headline.
3. Does pricing scale with value delivered? Per-seat licensing on a tool used by three partners a month is a different ROIROIReturn on Investment: das Verhältnis von Nettogewinn zu den Kosten einer Investition. Ein ROI von 300 % bedeutet, dass jeder investierte Dollar 3 Dollar zurückbringt.Vollständige Definition ansehen → story than usage-based pricing tied to documents processed.

A simple worked example: if TAR-assisted first-pass review cuts a 100,000-document review from 500 contract-attorney-hours to 150 hours at a blended rate of $75/hour, that's a gross saving of about $26,250 on that phase alone (350 hours x $75), before subtracting the AI tool's licensing or per-document cost. Always run this calculation on your own volumes rather than trusting vendor case studies, since document complexity and review standards vary enormously across matter types.

For a rigorous, freely available primer on how courts have treated AI-assisted review standards, see the Sedona Conference's public commentary on TAR.

🎬 [VIDEO: "How AI Is Changing the Legal Profession" — youtube.com/@Reuters — a Reuters Institute-style explainer on generative AI adoption across law firms, covering both efficiency gains and professional responsibility risks]

Key Takeaways

• Eleven of fourteen typical matter stages (intake through appeal) show genuine AI-driven economic change: hours removed, errors reduced, or new price points made viable.
• Legal research and contract review show the clearest, most documented ROIROIReturn on Investment: das Verhältnis von Nettogewinn zu den Kosten einer Investition. Ein ROI von 300 % bedeutet, dass jeder investierte Dollar 3 Dollar zurückbringt.Vollständige Definition ansehen →; matter budgeting prediction and some intake tools are earlier-stage and less proven.
• Citation-checking is now a mandatory workflow step, not optional, following real sanctions cases tied to fabricated AI-generated citations.
• Watch for relabeled search: static clause libraries, rules-based docketing, and basic full-text search rebranded as "AI" with no genuine model-driven behavior underneath.
• Evaluate every tool on net hours saved after verification cost, not headline speed claims, and always test on your own document volumes before trusting vendor benchmarks.