AI performance is task-specific. A measured gain in one workflow is evidence to test the next workflow—not permission to generalize across legal work.
Flintwood Research
Evidence before
opinion.
Research for legal leaders deciding where technology belongs, what the evidence supports, and how capability becomes accountable practice.
Publication standard
Readable enough to use. Rigorous enough to trust.
Flintwood Research distinguishes controlled findings from inference, names limitations, links to primary sources and DOIs, and translates evidence into operating decisions without pretending that synthesis is original experimentation.
What the evidence says
Numbers with their boundaries left intact.
These figures come from different studies and from Flintwood’s jurisdiction review. They are not a blended performance claim; each number retains its task, population, system, and date.
51jurisdictions mapped50 states and the District of Columbia
198authority entries54 AI and 144 client-funds sources, reviewed August 2026
≈30%shorter task timeOne controlled highlighting task; 206 law students
58–88%hallucination rangeOpen-domain federal-case questions using 2023-era public models
Key insights
Review must be designed. “Human in the loop” becomes a control only when the reviewer, source, acceptance standard, escalation path, and record are explicit.
Authority changes by jurisdiction. Fee labels alone do not determine ownership, trust-account treatment, earning, or refund obligations.
Published
Evidence Review · Report 001
Generative AI and Legal Work
An evidence review of productivity, quality, accuracy, and professional judgment—now available as a complete HTML report and a designed PDF edition.
Read the complete reportInteractive field guide · 50 states + DC
Legal AI, Client Funds, and Economics
Public AI authorities, advance-fee and trust-account rules, federal context, and government legal-market measures—mapped jurisdiction by jurisdiction.
Explore the mapField notes
Why legal AI pilots stall after the demonstration
Access is not adoption. A practical account of the operating decisions that turn a promising pilot into repeatable capability.
Read the analysisLatest Field Notes
The blog is the applied layer.
AdoptionWhy legal AI pilots stall after the demonstrationSix operating decisions that turn access into repeatable capability.
Governance“Human in the loop” is not a control until the loop is designedA practical model for review, evidence, escalation, and records.
EconomicsAI efficiency creates a pricing decision for law firmsMeasured time savings do not settle the strategic or ethical pricing choice.
Research agenda
Forward-descriptive work, grounded in the real questions legal teams face.
Planned publications are developed from verified source sets. Titles describe the decision the work will help a reader make.
How Legal Practices Have Adopted New Technology—and What AI Programs Can Learn
A history of legal research databases, e-discovery, cloud systems, and other adoption shifts, focused on the institutional conditions that moved tools into routine practice.
Shadow AI in Law Firms and Legal Departments: Why Unapproved Use Spreads and How to Govern It
An evidence-led account of demand, access, organizational incentives, data risk, and why governed pathways can outperform categorical prohibition.
How AI Changes the Billable Hour: Efficiency, Ethics, Pricing, and Value-Based Legal Work
A careful separation of measured efficiency from normative pricing decisions, professional duties, client value, and law-firm operating economics.
A Legal AI Operating Model for Law Firms and Legal Departments
A framework connecting authority, workflow, tooling, sources, human review, accountability, knowledge, and measurable outcomes.
How to Design Human Review for AI-Assisted Legal Work
A practical control model defining the output, reviewer, evidence, acceptance standard, escalation path, and retained record.
The research should lead somewhere useful.
Each evidence review creates a public HTML resource, a designed PDF, practical blog analysis, and one operational artifact. The path is deliberate: evidence → interpretation → working method → relevant service.