Buyer resource · version 1.0 · July 2026

Legal AI evaluation
scorecard.

Forty criteria for deciding whether a legal AI platform is accurate, governable and useful on your actual work. Score evidence—not demos, feature counts or vendor promises.

Apply it in a live pilot →
Scoring: rate each criterion 0–4. Multiply the category average by its weight. Require documentary or observed evidence for any score above 2.

Before you score

01

Use one live workflow

A representative, bounded matter exposes workflow gaps that a curated demo hides.

02

Freeze the inputs

Give every shortlisted vendor the same documents, authorities and instructions.

03

Use the same reviewer

Consistency matters more than the reviewer’s enthusiasm for a particular product.

04

Measure correction

Time saved disappears when lawyers must rebuild fast but unreliable output.

01 · WEIGHT 15%

Matter and workflow fit

____ / 4
Evaluation criterion0–4Evidence / finding
The platform completes the chosen workflow end to end, not only isolated prompts.____
The workflow accepts the file types, sources and matter size used in production.____
Firm playbooks, precedents and escalation rules can be applied consistently.____
Lawyers can review, correct and approve work without leaving the normal work surface.____
The vendor distinguishes generally available features from configuration, services and roadmap.____
02 · WEIGHT 20%

Legal accuracy and verification

____ / 4
Evaluation criterion0–4Evidence / finding
Material propositions link to the exact supporting source passage.____
The system surfaces uncertainty, missing authority and conflicting sources.____
A reviewer can trace how an answer was produced without relying on hidden reasoning.____
Adverse treatment, amendments and jurisdictional hierarchy are tested explicitly.____
Correction time is measured alongside first-output speed.____
03 · WEIGHT 12%

Source coverage and currency

____ / 4
Evaluation criterion0–4Evidence / finding
The vendor names the primary sources relevant to your jurisdictions and practice.____
Update cadence, last-ingested date and failed-source handling are observable.____
Uploaded matter documents remain distinguishable from public legal sources.____
Source licenses permit the proposed use, storage and downstream work product.____
Exported citations remain usable outside the platform.____
04 · WEIGHT 15%

Security and confidentiality

____ / 4
Evaluation criterion0–4Evidence / finding
Encryption, tenant isolation, access controls and audit logging are demonstrated.____
Ethical walls and matter permissions match the firm’s operating model.____
Model providers, subprocessors and zero-retention commitments are documented.____
Incident response, vulnerability handling and breach notification terms are acceptable.____
Required certifications are complete today or explicitly treated as a gap.____
05 · WEIGHT 12%

Data governance and deployment

____ / 4
Evaluation criterion0–4Evidence / finding
Storage and inference locations are documented for every relevant data category.____
Customer data is not used to train shared models without explicit authorization.____
Retention, deletion, backup deletion and legal-hold behavior can be tested.____
The customer can export matter data, work product, prompts and playbooks in usable formats.____
The DPA and deployment design align with professional duties and applicable law.____
06 · WEIGHT 8%

Adoption and controls

____ / 4
Evaluation criterion0–4Evidence / finding
The product fits Word, email, DMS and other daily systems used by the target team.____
Training addresses verification and professional judgment, not only button clicks.____
Usage policy, approval boundaries and prohibited workflows can be enforced.____
Admins can monitor adoption and risky behavior without reading privileged content unnecessarily.____
The pilot includes representative skeptics, not only innovation champions.____
07 · WEIGHT 8%

Implementation and service

____ / 4
Evaluation criterion0–4Evidence / finding
Named owners, dependencies and realistic dates exist for migration and rollout.____
Integration scope and professional-services effort are stated before signature.____
Support coverage, escalation paths and service levels match matter urgency.____
The vendor can explain failure recovery, continuity and customer exit procedures.____
Success criteria have accountable owners on both sides.____
08 · WEIGHT 10%

Economics and evidence

____ / 4
Evaluation criterion0–4Evidence / finding
Baseline lawyer time and current process cost are measured before the pilot.____
Value includes review and correction effort, not only generation time.____
Pricing, minimums, implementation, model usage and renewal mechanics are understood.____
The business case separates demonstrated pilot evidence from projected benefits.____
A go, conditional-go or no-go decision can be defended to management and clients.____

Decision record

Weighted score____ / 100
Critical unresolved risks________________
DecisionGO / CONDITIONAL / NO-GO
Decision owner and date________________

Suggested gates

  • No critical confidentiality, privilege or data-governance gap remains unresolved.
  • Every material citation in the evaluation set is traceable and reviewed.
  • The vendor meets the agreed workflow threshold, not only an aggregate score.
  • Implementation effort, recurring cost and exit terms are understood.
APPLY THE SCORECARD

Evaluate Gotham on one live workflow.

Two weeks, agreed criteria, source-linked work and a written go/no-go review.

Request a structured pilot →

Published by Gotham, a legal AI vendor. Adapt the scorecard to your professional duties, client commitments, security requirements and jurisdiction. It is an evaluation aid, not legal or security advice.