The pre-deployment checklist partners actually sign off on

A managing partner at a mid-size litigation firm was asked in late 2025 to approve a generative AI tool for first-pass deposition summaries. She refused to sign off until the vendor produced something no marketing deck had: a packet showing exactly what data trained the model, how it performed against summaries her own associates had already written, and what happened when it got something wrong. That packet, not the product demo, is what actually gets AI tools into production at law firms.

This lesson builds that packet, section by section.

Why law firms need an associate-style clearance process

Firms don't let a first-year associate file a brief unsupervised. They test drafting on low-stakes matters, check citations, and build trust incrementally before granting autonomy.

AI deployment should follow the same logic. The American Bar Association's Formal Opinion 512 (2024) on generative AI makes this explicit: lawyers using AI tools retain full competence and supervision duties under Model Rule 1.1 (competence) and Model Rule 5.3 (supervision of non-lawyer assistance), regardless of whether the "assistant" is a person or a model. A checklist is how a firm documents that it met that duty.

Section 1: Data provenanceData provenanceData Lineage zeigt, wie Daten sich durch Systeme bewegen und dabei transformiert werden, von der Quelle bis zur Nutzung: woher sie kommen, was sie verändert hat und wohin sie gehen.Vollständige Definition ansehen →

Before anything else, the sign-off packet must answer: what data trained or fine-tuned this model, and who owns it.

Key questions to document:

• Training data source. Is the base model trained on public web data (like most general-purpose LLMs, large language models), licensed legal databases (Westlaw, LexisNexis content), or the firm's own matter files?
• Confidentiality boundary. Does firm data entered into the tool get used for further training, or is it walled off? This matters directly under ABA Model Rule 1.6 (confidentiality of client information). A tool that silently retains prompts for model improvement is a malpractice risk, not a convenience.
• Jurisdictional exposure. If the vendor processes data in the EU, the EU AI Act (in force since August 2024, with phased obligations through 2027) classifies certain legal-assistance uses as limited or, in specific contexts, higher risk, triggering transparency obligations. In the US, no single federal AI statute exists yet; state laws (e.g., Colorado's AI Act, effective 2026) and existing tort/privacy law fill the gap.
• Data retention and deletion. Get contractual language, not a sales assurance, on retention periods and deletion rights.

A one-page provenance memo, signed by the vendor and reviewed by the firm's general counsel, is the deliverable here. If a vendor cannot produce this, that itself is the answer.

Section 2: Output testing against known-answer benchmarks

This is the associate's "trial brief" equivalent. Before granting any autonomy, test the model against questions with a verified correct answer.

Build a benchmark set from real, closed matters:

1. Pull 30 to 50 past matters where the correct output is already known (a contract clause that was in fact enforceable, a case citation that was in fact still good law, a deposition summary an experienced associate already wrote and partners approved).
2. Run the AI tool on the same inputs.
3. Score outputs on three axes: factual accuracy, citation validity, and omission of material facts.
4. Set a numeric pass threshold before you see results (for example, no more than 2% of citations may be fabricated or "hallucinated," a known failure mode where a model generates plausible but nonexistent case law).

Why this matters concretely: in *Mata v. Avianca* (S.D.N.Y. 2023), attorneys submitted a brief with six fabricated case citations generated by ChatGPT, resulting in sanctions. Multiple similar sanctions cases followed through 2024 and 2025 across US federal and state courts. A benchmark test run in-house, before deployment, catches this failure mode on your own matters rather than in front of a judge.

A simple worked example of how to score a batch:

Benchmark set: 40 closed matters, known correct citations = 40
AI-generated citations flagged as fabricated or non-existent = 3
Hallucination rate = 3 / 40 = 7.5%

Firm threshold for autonomous use = 2%
Result: FAIL. Tool requires mandatory human citation-check step before any use.

That single calculation is what turns "the vendor says it's 95% accurate" into a defensible, firm-specific number a partner can cite in a malpractice defense if something later goes wrong.

Section 3: Fallback protocols

Every checklist needs a plan for when the model is wrong, not just a hope that it won't be.

Minimum fallback elements:

• Human-in-the-loop gate. Define which outputs require a licensed attorney's sign-off before client delivery (nearly everything, at least initially). No AI output leaves the building without a named reviewer.
• Escalation trigger. A defined confidence or complexity threshold above which the tool must flag "send to senior associate" rather than answer. Most enterprise legal AI tools (e.g., Harvey, CoCounsel) now support this natively as of 2025 and 2026 releases.
• Kill switch. A documented process to pull the tool from active matters within hours, not days, if a systemic error is found (comparable to a firm freezing a junior associate's filings after a serious error is discovered).
• Audit trail. Every AI-generated draft is logged with timestamp, model version, and reviewing attorney, mirroring billing and conflict-check recordkeeping firms already maintain.

For a general framework on structuring this kind of governance, the NIST AI Risk Management Framework (US National Institute of Standards and Technology, 2023, free) is a solid non-legal-specific reference many firm risk committees now cite directly in their internal policies.

Assembling the sign-off packet

Put these three sections into one document a managing partner or general counsel can actually read in one sitting:

1. Provenance memo (1 page): data sources, confidentiality treatment, jurisdiction, retention terms.
2. Benchmark report (1 to 2 pages): sample size, accuracy and 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 → rates, pass/fail against pre-set thresholds.
3. Fallback protocol (1 page): human review gates, escalation triggers, kill switch owner, audit logging method.

This is deliberately short. A 40-page vendor security questionnaire gets filed and ignored. A 3 to 4 page packet with numbers gets read and signed.

🎬 [VIDEO: "How Law Firms Are Actually Using AI in 2025" — youtube.com — search for recent Thomson Reuters or Stanford CodeX panel discussions covering real adoption data and governance practices at large firms]

Key Takeaways

• Treat AI deployment like onboarding an unsupervised associate: require provenance documentation, a graded test on known-answer benchmarks, and a fallback plan before granting autonomy.
• Data provenanceData provenanceData Lineage zeigt, wie Daten sich durch Systeme bewegen und dabei transformiert werden, von der Quelle bis zur Nutzung: woher sie kommen, was sie verändert hat und wohin sie gehen.Vollständige Definition ansehen → must answer who trained the model, whether firm data is used for retraining, and what jurisdictional rules (EU AI Act, state AI laws, ABA Formal Opinion 512) apply.
• Run in-house benchmark tests on closed matters with a pre-set pass threshold (e.g., 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 → rate under 2%); do not rely solely on vendor-reported accuracy claims.
• Build explicit fallback protocols: human review gates, escalation triggers, a kill switch, and an audit trail, mirroring existing supervision duties under Model Rules 1.1, 1.6, and 5.3.
• A short, numbers-based sign-off packet (3 to 4 pages) gets reviewed and signed; a long generic vendor questionnaire typically does not.