The Legal AI Vendor Audit Rights Report 2026: What Law Firms and Legal Departments Are — and Are Not — Contractually Entitled to Examine in Their AI Vendor Relationships — and How Often They're Exercising That Right

Research briefing — a review of public sources. The Legal Stack did not conduct a survey, benchmark or interviews for this piece.

What can and cannot be documented

This report asked two questions: what audit rights legal buyers hold over their AI vendors, and how often they use them. The second question cannot be answered from public sources. We found no published study of how often law firms or legal departments exercise AI vendor audit rights. Vendor contracts are mostly confidential, so there is also no public dataset on how common particular clauses are. An earlier version of this article described patterns in vendor terms and a sample "precision" gap as findings. It also named specific vendors' DPA provisions and specific banks' outside counsel guidelines, and we could not verify those. We have removed all of it.

What is documented is the legal and standards baseline that sets what a buyer can demand. That baseline includes GDPR processor rules, the EU AI Act's transparency and deployer duties, US ethics guidance on vetting AI vendors, and the assurance reports vendors typically offer. This briefing covers that baseline and then gives our analysis of how to close the gaps.

The statutory floor: GDPR Article 28

Where an AI vendor processes personal data on a legal department's behalf, GDPR Article 28 already requires audit-type rights in the processing contract.

Under Article 28(3)(h), the processor must make available "all information necessary to demonstrate compliance" and must "allow for and contribute to audits, including inspections, conducted by the controller or another auditor mandated by the controller."
Under Article 28(2), the processor may not engage a sub-processor without the controller's prior specific or general written authorisation.
Under Article 28(4), the same data protection obligations must flow down to sub-processors.

Analysis: these rights concern data handling. They do not give the buyer a right to examine model accuracy, training data provenance or output quality. That distinction is the core of this topic. A buyer can hold a fully compliant Article 28 DPA and still have no contractual window into how the model performs on its own matters.

The EU AI Act: transparency and deployer duties, on a later timetable

The original article treated Article 13 of the EU AI Act as already reshaping vendor terms. The timeline matters here.

Article 13 requires providers of high-risk AI systems to supply instructions for use that let deployers interpret and use the system appropriately. Article 26 then imposes duties on deployers. They must:

use the system in line with those instructions
assign human oversight to competent people
monitor operation and inform the provider and authorities of risks or serious incidents
keep automatically generated logs under their control "for a period … of at least six months"

There are two important limits.

Timing. Following the Digital Omnibus on AI, in force since 27 July 2026, stand-alone Annex III high-risk obligations apply from 2 December 2027, not now.
Scope. Annex III point 8 covers AI used by or for judicial authorities (and alternative dispute resolution). In our reading, lawyer-facing legal AI generally falls outside it, so most contract review and research tools are unlikely to be "high-risk".

Analysis: the AI Act is a useful reference point for negotiation rather than a present obligation for most legal AI. A buyer can reasonably ask a vendor for Article 13-style documentation of capabilities, limitations and logging, and for log access that would let the buyer meet Article 26-style monitoring duties. The buyer should not assume the law compels the vendor to agree. We could not verify claims that particular vendors have already adopted AI Act-specific terms, and we do not repeat them.

US ethics guidance: vetting vendors is part of competence

ABA Formal Opinion 512 (29 July 2024) applies the duties of competence, confidentiality, communication, candour, supervision and reasonable fees to generative AI. As summarised by Katten in the National Law Review, the opinion expects lawyers to take several steps:

read and understand a tool's terms of use, privacy policy and related contract terms, or consult people who can explain them
gain a "reasonable understanding" of the tool's benefits and risks
obtain client informed consent before inputting information relating to the representation into a "self-learning" tool whose output could disclose it

State guidance points the same way. Examples include the NYC Bar's Formal Opinion 2024-5 (August 2024), Florida Bar Ethics Opinion 24-1 (January 2024) and North Carolina's 2024 Formal Ethics Opinion 1.

Analysis: audit rights are one practical way to meet these duties. If a firm cannot find out whether its vendor trains on its inputs, which sub-processors see the data, or how errors are logged, it will struggle to show the "reasonable understanding" the opinion expects.

Client pressure is real but poorly documented. Bloomberg Law reported in November 2023 that Paul Weiss had clients who asked the firm not to use generative AI without first communicating with them. The same reporting found firms unsure how much to disclose. We found no public, verifiable outside counsel guidelines that require firms to hold specific audit rights over their AI tools.

The assurance gap: what SOC 2 and ISO 42001 do and don't cover

For many buyers, the practical "audit right" is receiving a SOC 2 Type II report. A Cloud Security Alliance research note (30 August 2026) argues this falls short for AI. It makes the following points:

The AICPA's Trust Services Criteria were last substantively revised in 2017, with a 2022 refresh of the points of focus.
The AICPA has published no AI-specific criteria.
A SOC 2 Type II "does not reliably attest to model-specific properties." Training data provenance, evaluation methodology, bias testing and explainability are tested only if the auditor chooses to test them.
Two vendors can therefore each hold clean reports covering different AI controls.

CSA promotes its own AI Controls Matrix and STAR for AI as an interim baseline, which buyers should bear in mind.

ISO/IEC 42001, published in December 2023, is the first international standard for an AI management system. Organisations can be certified against it by accredited bodies. As CASRAI notes, it attests to governance processes. It does not attest that a specific model is accurate or unbiased, and it does not by itself establish compliance with the AI Act.

Analysis: neither report replaces contract rights over model behaviour. Both are useful inputs to a review, and buyers should ask what scope and controls each one actually covered.

Analysis: a framework for legal ops teams

The following recommendations are The Legal Stack's analysis.

Inventory by right type. For each AI vendor, record separately:
data-handling rights (GDPR Article 28 audits, sub-processor approval and flow-down)
assurance artefacts (SOC 2 scope, ISO 42001 certification)
model-behaviour rights (performance reporting on your data, training-data use, incident and error logs)
Close the model-behaviour gap explicitly. Ask for:
a written no-training-on-customer-data commitment
notice of material model changes
access to error and incident logs relating to your matters
a right to run your own evaluation sets
Use the AI Act as a reference point, not a mandate. Request Article 13-style documentation and Article 26-style log retention (at least six months) whatever your jurisdiction, while recognising that most legal tools are not in scope.
Read the assurance reports properly. Confirm which trust services criteria and which AI-related controls the SOC 2 auditor actually tested.
Write an exercise protocol. Assign an owner for each right, set triggers (a scheduled annual review, incidents, client requests, model changes) and set an escalation path. A right nobody owns will not be exercised.
Line up with client and ethics expectations. Map ABA 512 and the relevant state guidance to your vendor diligence file, so you can show clients what you checked.

Sources

GDPR-Info, "Art. 28 GDPR – Processor" (Regulation (EU) 2016/679): https://gdpr-info.eu/art-28-gdpr/
Future of Life Institute, EU AI Act Explorer, "Article 26: Obligations of Deployers of High-Risk AI Systems": https://artificialintelligenceact.eu/article/26/
National Law Review (Katten), "ABA Weighs in on Generative AI Use in Legal Practice" (14 Aug 2024): https://natlawreview.com/article/aba-weighs-generative-ai-use-legal-practice
Bloomberg Law, "Law Firms Wrestle With How Much to Tell Clients About AI Use" (29 Nov 2023): https://news.bloomberglaw.com/us-law-week/law-firms-wrestle-with-how-much-to-tell-clients-about-ai-use
Cloud Security Alliance, "The SOC 2 AI Gap: Auditors Improvise Where AICPA Has Not Acted" (30 Aug 2026): https://labs.cloudsecurityalliance.org/research/csa-research-note-soc2-ai-controls-gap-20260830-csa-styled/
CASRAI, "ISO/IEC 42001: The First AI Management-System Standard": https://casrai.org/wp/?p=1468
American Bar Association, "Formal Opinion 512: Generative Artificial Intelligence Tools" (29 Jul 2024): https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf
Ezel, summary of New York City Bar Formal Opinion 2024-5, "Generative AI in the Practice of Law" (7 Aug 2024): https://ezel.ai/ethics-opinions/nyc-bar/2024-5-generative-ai-in-the-practice-of-law
The Florida Bar, "Ethics Opinion 24-1" (19 Jan 2024): https://www.floridabar.org/etopinions/opinion-24-1/
North Carolina State Bar, "2024 Formal Ethics Opinion 1" (2024), copy hosted by UNC School of Government: https://www.sog.unc.edu/sites/default/files/course_materials/2024%20FEO%201%20%28State%20Bar%20re%20AI%29_0.pdf
National Law Review (K&L Gates), "EU Digital Omnibus on AI Enters Into Force" (31 Jul 2026): https://natlawreview.com/article/eu-digital-omnibus-ai-enters-force
Future of Life Institute, EU AI Act Explorer, "Annex III: High-Risk AI Systems Referred to in Article 6(2)" (accessed Oct 2026): https://artificialintelligenceact.eu/annex/3/