EFROS Research / Public Index
US AI Vendor Governance Index
An operator-grade, source-cited scorecard for 20 AI vendors evaluated on 12 US AI governance axes — NIST AI RMF, Colorado AI Act, HIPAA BAA, SR 11-7, ABA Formal Op 512, Section 1557, SOC 2, ISO/IEC 42001, US data residency, training-data opt-out, subprocessor transparency, and trust-center maturity.
Public • Source-Cited • Updated Quarterly
What this is
Every regulated buyer faces the same question — is this AI vendor safe for our use case? The trust-portal claim, the vendor sales deck, and the auditor's checklist rarely line up. The Index closes that gap by translating public vendor documentation into a normalized score against the US frameworks operators are actually measured against during audit: NIST AI RMF, the Colorado AI Act, the Microsoft / HHS / FRB sectoral overlays, and the ABA opinion that defines responsible use of AI in legal practice.
The Index is free, public, and ungated. Every cell is source-cited to a vendor trust portal, public BAA, SOC report cover page, or published methodology document. Scoring is "yes" / "partial" / "no" / "na" — N/A axes (sector overlays that don't apply to the vendor's deployment category) are excluded from the composite denominator so vendors aren't penalized for axes outside their scope. Composite scoring is sector-weighted: a healthcare-vertical vendor carries 2× weight on Section 1557, a legal-vertical vendor 2× on ABA Op 512, a banking-vertical vendor 2× on SR 11-7.
How we score
- 12 axes per vendor scored "yes" / "partial" / "no" / "na" with source citation per cell.
- Sector-weighted composite — healthcare 2× Section 1557, legal 2× ABA Op 512, banking 2× SR 11-7, plus 1.5× BAA for any regulated sector.
- Trust-center maturity scored 1-5 separately and contributes 10% of the final composite on top of the 12-axis average.
The full scorecard
20 vendors, 12 axes, ranked
Click any vendor name for the source-cited deep dive (per-cell notes, strengths, weaknesses, best/avoid use cases). Status badges are color-coded: green = yes, amber = partial, rose = no, zinc = not applicable.
| # | Vendor | Category | Sector | Score | Grade | BAA | Opt-out | US Res | SOC 2 | ISO 42001 | NIST AI | CO AI | §1557 | SR 11-7 | ABA 512 | Subproc |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Abridge | Healthcare | healthcare | 87 | A | Yes | Yes | Yes | Yes | Partial | Partial | Partial | Yes | N/A | N/A | Yes |
| 2 | Thomson Reuters CoCounsel | Legal | legal | 80 | B | Yes | Yes | Yes | Yes | No | Partial | Partial | N/A | N/A | Yes | Yes |
| 3 | FICO Falcon Fraud Manager + FICO Score AI | Banking | banking | 80 | B | Yes | Yes | Yes | Yes | No | Partial | Partial | N/A | Yes | N/A | Yes |
| 4 | Lexis+ AI | Legal | legal | 76 | B | Yes | Yes | Yes | Yes | No | Partial | No | N/A | N/A | Yes | Yes |
| 5 | Westlaw Precision AI | Legal | legal | 76 | B | Yes | Yes | Yes | Yes | No | Partial | No | N/A | N/A | Yes | Yes |
| 6 | Microsoft 365 Copilot | Productivity | general | 75 | B | Yes | Yes | Yes | Yes | Partial | Partial | Partial | Partial | Partial | Partial | Yes |
| 7 | Harvey | Legal | legal | 74 | B | Yes | Yes | Yes | Yes | No | Partial | Partial | N/A | N/A | Yes | Partial |
| 8 | Zest AI | Banking | banking | 74 | B | Yes | Yes | Yes | Yes | No | Partial | Partial | N/A | Yes | N/A | Partial |
| 9 | Upstart | Banking | banking | 74 | B | Yes | Yes | Yes | Yes | No | Partial | Partial | N/A | Yes | N/A | Partial |
| 10 | Suki AI | Healthcare | healthcare | 72 | B | Yes | Yes | Yes | Yes | No | Partial | Partial | Partial | N/A | N/A | Yes |
| 11 | Nuance DAX Copilot (Microsoft) | Healthcare | healthcare | 70 | B | Yes | Yes | Yes | Yes | No | Partial | No | Partial | N/A | N/A | Yes |
| 12 | Salesforce Einstein / Agentforce | Productivity | general | 69 | C | Yes | Yes | Yes | Yes | No | Partial | No | Partial | Partial | N/A | Yes |
| 13 | Glean | Productivity | general | 69 | C | Yes | Yes | Yes | Yes | No | Partial | No | N/A | N/A | N/A | Yes |
| 14 | Arctic Wolf | Security MSSP | general | 69 | C | Yes | Yes | Yes | Yes | No | Partial | No | N/A | N/A | N/A | Yes |
| 15 | Huntress | Security MSSP | general | 69 | C | Yes | Yes | Yes | Yes | No | Partial | No | N/A | N/A | N/A | Yes |
| 16 | eSentire | Security MSSP | general | 69 | C | Yes | Yes | Yes | Yes | No | Partial | No | N/A | N/A | N/A | Yes |
| 17 | Sophos | Security MSSP | general | 69 | C | Yes | Yes | Yes | Yes | No | Partial | No | N/A | N/A | N/A | Yes |
| 18 | Unit21 | Banking | banking | 68 | C | Yes | Yes | Yes | Yes | No | Partial | No | N/A | Partial | N/A | Yes |
| 19 | Ironclad AI | Legal | legal | 63 | C | Yes | Yes | Yes | Yes | No | No | No | N/A | N/A | Partial | Yes |
| 20 | Anthropic Claude | Foundation | general | 58 | C | Partial | Yes | Partial | Yes | No | Partial | No | N/A | N/A | N/A | Yes |
| 21 | Google Gemini for Workspace | Foundation | general | 58 | C | Partial | Partial | Yes | Yes | No | Partial | No | N/A | N/A | N/A | Yes |
| 22 | Hummingbird | Banking | banking | 56 | C | Yes | Yes | Yes | Yes | No | No | No | N/A | Partial | N/A | Partial |
| 23 | OpenAI ChatGPT & API | Foundation | general | 53 | D | Partial | Partial | Partial | Yes | No | Partial | No | N/A | N/A | N/A | Yes |
| 24 | ConnectWise | Security MSSP | general | 50 | D | Partial | Yes | Partial | Yes | No | No | No | N/A | N/A | N/A | Yes |
| 25 | Spellbook | Legal | legal | 45 | D | Yes | Yes | Partial | Partial | No | No | No | N/A | N/A | Partial | Partial |
| 26 | Heidi Health | Healthcare | healthcare | 45 | D | Yes | Yes | Partial | Partial | No | No | No | Partial | N/A | N/A | Partial |
| 27 | Notion AI | Productivity | general | 33 | F | No | Partial | No | Yes | No | No | No | N/A | N/A | N/A | Yes |
| 28 | Meta Llama | Foundation | general | 25 | F | No | Yes | Yes | No | No | No | No | N/A | N/A | N/A | No |
| 29 | Otter.ai | Productivity | general | 25 | F | No | Partial | No | Yes | No | No | No | N/A | N/A | N/A | Partial |
| 30 | Perplexity AI | Foundation | general | 19 | F | No | Partial | No | Partial | No | No | No | N/A | N/A | N/A | Partial |
Axis legend: BAA = HIPAA Business Associate Agreement; Opt-out = training-data opt-out default; US Res = US data residency configuration option; SOC 2 = SOC 2 Type II report; ISO 42001 = ISO/IEC 42001 attestation; NIST AI = NIST AI RMF self-attestation; CO AI = Colorado AI Act readiness; §1557 = HHS-OCR Section 1557 readiness; SR 11-7 = FRB SR 11-7 readiness; ABA 512 = ABA Formal Opinion 512 readiness; Subproc = subprocessor list public.
By category
Category leaders
Top vendors within each category, ranked by sector-weighted composite. Sector overlays (Section 1557, SR 11-7, ABA Op 512) are primary differentiators in healthcare and legal categories; the foundation and productivity categories compete on platform fundamentals (BAA, residency, SOC stack, trust-center maturity).
Foundation models
General-purpose LLM providers. Sector overlays (Section 1557, SR 11-7, ABA Op 512) are deployer responsibility.
Anthropic Claude
CClaude foundation model family delivered via claude.ai (Free/Pro/Team/Enterprise) and a developer API. Differentiated on Constitutional AI training and safety research orientation.
Google Gemini for Workspace
CGemini foundation models delivered through Google Workspace integration (Docs, Gmail, Drive) and the Vertex AI developer platform. Highest pull for Workspace-standardized organizations.
OpenAI ChatGPT & API
DGPT-class foundation models delivered via ChatGPT consumer/enterprise tiers and a developer API. The most-deployed generative AI vendor in US enterprise.
Meta Llama
FOpen-weight foundation model family (Llama 3.x, Llama 4) distributed under a community license. Used primarily as a self-hosted or partner-hosted alternative to API-only vendors.
Perplexity AI
FAnswer engine combining proprietary retrieval with multiple foundation models (GPT, Claude, Sonar). Differentiated on citation-grounded responses over raw chat.
Productivity / general enterprise
AI overlays on the productivity stack (M365, Workspace, Notion, CRM). Governance posture inherits from the platform.
Microsoft 365 Copilot
BGenerative AI overlay on the Microsoft 365 stack — Outlook, Word, Excel, PowerPoint, Teams. Available exclusively to commercial M365 tenants.
Salesforce Einstein / Agentforce
CAI and agent infrastructure built into Salesforce CRM. The Einstein Trust Layer enforces no-train, masking, and audit logging at the platform level.
Glean
CEnterprise generative search and AI agent platform that indexes the SaaS stack (Drive, SharePoint, Slack, Confluence, Salesforce, etc.) and returns permission-aware AI answers.
Notion AI
FAI overlay on Notion's collaborative workspace. Used for summarization, drafting, semantic search, and database automation within Notion content.
Otter.ai
FReal-time meeting transcription and summarization. Common deployment in sales/CS, sometimes leaking into clinical or legal meeting workflows where governance gaps matter.
Legal vertical
Purpose-built for law firms — privilege handling, matter walls, ABA Formal Opinion 512 alignment.
Thomson Reuters CoCounsel
BLegal AI assistant from Thomson Reuters (the parent of Westlaw and Practical Law). Acquired Casetext in 2023. Tightly integrated with Westlaw and Practical Law content.
Lexis+ AI
BLexisNexis's legal AI assistant integrated with the Lexis content corpus. Differentiated on citation grounding from the Lexis case-law and secondary-source database.
Westlaw Precision AI
BWestlaw's AI-assisted research layer — natural-language query, AI-generated case summaries, and AI memo drafting grounded in Westlaw's primary-source database.
Harvey
BGenerative AI platform purpose-built for law firms. Backed by OpenAI; primarily deployed at Am Law 100/200 firms for drafting, research, and matter-aware workflows.
Ironclad AI
CContract lifecycle management platform with AI features for contract drafting, review, and metadata extraction. Targets in-house legal teams.
Healthcare vertical
Clinical AI documentation and decision support. HIPAA BAA, Section 1557 algorithmic non-discrimination, HITRUST.
Abridge
AAmbient clinical AI documentation. Differentiated on clinician-experience design, citation-grounded notes, and deep EHR integration (notably Epic).
Suki AI
BClinical AI voice assistant for ambient note generation, dictation, and EHR navigation. EHR-integrated (Epic, Athenahealth, Cerner, Meditech, NextGen).
Nuance DAX Copilot (Microsoft)
BAmbient clinical AI scribe — captures clinician-patient encounters and generates structured clinical notes. EHR-integrated (Epic, Cerner, athenahealth, others).
Heidi Health
DClinical AI documentation assistant — Australia-headquartered with US market expansion. Used heavily in solo and small-practice deployments due to lower price point.
Banking vertical
Credit decisioning, fraud detection, BSA/AML, transaction monitoring. SR 11-7 model risk management is the binding constraint.
FICO Falcon Fraud Manager + FICO Score AI
BDecades-deep machine-learning portfolio across fraud detection (Falcon) and credit decisioning (FICO Score 10 T). The reference SR 11-7 documentation in the industry; most US banks already operate against FICO's validation patterns.
Zest AI
BAI-driven credit underwriting platform with strong fair-lending documentation. Differentiated on explicit ECOA/Reg B + adverse-action explainability output, designed for examiner-facing defensibility.
Upstart
BAI lending platform with CFPB no-action letter history. Operates as a partner for community banks and credit unions that want AI-driven origination without building it internally. CFPB scrutiny + fair-lending audit history is unusually deep.
Unit21
CModern transaction-monitoring + fraud detection platform. Deployed at fintech-adjacent banks, neobanks, payments processors, and crypto-aligned institutions where legacy AML vendors don't fit.
Hummingbird
CModern compliance operations platform — BSA/AML case management, investigations, SAR filing, transaction monitoring overlay. Used by community banks, credit unions, and crypto-adjacent institutions for examiner-ready AML workflow.
Security MSSP / MSP
MSSPs and MSPs with AI-augmented detection, response, and IT operations. Cross-cutting scoring — sector overlays N/A because MSSPs serve all sectors.
Arctic Wolf
CConcierge MDR with named-team accountability and AI-augmented threat detection across endpoint, cloud, network, and identity. AI features primarily as detection acceleration rather than autonomous decisioning.
Huntress
CEndpoint and M365 identity threat detection with AI-augmented threat hunting, sized for SMB-to-mid-market organizations without enterprise MDR budget. Decision-support AI rather than autonomous response.
eSentire
CEnterprise MDR with proprietary threat hunting depth and the most explicit AI-platform branding (Atlas AI) in the MDR category. Threat hunt depth is the differentiator over breadth-first competitors.
Sophos
CVendor-integrated endpoint AI with the longest-running deep-learning malware detection lineage in the category (Invincea acquisition, 2017). Sophos MDR overlays managed detection on top of the platform.
ConnectWise
DRMM + PSA platform with AI features for ticket automation, asset insights, and IT workflow acceleration. MSP-centric — sold to managed service providers who deliver downstream services to end customers.
Operator signal
Trust-center maturity (1-5)
Trust-center maturity captures how much of a vendor's governance posture is self-serve verifiable versus locked behind an NDA or a sales conversation. A score of 5 means a public certificate library, audit reports under NDA, granular subprocessor and residency documentation, and AI-specific governance pages. A score of 1 or 2 means a thin security page and direct-request documentation only.
- Abridge5/5
- Microsoft 365 Copilot5/5
- Nuance DAX Copilot (Microsoft)5/5
- Salesforce Einstein / Agentforce5/5
- Thomson Reuters CoCounsel4/5
- FICO Falcon Fraud Manager + FICO Score AI4/5
- Lexis+ AI4/5
- Westlaw Precision AI4/5
- Suki AI4/5
- Glean4/5
- Arctic Wolf4/5
- Huntress4/5
- eSentire4/5
- Sophos4/5
- Unit214/5
- Ironclad AI4/5
- Anthropic Claude4/5
- Google Gemini for Workspace4/5
- OpenAI ChatGPT & API4/5
- Harvey3/5
- Zest AI3/5
- Upstart3/5
- Hummingbird3/5
- ConnectWise3/5
- Notion AI3/5
- Spellbook2/5
- Heidi Health2/5
- Meta Llama2/5
- Otter.ai2/5
- Perplexity AI2/5
Microsoft (M365 Copilot, DAX Copilot) and Salesforce share the highest trust-center maturity scores — public certificate libraries, granular subprocessor lists, and audit-report distribution at scale. Abridge is the only clinical-AI vendor at parity. The bottom of the distribution is dominated by smaller vendors and foundation models in earlier governance stages.
What the data shows
Key findings
- 1. ISO/IEC 42001 is the biggest market-wide gap.No vendor in the Index holds a completed ISO/IEC 42001 AI management system attestation as of May 2026. Microsoft and Abridge are the closest — both publicly state alignment work in progress. For buyers asking "what's the AI-specific certification floor," there isn't one yet — and that's true even for the vendors with the most mature broader compliance stacks.
- 2. Sector overlays sort the leaderboard.Healthcare-vertical vendors (Abridge, DAX Copilot, Suki) lead on Section 1557 readiness; legal-vertical vendors (Harvey, CoCounsel, Lexis+ AI, Westlaw Precision AI) lead on ABA Op 512. Foundation models score lower on the composite because sector overlays are deployer responsibility — Section 1557, SR 11-7, and ABA Op 512 are all "N/A" at the vendor layer for general-purpose LLMs.
- 3. BAA gaps are the #1 disqualifier for shadow-AI workloads.Notion AI, Otter.ai, and Perplexity are widely deployed in regulated organizations but none of them sign BAAs. The audit finding we see most often involves PHI or attorney-client content flowing into one of these three tools without the deploying organization realizing the BAA gap. DLP at the upload boundary is the right preventive control.
- 4. Trust-center maturity correlates strongly with composite.Vendors at trust-center score 5 (Microsoft M365 Copilot, DAX Copilot, Salesforce, Abridge) cluster near the top of the composite. Vendors at score 2 (Meta Llama, Perplexity, Otter.ai, Spellbook, Heidi Health) cluster near the bottom. The 10% contribution to the composite formula matters less than the underlying signal — vendors that invest in self-serve trust documentation also tend to invest in the underlying controls.
- 5. Colorado AI Act posture is structurally vendor-side weak.Even at the top of the Index, Colorado AI Act readiness is "partial" at best — most vendors document the deployer responsibility model but do not claim full developer-side compliance. SB 24-205 places most of the burden on the deployer regardless, so this is structurally honest, but it means buyers cannot offload Colorado AI Act exposure to a vendor selection. The deployer obligations (impact assessments, consumer notice, right to appeal) remain in-house regardless of which vendor is chosen.
Methodology + disclaimer
Scoring as of 2026-05-13 from public information — vendor trust portals, published BAAs and DPAs, SOC report cover pages, model cards, and vendor governance documentation. Each cell is source-cited; methodology is auditable. Posture changes frequently — re-verify with the vendor's trust center before contract. Updated quarterly. The Index is an EFROS research artifact, not legal or compliance advice. Nothing in this Index constitutes an endorsement of any vendor; vendors with strong composite scores may still be poor fits for a specific deployment context.
Full methodology and source list →Three ways forward
Use the Index as a starting point. Reserve a fixed-fee AI Governance Audit if a specific deployment needs operator-grade verification, run the self-serve AI Risk Score to benchmark your own posture in five minutes, or read the sector-specific playbooks for your regulatory context.
Sector-specific AI governance
By sector
By US state (state-specific framework relevance)
Get the next edition first
Subscribe to EFROS Research Updates. Quarterly editions, mid-quarter changelog when material vendor governance changes ship. No marketing drip — research only. Unsubscribe with one click.
We never sell or share emails. Your address is logged in our research list only — see our privacy policy.
Save, cite, or integrate this edition