Harvey

AI Centrality

How much of the product is actually AI. Whether the machine learning is the mechanism the buyer is paying for or a feature layered onto conventional software, and whether the vendor is specific about which is which.

The models are the product. Assistant, Vault, Knowledge and Workflow Agents are all generative systems, and there is no underlying document or workflow system that would stand without them. Vendor material describes every module in model terms.

Citation Accuracy and Hallucination Disclosure

Whether the vendor publishes measured accuracy on citations and assertions, grounds output to primary sources, and says plainly what its system does when it does not know. Legal has a documented public record of fabricated citations reaching filed briefs, so an untested claim of accuracy is not evidence.

CORRECTED 29 Aug 2026, second correction to this row. Previously graded B on two stated grounds, both of which are now resolved. Ground one was that BigLaw Bench is the vendor's own framework rather than independent evidence. That remains true of BigLaw Bench, but independent evidence also exists and was missed: Harvey Assistant participated in the February 2025 Vals Legal AI Report, a third party benchmark against a measured lawyer baseline, and was evaluated across six tasks scoring between 65.0 and 94.8 percent, surpassing the lawyer baseline on five of the six, with 94.8 percent on document question answering at 24.7 points above baseline and 77.8 percent on scanned and messily formatted court transcripts at 24.1 points above baseline, all at sub minute response times. Those figures sit on the evaluator's own site and are checkable without reference to any vendor claim. Ground two was that the citator and refusal limbs failed. That was a double count and is withdrawn: both are separately measured by their own signal rows on this record, and applying them again to the capability grade penalised the same absence twice. It was also inconsistent, since three legal research vendors on this index hold an A on this axis with both of those signals recorded as not addressed. The vendor's own disclosure is unchanged and remains substantial: BigLaw Bench with task categories and grading rubrics on a public repository, measured hallucination rates and source scores by model, a hallucination defined as a factual claim disprovable against a source of truth with reasoning errors tracked separately, and output linking to the specific document passages supporting each assertion. Two limits recorded rather than deducted for: the February 2025 study measured task accuracy rather than citation validity or hallucination rate specifically, and this vendor did not participate in the later Vals study that measured citation authoritativeness. The full BigLaw Bench dataset also sits behind a direct request rather than open publication.

Autonomy and Oversight Model

What the system decides on its own, what a lawyer must approve, and whether the vendor documents where the review point sits. A tool that drafts under review and a tool that files without one are different products and different risks.

States in published material that the product is designed to assist lawyers rather than replace them and that it is built to make verification easy. Review surfaces are real and documented: inline links to source passages, role based permissions and conditionals in the workflow builder, and admin level workspace governance. Not located as of 29 Aug 2026: the threshold at which an agent stops and hands back to a lawyer, or what the vendor commits to when an agent is wrong.

Operational and Outcome Evidence

Named, dated evidence that the product works in production at real firms or legal departments. Case studies with figures and identified customers count. Unattributed testimonials and launch announcements do not.

Named customers appear in vendor material with attributed quotes, including Blank Rome on the iManage integration and a published Burges Salmon selection story. Vendor states 700 plus customers across 58 plus countries. Not located as of 29 Aug 2026: dated outcome figures with a stated method a reader could assess, which is what separates this from an A.

Privilege and Confidentiality Posture

How client confidences are handled: attorney client privilege and work product treatment, segregation of one client matter from another, whether client data trains any model, and what the vendor commits to in writing rather than in marketing.

Substantive published commitments: no training on customer data by default, a contractual prohibition on model providers training, zero data retention enforced on model providers, logical workspace separation, role based access, ethical wall sync with the firm's own walls provider, and processing in the EU, Switzerland or Australia. Two gaps keep this off an A. The security page defines customer data as uploaded documents and customer content as queries and responses as separate contractual terms, so the no training commitment reads plainly on one and not on both. Attorney client privilege and work product handling is not addressed directly in located public material as of 29 Aug 2026.

UPL and Professional Responsibility Posture

Whether the vendor is clear that it supplies a tool rather than legal advice, who its audience is, and how it addresses unauthorized practice of law, competence and supervision duties, and jurisdiction limits. ABA Formal Opinion 512 is the reference point.

One sentence in a security blog post states the product is designed to assist lawyers rather than replace them. Checked the vendor site, security page, security addendum and help center on 29 Aug 2026 and did not locate a published position on the advice line, on competence and supervision duties, or on jurisdiction limits. The intended audience is unambiguously lawyers and legal departments, which is why this sits at C rather than lower.

AI Governance and Bias Disclosure

Published governance over model behaviour: who owns it inside the vendor, what is tested before release, and what is disclosed about disparate output across matter types, parties, or populations.

CORRECTED 29 Aug 2026, third correction to this record. Previously graded C on the finding that no AI specific governance regime was located and that the published testing was security testing rather than model behaviour testing. That was wrong, and it came from reading the vendor's marketing surfaces rather than opening its trust centre, where the governance material actually sits. What is published on the trust centre, publicly and without a request: ISO/IEC 42001:2023 certification, the international standard for AI management systems, accompanied by a published Statement of Applicability, which is the document identifying which controls apply and why and is therefore a published scope rather than a bare badge. Alongside it, AIUC-1 certification, an AI specific assurance standard, conducted by Schellman, which the vendor states is the first accredited AIUC-1 certification body, and which the vendor describes as validating adversarial testing and its AI security programme specifically. EU AI Act conformity is separately listed. The trust centre carries a dedicated AI section with AI Governance, AI Monitoring and AI Overview items, and an AI Acceptable Use Policy sits in the published policy set. ISO 27701 for privacy information management and an IRAP attestation are also held. Two independent AI specific certifications, one of them adversarially tested, with published statements of applicability and a named accredited certifier, is the strongest AI governance position on this index, ahead of the four other A grades on this axis, each of which rests on ISO 42001 alone or on a single certification plus a framework document. One gap remains and is recorded rather than waived: no disclosure was located about uneven output across matter types, parties or populations, so bias specifically is still unaddressed, and no named individual owner of model governance was located.

AI Safety and Data Stewardship

Retention, deletion, access control, and what happens to prompts and documents after they are processed. Whether the vendor states its subprocessors and its incident practice, or leaves the buyer to assume.

Publishes retention under customer control with documented vault retention triggers and deletion timelines, role based access control, logical workspace separation, encryption in transit and at rest, and a current named subprocessor list with an update FAQ. Not located as of 29 Aug 2026: a published incident and breach notification practice, which is the remaining element of the A bar.

AI Liability and Recourse

What the vendor stands behind contractually when its output is wrong. Indemnities, caps, carve outs, insurance, and whether any of it is published or only reachable through a negotiated agreement.

Checked the vendor site, its published legal pages including the security addendum and the subprocessor FAQ, and the trust center on 29 Aug 2026. No published indemnity scope, liability cap, carve out, warranty or insurance position located. Commercial terms appear to be reached through a negotiated enterprise agreement rather than published.

Practice Systems Integration Depth

How deeply the product reaches into the systems legal work already lives in: document management such as iManage and NetDocuments, Word and Outlook, contract lifecycle management, matter management, e-billing, and court filing systems.

Documented native integrations with iManage, NetDocuments, SharePoint and OneDrive, Google Drive, Box, Microsoft Word, Outlook, EDGAR and PitchBook, plus an MCP connector library. Help center articles describe what each integration moves, in which direction, what an admin must configure, and what a given integration does not support. The iManage connection is a direct OAuth integration with an embedded web extension rather than third party middleware.

Deployment Model and Data Residency

Where the software runs and where the data sits. Multi tenant cloud, single tenant, private deployment, on premises, and whether region of residence is a published option or an enterprise conversation.

States processing in the EU and Switzerland or Australia for customers with data localization requirements, and states that this applies to subprocessors as well. Tenancy is multi tenant with logical workspace separation and enforced role based access. Not located as of 29 Aug 2026: where data is stored as distinct from where it is processed, and what changes between tiers.

Security Certifications and Trust Center

Independent attestation a buyer can pull without a sales call: SOC 2, ISO 27001, penetration test summaries, a trust center with current reports and named scope rather than a badge image.

SOC 2 Type II attestation and ISO 27001 certification with the auditor named as Schellman, renewed annually, and the 2026 cycle announced publicly. Penetration testing and red teaming partners are named as NCC Group and Bishop Fox. Certified under the EU US Data Privacy Framework. A live trust portal at trust.harvey.ai carries the current reports. Reports sit behind a portal request rather than an open download, which is a request flow rather than a sales call.

Model Supply Chain Disclosure

Which models sit underneath, whose they are, where they run, and whether the vendor commits to telling customers when that changes. A legal buyer inherits every dependency it cannot see.

Publishes a subprocessor list naming model and infrastructure providers including OpenAI, Anthropic, Google Cloud, AWS and Microsoft, alongside a subprocessor update FAQ and a security diagram showing model access through Bedrock and Vertex AI. States zero data retention and ephemeral processing at the model providers. Not located as of 29 Aug 2026: a published commitment to notify customers before the model supply chain changes, as distinct from an FAQ describing a change that has already happened.

Commercial Transparency

Whether a buyer can learn what this costs without entering a sales process: published rates, the unit being charged, what sits behind an enterprise tier, and what implementation adds.

Checked the vendor home page, the platform and product pages, the help center and the trust center on 29 Aug 2026. No pricing page, no published rate, no stated unit of charge and no published tier structure located. Access to pricing runs through a demo request, which is sales gated and earns no credit. Third party per seat estimates exist in trade coverage but are not vendor published and do not move this axis.

Firm and Practice Coverage

Who the product is actually built for. AmLaw, midlaw, small firm and solo, in house departments, government and courts, and which practice areas are supported rather than merely claimed.

Describes its segments with substance: large law firms first, expanding into corporate legal departments and professional services, with 700 plus customers across 58 plus countries and practice coverage spanning litigation, transactional diligence, regulatory and tax. Not located as of 29 Aug 2026: a statement of the boundaries, meaning which firm sizes or practice areas the product is not built for, which is what the A bar asks for.