Harvey

Harvey combines an Assistant, document-oriented Vaults, Workflow Agents, and reusable workspace resources for legal work. Its current support material says Assistant can work with uploaded documents or selected knowledge sources, link answers to cited sources, and produce editable drafts; Vault is described for storing and analyzing large document sets. [S1] Its public platform page describes agents, pre-built or buyer-built workflows, and review-ready cited results. [S2]

The useful fit is a legal team that can name the matter owner, select authoritative sources, define a narrow task, control access to sensitive documents, and have a lawyer check every material output before it is used. It is not evidence that a citation is complete or current, that output is legal advice, or that a firm can delegate legal judgment, client communication, conflict controls, or final approval to the product.

The core trade-off is that Harvey can make analysis and first drafts faster while making source selection, permissions, context quality, legal review, and exception handling central to the operating model. Public materials must be confirmed against the subscription, jurisdiction, data, integrations, settings, and contract a buyer actually intends to use.

Best for: legal departments and firms with repeatable, bounded work; defined source and document controls; lawyer reviewers; and a measured rollout plan.

Not for: a team seeking autonomous legal advice or sign-off; a matter without a clear owner and review path; or a buyer that cannot validate jurisdictional coverage, citations, privilege, access controls, and data terms in its own environment.

Research a defined legal question with selected sources

• Trigger: A lawyer or approved business user needs a preliminary answer, issue map, or cited research path for a bounded question.
• Inputs: Matter facts, jurisdiction, date and scope limits, approved knowledge sources, prompt, source-selection rule, and review standard.
• Output: Harvey's support material describes Assistant answers that link back to selected sources and drafts that can be revised, cited, and exported. [S1]
• Human checkpoint: The responsible lawyer checks the cited materials, authority, currentness, jurisdictional fit, factual assumptions, omitted issues, and conclusion before relying on or sharing the work.
• Success measure: Citation-verification rate, material correction rate, unanswered or escalated questions, research cycle time, source-coverage gaps, and reviewer time.

Review a bounded document set

• Trigger: A team needs to identify defined terms, clauses, deviations, or diligence issues across an approved document collection.
• Inputs: Authorized documents, a defined review question or table, approved taxonomy, relevant playbook, access policy, source links, and quality sample.
• Output: Harvey describes Vault as a place to upload, store, and analyze thousands of documents; it also describes review tables and workflows for structured analysis. [S1][S2]
• Human checkpoint: Counsel samples documents and cited passages, resolves ambiguous or missing evidence, approves the issue list, and decides what is communicated or escalated.
• Success measure: Recall and precision against a lawyer-reviewed sample, unsupported findings, missed critical issues, document throughput, rework, exception rate, and reviewer effort.

Prepare a lawyer-reviewed draft or comparison

• Trigger: A legal team needs a first draft, summary, clause comparison, or structured explanation from approved matter inputs.
• Inputs: Verified facts, instructions, precedent or playbook, jurisdiction, approved documents and sources, formatting rules, and a named owner.
• Output: Harvey's support documentation says Assistant can create and refine drafts, edit text, summarize documents, and generate results that can be revised, cited, and exported. [S1]
• Human checkpoint: The matter owner verifies facts, legal reasoning, quoted language, citations, client instructions, privilege, style, and final advice before release.
• Success measure: Draft-to-approval time, material legal and factual corrections, citation errors, reviewer changes, turnaround, and client or stakeholder rework.

Standardize a recurring legal workflow

• Trigger: A team identifies a repeatable task with an agreed scope, input format, output, review point, and exception route.
• Inputs: Workflow design, approved instructions, selected Vaults or knowledge sources, embedded files where applicable, builder and admin roles, test cases, and a publication rule.
• Output: Harvey's platform says teams can run pre-built workflows or build their own. Its release notes record workflow-builder permissions, admin oversight, and workflows that can use Vault, regional knowledge, local, and DMS files. [S2][S5]
• Human checkpoint: An admin controls availability; a lawyer or designated owner reviews outputs, monitors exceptions, changes the workflow deliberately, and disables it if it exceeds the approved boundary.
• Success measure: Completion rate, exception and override rate, correction rate, time saved after review, adoption by approved users, and percentage of results within the agreed scope.

1. The buyer chooses one legal task and writes the scope: jurisdiction, permitted data, source set, expected output, responsible lawyer, quality standard, confidentiality rule, exception path, and stop condition.
2. An authorized user works in Assistant, a Vault, or a configured workflow. Harvey says Assistant can use uploaded documents or chosen knowledge sources; Vault supports document analysis; its platform describes pre-built and custom workflows. [S1][S2]
3. The user or workflow produces a cited answer, draft, table, or other reviewable work product. The citation is a navigation aid, not proof that the analysis is complete or correct.
4. The accountable lawyer checks source passages, assumptions, jurisdiction, client instructions, conflicts, omitted issues, and the final work product. Missing, contradictory, sensitive, or out-of-scope work enters a manual exception route.
5. The team records quality and operational results, restricts or changes the workflow only through its defined control path, and expands only after the pilot proves both value and safe recovery in the actual subscription.

Verified facts from current primary sources

• Harvey's support article, last updated September 15, 2026, describes Assistant, Vault, Workflow Agents, History, and Library. It says Assistant can use uploaded documents or selected knowledge sources, links answers to sources used, and creates editable drafts; it describes Vault for uploading, storing, and analyzing document collections. [S1]
• Harvey's platform page describes Agents as delivering a cited, review-ready result; Vault for bulk legal-document analysis; and Workflows that are pre-built or tailored by the buyer. [S2]
• The platform page publicly lists Word, Outlook, iManage, LexisNexis, NetDocuments, SharePoint, Google Drive, Aderant, Ironclad, APIs, and regional knowledge sources. It does not establish a buyer's entitlement, connector scope, or configuration. [S2]
• The Privacy Center, last updated July 13, 2026, defines Customer Data and Content and says Harvey does not use them to train AI models or improve its products. It also says AI-provider subprocessors are contractually prohibited from using them to train, develop, or improve models or services. [S3]
• Harvey's security page describes SAML SSO, audit logs, IP allow-listing, lifecycle controls, encryption at rest and in transit, role-based access, and regional processing options. The page directs buyers to the security portal and legal agreements for details. [S4]
• Harvey's Agents release notes record central admin oversight over builder publication and current workflow capabilities. [S5]

B2Bagents assessment

Harvey looks most useful when it shortens the distance from approved legal materials to a reviewable draft, table, or research path. The public evidence supports a platform with cited assistance, document collections, configurable workflows, and a range of named integrations. The buyer still owns source selection, access control, factual and legal validation, final advice, and the boundary between an assistive workflow and an unsafe automation.

• No public price, licensing unit, term, implementation cost, support commitment, trial terms, cancellation terms, or data-export commitment was established in the reviewed sources. Request a written proposal for the intended subscription. [S2]
• The reviewed materials do not prove output accuracy, complete authority coverage, currentness of each source, correct jurisdictional interpretation, legal advice suitability, or performance for a buyer's matters. Test those questions against a lawyer-reviewed sample.
• The public integration list does not establish which connectors, knowledge sources, regions, models, features, guest access, or workflow capabilities are available in the buyer's edition or how they are configured. [S2][S5]
• Harvey's public security and privacy statements are a useful diligence starting point, not a substitute for buyer-specific reports, contracts, roles, retention, deletion, data flows, support access, incident obligations, or subprocessors. [S3][S4]
• The sources do not establish how the buyer's playbooks, prompts, workflows, publications, exceptions, ethical walls, exports, audit records, or emergency disablement behave in practice. Verify them in a controlled tenant test.

Good-fit conditions: a legal team has a bounded recurring task, reliable approved materials, named reviewers, a source-verification practice, permissions that match matter access, and a measured pilot boundary.

Poor-fit conditions: a task demands unsupervised legal judgment; the relevant jurisdiction or source base is unclear; documents cannot be safely scoped; reviewer capacity is absent; or the team lacks a controlled route for exceptions, corrections, and access changes.

Predictable failure modes and controls:

• A response cites material that is incomplete, stale, misapplied, or missing a controlling authority. Define sources and jurisdiction, inspect primary citations, sample outputs against a lawyer-reviewed baseline, and escalate uncertainty.
• A draft carries incorrect facts, assumptions, or legal conclusions into a client or internal deliverable. Keep a named lawyer accountable for fact checking, legal analysis, client instructions, and final approval.
• A workflow handles a document, user, matter, or action outside scope. Use least-privilege, matter and resource controls, approved input criteria, tests, exception queues, logging, and an immediate manual fallback.
• A shortcut becomes a hidden decision system. Measure corrections, overrides, unsupported citations, missed issues, and review time—not only raw throughput or prompt volume.
• An integration or retention configuration expands access unexpectedly. Review the actual connector, data flow, roles, lifecycle settings, exports, termination, and recovery process before use with live matters.

Begin with one task such as an internal issue map from a fixed document set, a research outline against an approved source set, or a first draft that always receives lawyer review. Assign a matter owner, legal reviewer, workspace administrator, knowledge-source owner, security and privacy reviewer, integration owner, incident contact, and executive sponsor.

Harvey's public materials describe resource-level controls in its Spaces product and, on the security page, SAML SSO, audit logs, IP allow-listing, data lifecycle controls, role-based access, and regional processing options. [S2][S4] Those descriptions do not prove the buyer's policy or tenant setup. Demonstrate current roles, workspace and matter separation, ethical walls where applicable, guest access, source scopes, workflow publication, logging, deletion, export, credential revocation, and disablement against representative failure cases.

Harvey's Privacy Center says it does not train on Customer Data or Content and says its AI-provider subprocessors are contractually prohibited from using that material to train or improve models. [S3] Obtain and review the current agreement, DPA, subprocessor list, Security Addendum, applicable reports, data flows, regional options, retention and deletion terms, support-access controls, and incident commitments for the selected deployment.

No public Harvey price or pricing model was verified in this research. The platform page offers a demo request, but that does not establish product entitlement or commercial terms. [S2]

Request a proposal that identifies the product modules, users or usage metric, permitted knowledge sources, integrations, implementation and migration work, training, support, pilot scope, region, billing, overages, term, renewal, cancellation, audit and export access, retention and deletion, and transition assistance. Confirm whether any workflow, agent, Vault, integration, source, security, or collaboration capability required by the pilot is included.

1. Scope: Pick one low-risk, repeatable internal task, a fixed document set or approved source set, one jurisdictional boundary, and one accountable lawyer. Keep advice, filings, signature authority, client communications, negotiations, conflict decisions, and broad matter access outside the first scope.
2. Baseline: Record current research or review time, reviewer time, turnaround, material corrections, missed issues, citation-check results, rework, escalations, source gaps, and stakeholder satisfaction.
3. Test set: Include ordinary, incomplete, contradictory, stale, ambiguous-jurisdiction, irrelevant-source, missing-citation, sensitive-document, denied-access, wrong-matter, workflow-change, and manual-fallback cases.
4. Pass threshold: Agree in advance on a lawyer-reviewed quality standard, citation-verification rate, material-error and missed-issue limits, review time, turnaround, exception rate, access-control behavior, and improvement over the baseline after checking and correction time.
5. Stop conditions: A result is used as legal advice without review; a material source or fact cannot be verified; a workflow crosses its scope; permissions or ethical walls are uncertain; an error cannot be traced or corrected; or the pilot does not improve the full reviewed process.

• Spellbook: Compare contract-review and drafting scope, Word workflow, clause guidance, source and playbook controls, and reviewer responsibility.
• Luminance: Compare document-analysis and legal-review workflow, matter scope, source evidence, permissions, and the human decision process.
• Legora: Compare legal-work assistance, document and knowledge workflow, collaboration, integrations, governance, and procurement evidence.
• Crosby: Compare a buyer-operated legal-work platform with an AI-native legal service that includes attorney review and document-based commercial scope.

1. Which knowledge sources, jurisdictions, models, connectors, Vaults, features, regions, and workflow capabilities are available in our exact subscription, and how will we verify the source behind each material statement? [S1][S2]
2. Can a lawyer trace a result back to the underlying authority, identify omissions and source dates, correct it, and preserve a review record before it reaches a client or business user? [S1]
3. Who can upload, view, query, share, export, delete, connect, build, publish, edit, disable, or override workspaces, Vaults, sources, playbooks, workflow agents, and integrations? Show the roles and audit evidence in our tenant. [S4][S5]
4. How are matter boundaries, ethical walls, guest accounts, client collaboration, privilege, retention, deletion, logs, exports, emergency access, and offboarding handled in our legal and security model? [S2][S4]
5. Which customer data, content, metadata, and outputs are sent to Harvey or subprocessors; which agreement terms apply; and how do enabled features affect retention or access? [S3][S4]
6. What does our quote include, and what happens if we need to export work, change providers, reduce scope, or end the agreement? [S2]