Harvey connects legal agents with document review and shared matters

Harvey is an AI work platform for law firms and in-house legal teams. It combines agents that carry out multi-step tasks with document analysis, legal knowledge sources and shared matter workspaces. Its value depends on whether those pieces improve the route from source material to reviewed work product, while the professional team remains responsible for judgment.

01 / ProductAgents work alongside document collections and research sources

The current Agents page describes parallel tasks, plan review, agent construction and connectors. Harvey presents these as a way to delegate work while retaining approval and inspection of the result. The practical unit is therefore a task with a defined deliverable, such as an evidence table or draft briefing, rather than an unrestricted instruction to resolve a legal matter.

Vault supplies the document workspace. It supports organizing material, extracting comparable data into review tables and asking questions across the resulting information. That tabular route matters when a reviewer needs the same question answered for many documents. A narrative summary can hide a missing document or inconsistent treatment; a row-based view makes those omissions easier to investigate.

Knowledge connects research to legal databases, curated public sources and internal material. Its public page names several data partners. That does not mean every source is available under every agreement or that every jurisdiction receives identical coverage. The selected source set is part of the product configuration and should be visible when a team evaluates research results.

Spaces adds collaboration between firms, clients and internal teams around shared Vaults and work. This connects the information used to prepare a result with the people reviewing it. The organizational benefit is potentially substantial when work normally travels through separate email attachments, but it also makes access boundaries a central part of the implementation.

02 / AudienceA fit for teams with repeated evidence-heavy work

A transactional group comparing many agreements, an in-house team organizing contract information or a litigation team preparing material for review may have a useful Harvey evaluation. Each can define a bounded work product and a qualified reviewer. The common requirement is a repeatable relationship between the source material and the output, with enough consistency that corrections can improve a reusable workflow.

The product is less suitable as an unsupervised answer service for a person seeking a definitive legal conclusion. Nor does a broad agent platform necessarily improve a one-off task whose documents are still incomplete. A team should first know what belongs in the matter, which sources are authoritative for that task and what a satisfactory deliverable contains. Those decisions cannot be recovered reliably from a vague prompt alone.

For broader internal knowledge discovery, Glean is an adjacent comparison: the central need may be finding information across the business rather than creating specialized legal work product. Writer offers a broader enterprise AI and workflow approach for organizations evaluating reusable assistants across departments. These comparisons help identify whether legal-specific review and source integration are essential to the purchase.

03 / WorkflowA proposed contract inventory starts with rows and evidence

Consider a proposed evaluation by an in-house legal operations team building an inventory from a set of authorized, non-sensitive sample agreements. The initial goal is to organize factual fields and route ambiguities for professional review, not to decide enforceability. Define a small set of fields such as document identifier, named parties, stated dates and where each value appears. Keep business interpretation separate from extraction.

Prepare the document set before asking a model to summarize it. Include signed versions, amendments and known duplicates deliberately, with an expected relationship between them. A source file called “final” is not sufficient evidence that it is the controlling version. Record the inventory outside the generated output so the team can check whether every expected item was processed and identify gaps caused by unreadable scans or missing attachments.

Build a Vault review table with a clear definition for each field. Harvey’s review-table release notes document reference files for column context, persistent instructions, manual columns and cell-level refresh. A team could use those mechanisms to apply its approved extraction instructions, add a human review status and rerun a problematic cell without disturbing an entire completed table.

Inspect every consequential extracted field against the original document in the sample. A citation should point to the passage that supports the value, not merely a nearby page. Keep missing information and conflicting information distinct: “not found” is a different result from two amendments stating different dates. For the proposed evaluation, the reviewer resolves those cases and records the reason instead of allowing the system to select a convenient answer.

Only after the table is checked should the team ask for an inventory briefing. The briefing can describe the reviewed rows, open items and next actions, with an explicit boundary around what was examined. Avoid treating a fluent summary as evidence that the underlying extraction is complete. A useful acceptance test is whether another reviewer can reproduce the important entries from the cited material without rebuilding the entire analysis.

Measure preparation time, correction effort and handoff clarity separately. If document preparation dominates the task, better ingestion may matter more than faster drafting. If reviewers repeatedly rewrite the same column, the instruction or source structure may be the issue. If the final report obscures unresolved rows, revise the deliverable template. This proposed evaluation gives the team specific evidence about where the software helps.

04 / PricingThe public commercial route is a scoped sales discussion

A quoted platform price is meaningful only with a defined workload. For the contract inventory example, specify the expected document volume, concurrent reviewers, shared workspaces and source integrations. Ask which elements are included and which depend on additional agreements. This is a way to interpret an actual quote, not a claim that Harvey charges a particular fee for each item.

Compare the commercial proposal with the review effort observed in the sample. A reduction in draft preparation can be valuable even when final review remains unchanged, but those are different benefits. Do not convert vendor customer anecdotes into a guaranteed return for another organization. The useful financial model uses the team’s own task frequency, accepted outputs and measured work needed to reach them.

05 / DistinctionsStructured review creates a bridge from extraction to work product

Review tables provide a useful middle layer between reading individual files and generating a broad memo. They let a team see coverage, compare fields and preserve unresolved cases before synthesizing conclusions. Harvey’s combination of Vault, agents and reusable instructions is most compelling when that intermediate layer becomes a maintained part of the team’s practice rather than a disposable output from a chat session.

The public Vault release notes describe matter-level retention settings, exportable file logs and controls over actions available to view-only users. These are operational capabilities that can matter as much as generation. A document workflow needs a way to establish what was present, who could use it and what happens when the matter ends.

Harvey’s security page states that inputs, outputs and uploaded documents are not used to train underlying models, and describes retention controls and regional options. It also explains that ethical-wall policies remain under their existing system of record. These are vendor-described controls to verify in the proposed configuration and agreement, rather than evidence that every deployment is automatically configured correctly.

06 / QuestionsAvailability, provenance and collaboration need concrete checks

Does the account have the workflow being demonstrated?

The public site now emphasizes Harvey II and Agents, while some documentation still refers to Assistant and Workflows. Confirm the actual product experience and rollout for the account under evaluation. Several detailed help articles required authentication during this research; public release notes were readable. Consequently, this blueprint describes documented capabilities without claiming that every account already exposes an identical interface or feature set.

Can reviewers distinguish generated, extracted and human-entered information?

Keep those categories visible in the working process. A manually corrected cell should not become indistinguishable from a fresh extraction after the next run. Decide which values are approved, which need another look and which must be preserved across refreshes. This is especially important when a summary is generated from a table that contains both document facts and a reviewer’s interpretation.

What changes when another organization joins a Space?

Test collaboration with sample users before adding live matter content. A shared resource should be available to the intended reviewer and unavailable to a person outside the matter. Check the actual document, table and output, rather than only the workspace title. A successful invitation is not evidence that the entire permission model matches the team’s engagement structure.

07 / DecisionChoose a reviewable work product as the starting point

Harvey merits evaluation when a legal team wants specialized AI work organized around documents, knowledge and collaboration. Start with a recurring task that can be checked in detail. Expand only after the team understands the source coverage, review effort and account capabilities that produced the result.