Harvey AI

Harvey is AI for legal work, and the distinction that matters is not the model but everything wrapped around it. Any assistant can summarise a contract; almost none can do it inside a system a magic circle firm's risk committee has approved.

The product covers agents that carry out legal tasks, a Vault for documents, a Knowledge layer, Spaces for collaborative matters, and Contract Intelligence for the review work that dominates transactional practice.

This review covers what the platform does, why its compliance posture is the actual selling point, who is using it, and the limits that no amount of certification removes.

Harvey AI: What You Need To Know

• What it is: an AI platform built specifically for legal and professional services work, not a general assistant with a legal prompt.
• Why firms buy it: the compliance stack. SOC 2 Type II, ISO 27001, ISO 27701, and ISO 42001, with SAML SSO, audit logs, and IP allow-listing.
• Core pieces: Agents, Vault for documents, Knowledge, Spaces, Contract Intelligence, Command Center, and a mobile app.
• Adoption is real: Harvey cites 2,400+ organisations including A&O Shearman, Latham & Watkins, Norton Rose Fulbright, Bayer, and Deutsche Telekom.
• Not for individuals. There is no self-serve trial path advertised — this is an institutional purchase with an implementation behind it.

Is Harvey The Right Legal AI For You?

A Good Match When

• You are a firm or in-house team with a formal security review process
• Document review and due diligence consume significant hours
• Firm precedent exists but nobody can reliably find it
• ISO 42001 or comparable AI governance is being asked for
• There is budget and appetite for an implementation, not just a subscription

The Wrong Tool When

• You are a solo practitioner or a very small practice
• You want to try something today without a procurement process
• Your billing model cannot absorb faster document review
• Your jurisdiction is outside where the platform is strongest
• You expect output that does not need lawyer verification

Harvey AI Platform Profile

Confidentiality Is The Product

Legal work has a constraint most software categories do not: the material is privileged. A tool that leaks a draft, retains a document it should not, or cannot prove who accessed what is not a bad tool — it is a professional liability.

Harvey's certification list is unusually long for an AI vendor, and the composition is telling. ISO 42001 is the AI management standard specifically, which is still uncommon and directly answers the question a risk committee now asks.

Alongside it sit SOC 2 Type II, ISO 27001, ISO 27701 for privacy, GDPR and CCPA, with SAML SSO, audit logs, IP allow-listing, and data lifecycle management.

For most software, security is a checkbox on the way to a purchase. In legal AI, it is the purchase.

The Pieces And What They Do

Harvey is assembled from components that map onto how legal work is actually organised — by matter, by document set, by precedent.

• Agents — carry out multi-step legal tasks rather than answering one question
• Vault — the document repository the agents reason over
• Knowledge — the firm's own precedent and institutional material
• Spaces — collaborative areas, closest to how a matter is run
• Contract Intelligence — review and analysis across contract sets
• Command Center — oversight of what is running and who is using it
• Memory — context carried across sessions rather than re-established
• Harvey Mobile — access away from the desk

Where It Earns Its Keep

The work Harvey targets is high-volume, high-cost, and structurally repetitive — exactly the combination that makes legal services expensive.

Document Review At Scale

Due diligence and contract review are the clearest cases. Reading two thousand contracts for a specific clause is work that consumes junior time and produces no insight in the reading itself.

Contract Intelligence is aimed squarely at that, and it is where firms typically see a measurable change first.

Institutional Knowledge

The Knowledge layer addresses a problem every firm has: precedent exists but nobody can find it. Work gets redone because locating the previous version costs more than rewriting.

An AI that can reason across a firm's own history is doing something no general assistant can, because the material is not public.

In-House Legal Teams

For corporate legal departments the calculation differs — the goal is handling more internally rather than billing less. Named corporate users include Deutsche Telekom, Bayer, KKR, and Procter & Gamble.

That group is often the faster adopter, because there is no billable-hour tension to resolve first.

How A Firm Should Approach It

Legal AI rollouts fail on process far more often than on technology, and the sequence below reflects where firms typically get stuck.

1. Get risk and compliance involved before the demo, not after — they will decide this regardless.
2. Pick one practice area and one task type, ideally contract review, where the output is checkable against a known answer.
3. Run Harvey alongside the existing process on real matters and compare, rather than testing on synthetic examples.
4. Establish the verification rule explicitly: who checks the output, and what happens before anything reaches a client.
5. Load the Knowledge layer with genuine precedent — thin institutional context produces generic legal answers.
6. Only then widen to other practice areas, and revisit the billing model, because the economics change.

The Adoption Signal

Harvey states more than 2,400 organisations, and the named list is unusually specific for a vendor page: A&O Shearman, Latham & Watkins, Norton Rose Fulbright, Reed Smith, Ashurst, Cuatrecasas, and O'Melveny among the firms.

That matters as evidence rather than as marketing. Firms of that size do not deploy tools into privileged work without a security review, and each named deployment is effectively a passed audit.

It is still a vendor claim, and we report it as one — but it is the kind that would be conspicuous if wrong.

The Caveats No Certification Fixes

Verification remains a human obligation. AI output in legal work must be checked, and the profession has already produced cautionary cases of filings containing fabricated citations. Harvey does not remove that duty; it changes what you are checking.

The economics are genuinely awkward for firms billing by the hour. A tool that makes document review dramatically faster reduces billable time, and resolving that tension is a business model question rather than a software one.

There is no advertised self-serve trial. Harvey is an institutional purchase with implementation and training behind it, which puts it out of reach of solo practitioners and very small firms.

And jurisdiction matters. Legal reasoning does not generalise across systems the way general knowledge does, so coverage quality varies by where you practise.