Harvey, DeepJudge, and the two retrieval problems inside a law firm

Harvey shipped three things in a single week: a partnership with DeepJudge, a product called Command Center, and a waitlist for Contract Intelligence aimed at in-house legal teams. The DeepJudge piece is the one worth slowing down on, because it quietly answers a question that has been hanging over legal AI for two years: why hasn't a fine-tuned model on top of a giant frontier LLM just solved knowledge retrieval inside law firms already?

The short answer is that there are two retrieval problems inside a law firm, not one, and they want different architectures. Harvey is good at one of them. DeepJudge is built for the other. The partnership is, in effect, an admission that you cannot collapse them into a single index and expect the result to be usable.

I want to walk through what those two problems actually are, why a naive RAG (retrieval-augmented generation — fetching relevant documents and feeding them to a language model as context) setup falls over on the second one, and what the partnership tells us about where the field is heading.

The two corpora

External legal knowledge. This is case law, statutes, regulations, published commentary, treatises. It is enormous, mostly public or licensable, relatively static (a statute does not change overnight), and the same for every firm in a jurisdiction. Harvey's core competence sits here: a frontier model fine-tuned on legal corpora, with reinforcement learning from lawyer feedback, that can read a question and produce a grounded legal answer.

Internal firm knowledge. This is everything the firm itself has produced. Past memos, deal bibles, closing checklists, negotiation notes, redlined drafts, advice given to clients three years ago that turned out to be exactly right. It is smaller in volume than Westlaw but vastly more valuable for a specific matter, because it encodes how this firm thinks, what positions it has taken, and what it has learned the hard way.

A partner who has been at a firm for twenty years carries a mental index of the second corpus. A new associate does not. The promise of internal retrieval is that the associate can ask the index the same question they would ask the partner, and get something close to the same answer.

Why one index does not handle both

If you have spent any time with RAG, your instinct will be: just index everything, embed it, search by similarity, hand the top results to the model. This works fine on a corpus of public documentation. It falls apart on firm-internal knowledge for reasons that are not really about retrieval algorithms at all.

Access control is per-document and per-user, not per-corpus. Lawyer A worked on Matter X for Client 1. Lawyer B is now working on Matter Y for Client 2, who happens to be Client 1's adversary in unrelated litigation. Lawyer B must not see anything from Matter X. This is an ethical wall, and it is not negotiable. A retrieval system that returns the most relevant document without checking whether the asking user is allowed to see it has just committed a serious breach. The wall has to live inside the retrieval layer, not be bolted on after.

The most useful knowledge is in the worst formats. Polished client memos are easy to index. The genuinely valuable artefact is often the tracked-changes Word document showing how a clause evolved across six rounds of negotiation, or the email thread where the partner explained to the associate why they rejected the counterparty's position. PDFs of scanned signed documents. Excel sheets with deal terms. None of this is clean prose ready for embedding.

Recency and authority matter in legally specific ways. A 2019 memo on a regulation that was amended in 2023 is not just stale, it is dangerous. The retrieval system has to understand that some documents supersede others, that some are drafts and some are final, and that "the partner signed off on this version" is metadata that changes the document's authority.

The query is rarely the question. When a lawyer asks "how have we handled MAC clauses in healthcare deals," they do not want a similarity search. They want: every healthcare deal we have done, the MAC clause from each, grouped by counterparty type, with the negotiation outcome. That is closer to a structured query over extracted entities than a vector search over text.

This is the gap DeepJudge is built into. Their pitch is that internal knowledge retrieval needs its own pipeline: ingestion that understands legal document types, extraction that pulls out the structured bits, an index that respects access controls at query time, and a query interface that maps natural language onto the structured-plus-unstructured hybrid underneath.

What Command Center probably is

The neutral coverage describes Command Center as an "intelligence layer" — analytics over Harvey usage across a firm, with natural-language queries, benchmarking against peer firms, and recommendations on adoption. Reading the announcements alongside Harvey's earlier product surface, I think the cleanest mental model is this: Command Center is not the orchestration layer between AI capabilities. It is the orchestration layer between humans and the deployment.

Innovation partners, knowledge management leads, and legal-ops people now have a dashboard that tells them which practice groups are using Harvey for what, where the power users are, where adoption has stalled, and how their firm compares to anonymised peers. That is genuinely useful in a firm of two thousand lawyers where adoption is uneven, but it is a different thing from the "agentic orchestrator routing tasks to specialised sub-agents" pattern that the word "Command Center" might suggest.

I am flagging this because the metaphor in the product name is doing work the product may not do. Worth watching for an actual technical spec before assuming multi-agent coordination is what is shipping.

The pattern, drawn out

Step back from Harvey specifically. The architecture that is emerging across legal AI, and arguably across enterprise AI generally, looks like this:

A frontier model, fine-tuned for the domain, sits at the centre. Around it, two distinct retrieval surfaces: one for public domain knowledge (curated, large, shared across customers) and one for customer-private knowledge (smaller, access-controlled, customer-specific, with its own ingestion pipeline). On top, a workflow layer that sequences calls to the model and the retrieval surfaces for specific tasks (draft this, review that, find precedent for the other). And alongside all of it, an observability layer so the humans buying the thing can see what it is doing.

Harvey is now visibly assembling that full stack. The model and the public retrieval were already there. DeepJudge plugs the private retrieval gap. Command Center is the observability. Contract Intelligence is a vertical workflow application of the same stack pointed at in-house teams, where the document mix shifts from bespoke advice to high-volume contract review and the integrations need to extend into CLM (contract lifecycle management — the systems in-house teams use to store, route, and track contracts through their life).

That is the financial context the partnership sits inside. A company valued like that does not partner on a gap it could trivially close itself. Harvey building a DeepJudge-equivalent from scratch is presumably possible. Choosing to partner is a statement that the second retrieval problem is its own discipline.

What to watch

Three things, if you are following this closely.

Whether the partnership goes deeper than distribution. A logo-swap arrangement where Harvey sells DeepJudge to its customers is different from a real integration where Harvey workflows call DeepJudge retrieval as a first-class capability. The press release does not distinguish. The actual product will.

Whether Contract Intelligence can land in the in-house market. Ironclad, Robin AI, DraftWise, and Spellbook have multi-year leads and CLM integrations Harvey will need to replicate. The "client-pull" finding from Litera, that corporate buyers now tell their outside firms which AI tools to use, cuts both ways. If corporates are choosing the tools, Harvey's law-firm relationships do not automatically translate.

Whether the dual-retrieval architecture becomes the default. If it does, every vertical AI company is going to need a partner or an acquisition for the private-knowledge side. The companies that figured out access-controlled, format-messy, legally-aware ingestion early will end up looking very well placed.

Further reading

• Artificial Lawyer, Harvey Launches Command Center, Partners With DeepJudge, 22 May 2026
• Legaltech News, Legaltech Rundown: Harvey Partners With DeepJudge, 22 May 2026
• Harvey, 2026 SKILLS Legal AI Survey: Q&A With Oz Benamram
• DeepJudge product overview: https://www.deepjudge.ai

¶