Harvey Wants to Build a Model That Thinks Like Your Firm

Most law firms using AI today are using shared tools. Harvey handles millions of queries. CoCounsel serves thousands of firms. Legora serves 1,200 organizations across 50 markets. Your firm uses the same AI as your competitors.

Harvey's co-founder has a different destination in mind.

What Harvey Confirmed

In June 2026, Harvey CEO Winston Weinberg confirmed that Harvey is conducting proof-of-concept studies with law firms to train open-source LLMs to encode the firm's way of working.

Not "encode legal text." Encode how the firm works.

The technical specifics, per Artificial Lawyer's June 18 report: Harvey is training models to capture workflows that "span months and take dozens of associates" — entire matter lifecycles, not individual tasks. The model learns to control legal tech tools, manage sub-agents, and know when to request human assistance. It incorporates firm-specific playbooks, reference data, and what Harvey's co-founder Gabe Pereyra calls "digital twins" of methodologies.

Client-matter-specific: the model learns how a firm works with specific longstanding clients. Not just how to do contract review — how your firm does contract review for the manufacturing client you've represented for fifteen years.

Pereyra's stated goal: "Enable law firms to build their own specialized models and own their own intelligence."

Why This Is Different From Everything Else

When you subscribe to Harvey today, you're renting access to Harvey's capability. Harvey answers legal questions. Harvey analyzes documents. Harvey conducts research. It does these things well, and the June 2026 update bundle — US case law from 9 million court opinions, MCP connections to iManage and NetDocuments, .pst archive support — made it significantly more capable.

But every law firm with a Harvey subscription has access to the same Harvey.

Firm-owned AI is different in kind, not just degree. Consider what a senior partner at your firm actually knows: the approach your firm uses for a particular contract negotiation, the risk tolerance of specific clients, the clause language that has worked in your jurisdiction, the way you structure arguments for a particular judge. That knowledge lives in the partner's head. It doesn't live in Harvey's training data.

The POC studies Harvey is running are testing whether that institutional knowledge can be encoded into a model — so that the approach your senior partners have developed over twenty years becomes a capability any attorney at your firm can access, consistently, at scale.

The Kirkland Signal

Kirkland & Ellis is the largest law firm in the world. In June 2026, it has been hiring GPU cluster infrastructure experts.

That is the extreme version of what Harvey is building toward from the middle. Kirkland isn't just exploring Harvey's POC — it may be building the infrastructure to train its own models in-house, using fifty years of fund formation precedents, deal data, and outcome records as training input.

That path is not available to a 15-attorney firm. But Harvey's program is building toward making the concept accessible at lower price points — eventually.

Thomson Reuters is running similar studies. Harvey has open-sourced its LAB benchmarks to measure legal agent capability across practice areas. The direction of travel in the legal AI market is clear: the next competitive frontier is not which firm has the best AI tool. It is which firm has encoded the most proprietary knowledge into the most capable model.

What This Means for Your Practice Today

The practical implication is not "go build a proprietary model." It is: start building the inputs.

Every firm has institutional knowledge that would be valuable to encode. Most firms have not documented it. The knowledge lives in how senior partners approach matters, in email threads that never became standard procedures, in preferences that get communicated verbally in case reviews but never written down.

The firms that benefit most from Harvey's capability — when it becomes available to smaller practices — will be the ones that have already documented their workflows, their client-specific approaches, and their matter history. The firms that scramble to assemble that institutional knowledge when the training capability arrives will find they don't have much to train on.

Three actions for a practice under 50 attorneys:

Document one workflow this month. Pick the highest-volume matter type your firm handles — NDA review, demand letter drafting, a specific type of contract negotiation — and write down how your firm does it. Not a general description. The specific steps, the clause language you use, the considerations that apply. If a competent new associate joined tomorrow and you needed them to handle this the way your firm handles it, what would they need to know? Write that down.

Activate Harvey Agent Builder. This is the accessible version of firm-owned AI available today. Build agents for your highest-volume workflow types. An NDA review agent that applies your firm's specific deviation checklist. A matter summary agent that outputs in your firm's preferred format. Each agent you build is an encoded piece of your methodology — the foundation of whatever firm-specific training becomes possible next year.

Audit your matter data. If your matter history is fragmented — files in different document management systems, some in email archives, key information living in attorneys' heads — start consolidating it. The firms that will have the richest training data are the ones whose matter history is structured and accessible, not scattered across systems.

The competitive moat in legal AI is shifting. For the past two years, it was having any AI at all. For the next two years, it will be having AI that knows how your firm works specifically.

Harvey's POC is building that capability. The question is what your firm will have documented when it becomes available.