01  /  EXECUTIVE SUMMARY: Proprietary AI in BigLaw is a hybrid model, not a from-scratch build

 “Proprietary AI” at large law firms usually does not describe a foundation model trained from scratch. The dominant architecture across the market is hybrid, combining four elements: licensed foundation-model capability from providers such as OpenAI, Anthropic or Microsoft; specialist legal applications such as Harvey, Legora, Spellbook or DeepJudge; a firm-built layer covering document retrieval, permissions, prompts, workflows and evaluations; and lawyers who own validation, risk and workflow design.

This hybrid approach is also the most defensible one for a firm operating at Kirkland's scale. Training a fully proprietary model is expensive, difficult to keep current against fast-moving foundation-model providers, and unlikely to outperform specialist vendors across every practice-area workflow. The advantage a large firm can sustain over time is its private precedent base, its workflow design, its security controls and its adoption data, not the underlying language model.

A factual limitation applies to this analysis. The underlying dataset does not identify a confirmed Kirkland partnership with a named foundation-model or legal-AI vendor. Kirkland's specific model providers, contract terms, usage levels and technical architecture should be treated as unverified until confirmed directly by the firm or a vendor. An internally branded interface, such as “Kirkland AI,” should not be read as evidence of an internally trained model.

02  /  BUILD VS. BUY Four sourcing strategies, grouped under one label

Firms and vendors often group four materially different strategies under the single heading of “proprietary AI.” Each carries a different ownership position, a different cost profile and a different risk. The table below compares them on what the firm actually owns at the end of the exercise.

Source: DreamLegal Market Intelligence, Build vs. Buy: Proprietary AI at Kirkland and Major Law Firms.

How to compare the four options

A firm choosing between these strategies is answering three separate questions at once: how much control it needs, how fast it needs to move and where its real competitive advantage sits. In practice, most large firms do not choose one strategy for the whole organization. They train nothing, fine-tune selectively where data is clean and the task is narrow, build the retrieval and workflow layer themselves because that is where firm-specific knowledge lives, and buy or configure a specialist product for everything else.

•     Control: highest with a trained or fine-tuned model, lowest with a configured off-the-shelf product.

•     Speed to deployment: fastest with a configured product, slowest with a trained foundation model.

•     Cost and maintenance burden: rises sharply from configuration, to retrieval-layer build, to fine-tuning, to full model training.

•     Source of durable advantage: for a law firm, this is almost always the retrieval and workflow layer, since it encodes precedent, taxonomies and playbooks a competitor cannot copy.

03  /  CASE STUDY Assessing Kirkland's proprietary AI strategy

Kirkland's practice concentration in private equity, M&A, restructuring and complex disputes points to a specific set of high-value AI use cases. These are the workflows most likely to justify internal investment, regardless of which vendors sit underneath them.

Likely high-value use cases

•     Portfolio-company and fund-document analysis

•     Purchase agreement and disclosure-schedule review

•     Diligence across large document sets

•     Debt and restructuring instrument analysis

•     Precedent retrieval and clause comparison

•     Litigation chronology and evidence analysis

•     Drafting from firm-approved language

•     Matter staffing, pricing and knowledge reuse

Where the proprietary value is likely to sit

Across these workflows, the elements a firm like Kirkland can actually own and defend are its precedent and transaction datasets, its matter-centric access controls and ethical walls, its taxonomies for funds, transactions, debt instruments and disputes, its lawyer-approved playbooks, its evaluation sets built from historical matters, and its integration with document management, Word, Outlook, timekeeping and client portals. None of these require a proprietary foundation model. All of them require sustained internal investment.

What vendors and buyers should verify before accepting a “build” claim

A firm-branded AI product can represent several very different technical arrangements: a custom interface calling an external model API, a Microsoft Azure OpenAI deployment inside the firm's own cloud environment, a retrieval system layered over a document management platform, a white-labeled or deeply configured specialist product, a multi-model gateway routing tasks across providers, or a genuinely fine-tuned model built for one or two narrow workflows. Branding alone does not distinguish between these. The questions below separate the claim from the architecture.

Until these questions are answered, Kirkland should be classified as a probable hybrid build-and-buy environment. The available evidence does not support treating Kirkland, or firms making similar claims, as proof that major firms are replacing external AI vendors with internally trained models.

04  /  MARKET PATTERNS How major law firms are actually sourcing AI

Across the firms for which public information exists, three patterns recur: vendor-led deployment at scale, a firm-owned interface layered over external infrastructure, and structured co-development between a firm and a vendor. These patterns are not mutually exclusive, and most large firms run more than one at the same time.

A. Vendor-led deployment

Linklaters and Legora

The strongest confirmed example in the underlying data is Linklaters' deployment of Legora across all 30 global offices in September 2025. Legora's platform covers agentic legal research, drafting and document analysis, large-scale tabular review, deep Microsoft Word integration and M&A due diligence capability. This is a broad purchase of an external legal-AI operating layer, not a pure internal build. Legora's subsequent partnership with Box, connecting its agents to an enterprise content layer, reflects a wider industry move away from isolated chat interfaces and toward systems that operate directly over governed firm documents.

Harvey

Harvey's product direction illustrates why firms continue to buy even when they run internal development teams. Recent additions to the platform include native audio and video transcription, image, chart and diagram interpretation, full-inbox Outlook search, matter-specific email drafting, and automatic identification of messages that require attention. Reproducing this internally would require multimodal model integration, Outlook permissioning, security review, interface development and ongoing maintenance. A vendor can amortize that cost across many firms; a single firm cannot.

B. A firm-owned interface over external infrastructure

Several major firms have built internally branded assistants or knowledge platforms that sit on top of external model infrastructure rather than replacing it. Commonly reported examples include Clifford Chance Assist, associated with Microsoft's Azure and OpenAI infrastructure; Dentons' fleetAI, an internally branded generative-AI environment; Linklaters' Laila, an internal assistant developed alongside the firm's broader specialist-vendor adoption; and Travers Smith's YCNBot, an internal generative-AI assistant.

The strategic point is that an internal brand and an external vendor relationship are not competing choices. Linklaters maintains Laila for general internal knowledge while deploying Legora firmwide for complex legal work. A single firm can run an internal assistant for general knowledge, a platform such as Harvey or Legora for complex legal work, a dedicated research platform for cited legal authority, Spellbook for Word-based contract review, DeepJudge for enterprise knowledge retrieval, and a separate workflow platform for governed process automation, all at once.

C. Co-development

Co-development is often the most realistic option for the largest firms, because it combines vendor engineering capacity with law-firm domain expertise without requiring the firm to become a software company. In a typical structure, the vendor supplies the product, the underlying models, integrations and support, while the firm supplies anonymized examples, lawyer feedback and workflow design. Both sides build task-specific evaluations together, and the firm negotiates early access, roadmap influence and data restrictions.

At the end of the arrangement, the vendor retains the reusable platform and the firm retains its documents, taxonomies, playbooks and client-specific configurations. This split only works if intellectual-property boundaries are negotiated carefully. A vendor should not be permitted to convert one firm's confidential precedents into a feature delivered to that firm's competitors.

05  /  PARTNERSHIP SCALE, IN NUMBERS Reading deal size against deal type

Reported deal and deployment figures span very different units, from lawyers on a rollout to dollars in a funding round. Read side by side, they show where the market is currently placing its largest commitments: capital consolidation at the top of the market, and workflow deployment inside individual firms below it.

Figure 1. Capital committed across three recent legal AI transactions. Source: DreamLegal Market Intelligence.

Figure 2. Deployment and adoption scale across three initiatives, each measured on its own metric. Source: DreamLegal Market Intelligence.

Deployment size correlates with maturity, not with intent. A rollout to more than 400 litigators or a deployment live across 30 offices indicates a vendor has already moved past pilot risk inside that firm. A capital partnership, such as Blackstone's investment in Norm AI, signals conviction earlier in a product's life, before workflow proof is fully established across a large customer base.

06  /  VENDOR-TO-VENDOR PARTNERSHIPS Why no single vendor covers every layer

Legal workflows cross document management, research, contract management, billing, practice management, eDiscovery and analytics. No vendor covers every layer of that stack, which makes vendor-to-vendor partnership a structural feature of the market rather than an optional extra.

07  /  PRODUCT LANDSCAPE Not one market, but several architectural layers

 The legal AI market should not be evaluated as a single list of interchangeable products. The underlying dataset tracks 372 products matched to legal operations management, but those products occupy different layers of the stack. The comparison below covers three general legal-AI workspaces referenced in the source data.

08  /  EVALUATION FRAMEWORK Not every announced partnership carries equal weight

 Before treating a partnership, whether buyer-vendor or vendor-vendor, as strategically significant, it is worth checking it against four tests. An arrangement that fails all four is most likely co-marketing dressed up as a partnership.

What this means by player type

•     Buyers gain the most by treating partnership as a governed process: fixed pilot windows, named metrics, and clear data and exit rights agreed before dependency builds.

•     Specialist and AI-native vendors need a defensible asset the platform partner cannot replicate, such as evaluation data or deep implementation expertise, since distribution alone is not a moat.

•     Platforms and incumbents should be explicit about which capabilities are native and which are partner-supplied, rather than using partnership as a permanent substitute for a roadmap decision.

•     Content providers keep leverage only by enforcing licensing, version control and visible citation in every downstream product.

•     Foundation-model providers lack workflow ownership, which is why they are moving into legal-specific partnerships rather than remaining generic infrastructure.

09  /  CONCLUSION What buyers and vendors should take from this

 For a firm the size of Kirkland, the realistic choice is not between building a proprietary model and buying an off-the-shelf product. It is a decision about which layers of the stack to own directly and which to license. The evidence across Linklaters, Harvey, Legora, Clifford Chance, Dentons and Travers Smith points to the same conclusion from different directions: firms retain ownership of their precedent data, taxonomies, playbooks and access controls, and they license or partner for the model, the interface and the workflow engineering that sits on top of that data.

The partnerships that will matter over the next several years are the ones that pass the workflow integration, learning loop, distribution and shared accountability tests set out in Section 08. Firms and vendors evaluating a new claim of “proprietary AI” or a newly announced partnership should apply the diligence questions in Section 03 and the four tests in Section 08 before treating either as strategically significant. The rest is marketing.

Source: DreamLegal Market Intelligence | Industry A