AI architecture and evaluation strategy for a multi-year program at one of the world's largest law firms, building a unified workspace for diligence, document negotiation, communications, and institutional knowledge across transactional matters.
Global AmLaw 100 Law Firm
8/25/2026
Deal execution ran across email, spreadsheets, and systems that did not talk to each other, leaving no unified view of matter status, no structured record of decision rationale, and no reliable way to judge whether an AI system was fit to run on live matters.
Led the AI architecture workstream, directing a team of ten to design a modular, self-hostable platform inside the firm's own cloud tenant, and established the benchmarking discipline used to decide what shipped.
Our Role
Team Led
Deployment
Status
We led the AI architecture workstream on a multi-year program to build an operating system for deal execution: one workspace combining email, diligence, documents, issues, tasks, rationale, fees, and time tracking, for one of the world’s largest transactional practices.
The remit was foundational rather than feature-level. It covered the platform architecture, the security and authorization model, and the evaluation discipline the firm would use to decide whether an AI system was fit to run on live matters.
Client: One of the world’s largest law firms, with a market-leading transactional practice.
Engagement: Lead AI Architect, directing a cross-functional delivery team of ten alongside the firm’s IT, security, and practice stakeholders.
Four consequences of fragmented tooling mattered most.
No unified operational view. Matter status, ownership, dependencies, and emerging risk lived in different places, and nothing showed all of them at once.
Institutional knowledge went uncaptured. Decision rationale, specialist input, and negotiation context were rarely recorded in structured form. Continuity depended on individual memory, and onboarding was slower than it needed to be.
Manual reconciliation. Attorneys reconciled disclosure schedules against data room contents, issue lists, and document edits by hand. The work was slow and a genuine source of error.
Isolated systems. The document management system, the ethical walls system, email, and matter metadata each operated on their own, which made standardized workflow and reliable supervision hard to achieve.
Firms usually assess AI on a curated demonstration: a few clean documents, a rehearsed question, a convincing answer. Transactional work is not decided on clean documents. It turns on buried provisions, cross-referenced definitions, and deep exhibit hierarchies.
We treated evaluation as an architectural question rather than a procurement one, and built the program around it. A system that has never been measured against a provision three levels down in an exhibit to a schedule cannot be trusted with a live matter, however well it demos. Systematic benchmarking against hard cases replaced anecdotal spot-checking, and measured performance became a condition for shipping.
The platform is modular, self-hostable within the firm’s own cloud environment, and built for high-fidelity document processing under strict data residency requirements.
Core application. A React front end and a Python orchestration backend, with self-hosted relational and vector databases, object storage, cache, and a durable task queue, all running on managed Kubernetes inside the firm’s tenant.
Identity and authorization. Users authenticate through the firm’s enterprise identity provider, but authorization is governed by matter assignments replicated from the business intake system. Before the platform renders anything, it checks matter access against governance flags covering AI use, cloud processing, export control, and data residency, and denies access outright where a user is walled from the matter. Wall enforcement happens before display, not after.
Document engine. A containerized stack handling ingestion (layout-aware conversion and OCR) and inference through a planner, retriever, extractor, and adjudicator pipeline, with GPU-backed nodes where throughput required them.
Hybrid email. Some users sat on on-premises mail infrastructure and others on cloud mail. We evaluated a self-hosted gateway against native hybrid application registration, weighing data retention posture, licensing cost, and operational complexity, under a firm requirement that no third-party service hold mail in transit.
Drafting engine. In-browser editing with round-trip fidelity to Microsoft Word, assessed across open-source and commercial options on self-hostability, fidelity, and license cost at firm scale.
Stated non-goals. The platform does not replace the document management system as the system of record, does not provide firm-wide chat, and does not draft without human review. Fixing these boundaries early did as much for stakeholder confidence as any capability we added.
The MVP reached production. Onboarding is underway, starting with an initial team of three and expanding toward the firm’s full M&A practice.
The engagement also produced the patterns the firm reuses for everything built next: identity, wall enforcement, integration scaffolding, deployment pipelines, logging, and audit. These were built once, for the M&A workflow, and are designed to carry to further practice areas without redesign.
Client identity, product details, implementation specifics, and commercial terms are withheld under confidentiality.
