Case 01: Building firm-owned AI infrastructure for a 30+ year law firm
Axys is delivering the engagement in three deliberate phases: first an AI-ready public foundation, then a governed knowledge and content system, then a private co-pilot designed to retrieve, research, and draft from the firm's own material. The engagement is active, so this case reports verified scope and delivery evidence — not premature ROI claims.
Client: an established Caribbean law firm founded in 1992. The firm remains anonymous here. Axys may reference the engagement in its portfolio, but we do not name the client or publish its confidential material without explicit approval.
The business problem
The firm's authority had accumulated over more than three decades: articles, opinions, case knowledge, legal source material, and the judgment of experienced lawyers. But much of that value lived in a legacy website, separate archives, and the memories of the people who knew where to look.
That created two related risks. Prospective clients increasingly ask AI answer engines before they visit a law firm's website, yet the firm's public expertise was not consistently structured for machine retrieval and citation. Inside the firm, valuable knowledge was difficult to reuse without manual search and senior staff involvement.
The design choice: tool, system, then teammate
We split the work into three phases because each one creates the conditions for the next. A co-pilot cannot be reliable if the underlying knowledge is unstructured. A knowledge system cannot compound if the public and private information architecture is weak. The sequence is part of the solution.
Phase 1 · AI as a tool
The first phase established a modern business-site implementation base and an AI-ready publishing architecture. The work included a faster, machine-readable web foundation, clearer entity and service structure, structured content templates, and a separate legal publishing layer designed for answer-first articles and citation-ready metadata.
AI accelerated research, scaffolding, and editorial production. Human judgment remained responsible for the information architecture, legal positioning, factual review, and what was ready to publish.
Phase 2 · AI as a system
The active second phase moves from pages to institutional memory: corpus assessment, knowledge-base architecture, and a content engine grounded in the firm's own source material. The objective is a governed retrieval layer that can support both public legal education and private knowledge work without treating a general-purpose model as the source of truth.
This phase began with commercial commissioning and a dedicated data-handling agreement covering confidentiality, purpose limitation, AI use, provider controls, ownership, deletion, privilege protections, and audit rights. For a law firm, those controls are part of the system — not paperwork added after the build.
Phase 3 · AI as a teammate
The third phase is the target operating layer: a private co-pilot with distinct research, retrieval, drafting, and review functions. It is designed to work from the firm's approved knowledge, preserve access boundaries, cite the material behind substantive outputs, and keep a lawyer in control of anything that may be published, filed, or delivered to a client.
This phase is not presented as a finished production outcome. Its requirements and trust boundaries have been designed; implementation follows the knowledge foundation and evaluation work.
What is verified today
- The relationship moved from proposal into an active client engagement.
- The public-site and content-hub architecture formed the Tier 1 delivery path.
- The AI content engine and legal knowledge base were commissioned as Tier 2 work.
- A project-specific data-handling and AI-use framework was negotiated before corpus ingestion.
- The private workspace requirements define human approval, retrieval scopes, versioning, review states, and auditability.
What we are not claiming yet
This is a live build, not a retrospective dressed up with projected numbers. We are not yet publishing time-saved percentages, retrieval-accuracy scores, citation growth, or revenue impact. Those claims require a stable baseline and a measured operating window.
The measurement plan covers site performance and schema coverage, AI citation appearances, documents ingested and residual error rates, retrieval accuracy on an evaluation set, and drafting or research time against a measured baseline. Results will be added when they are real.
The reusable lesson
For a domain-expert firm, the moat is not access to a model. It is the corpus, the workflow, the review standard, and the controls around them. The practical path is to make expertise legible first, retrievable second, and actionable third.
That is the same pattern Axys applies beyond legal services: AI as a tool, AI as a system, AI as a teammate — with ownership and governance carried through every phase.
What would your firm's knowledge make possible?
Start with a readiness assessment that maps the work, the knowledge, the controls, and the first build worth funding.