In my work with law firm marketing and business development professionals, one problem keeps coming up: the AI platform rarely knows enough about the firm.
A team may use Copilot, ChatGPT, Claude, Harvey, or another platform to draft a client alert, write an RFP response, or prepare a partner's LinkedIn post. The response can read well and still miss the mark. It might overlook the firm's voice, describe a practice too broadly, misjudge who the audience actually is, or ignore how a client relationship has grown over time.
The knowledge that would fix this already exists somewhere in the building. A CRM holds relationship details. An experience management platform contains lawyer and matter profiles.
iManage houses prior work. SharePoint and the intranet carry internal guidance. The website reflects approved market language. Each system knows part of the firm, but no single system knows the whole firm.
AI receives fragments instead of the full picture.
Google's Open Knowledge Format, known as OKF, offers an early framework for organizing those fragments into portable, structured knowledge. Its use in legal marketing remains largely unexplored.
Based on the research available, no law firm marketing or BD team has publicly documented an OKF implementation yet. That gap leaves room for thoughtful testing, along with several open questions worth working through.
AI platforms build their answers from whatever material sits in front of them at that moment. A detailed prompt can supply some direction, but prompts get cumbersome fast if a person has to attach the firm's voice guide, practice description, client background, approved terminology, experience records, and competitor context every single time.
Google describes this broader problem in its introduction to OKF. Organizational knowledge sits scattered across catalogs, wikis, shared drives, third-party systems, code comments, and the heads of individual employees. An AI agent has to piece together an answer from sources that each use different structures and labels.
A law firm makes that difficulty even sharper. The same client might appear under several different names across CRM records, billing systems, matter databases, and old proposal files.
Practice descriptions can vary from office to office or biography to biography. Approved website copy might even conflict with an outdated pitch deck sitting in someone's folder. Without a clear hierarchy for which source wins, the AI has almost no basis for deciding which version of the truth to trust.
Client pressure adds real urgency here. According to Litera's 2026 State of Legal AI research, 85% of law firms say their clients are the ones driving AI investment decisions.
That same research ranks people, talent, and expertise as the most valuable asset for AI differentiation at 24%. Custom workflows and use cases follow at 18.7%, with proprietary data and knowledge close behind at 13.3%.
Those numbers put institutional knowledge near the center of the whole AI conversation. Picking the right model is one decision. Getting the quality, ownership, and structure of firm knowledge right is a separate one, and arguably a harder one.
A useful starting point is a plain inventory of every source that shapes marketing and BD work. The goal here is visibility: which system holds which kind of knowledge, and who owns it? Update frequency counts too, though tracking it takes separate effort.
Legal Technology Hub describes experience management systems as a central place for professional experience, education, and bar admission information.
Litera's description of its Foundation platform adds client and lawyer profiles tied to specific matters.
These sources don't all carry equal weight. The website might hold the official public description of a practice. The experience platform might hold richer, more detailed matter data.
The CRM might have the most current relationship history. A proposal archive might contain language written specifically for one industry or one buyer.
A good inventory records those differences. A source's name reveals little on its own. Ownership, review status, confidentiality level, intended audience, and the date of the last update help an AI system, along with the humans checking its work, gauge how much weight each piece of material deserves.
Traditional systems organize information around files and folders. AI work often depends on relationships that cut straight across those folder lines.
Think about a request for an RFP draft aimed at a health care company expanding into three new states. The useful context might include attorneys with the right regulatory background, related firm experience, existing client relationships, approved industry language, geographic coverage, and recent thought leadership. Those pieces almost never live in one document.
The background research points toward a shift in how knowledge gets organized, where relationships become part of the stored information itself. A client links to matters. Matters link to lawyers and practices.
Lawyers link to industries, jurisdictions, publications, and credentials. Content links to audiences, topics, and approval status.
This changes the actual task an AI tool is being asked to do. Simple file retrieval might return a pile of proposals and biographies for a person to sort through. Connected knowledge can instead help the system pull together a bounded set of relevant concepts, tagged with labels that explain how those concepts relate to each other.
For marketing leaders, this work starts with something surprisingly basic: a shared vocabulary. Firms often use different terms for the same industry, service, office, or client.
Before anyone packages knowledge for AI, someone needs to settle on preferred names, accepted synonyms, and which source counts as the authority. Otherwise, the new format just carries the firm's old inconsistencies forward into a new home.
Google introduced OKF in June 2026 as an open specification for packaging metadata, context, and curated knowledge into a portable format. The design relies on Markdown, a plain text format that people can read easily and that many software tools already process without trouble.
An OKF package can live inside a file repository or a version-controlled environment. The format needs no dedicated runtime and no mandatory software development kit, according to Google.
Teams can open the files in a regular text editor and track changes over time. Moving the whole package between compatible environments takes little effort as well.
That portability matters for a practical reason. A knowledge package built tightly around one AI vendor can create painful migration work down the road if the firm ever switches tools.
Vendor-neutral is how Google presents OKF, though how well it actually works across major AI platforms remains unproven in the available research. Using it with Copilot, ChatGPT, Claude, Harvey, or any other system would still need real technical testing first.
OKF also serves a different purpose from llms.txt and AGENTS.md. TinyCommand's independent explanation places OKF at the deeper knowledge layer, while llms.txt works more like a map and AGENTS.md gives instructions to agents working inside a repository.
OKF carries no documented search-ranking benefit, either. Its immediate value lies in organizing knowledge for AI use and human review. Search visibility remains a separate question OKF does not address directly.
Converting a whole firm's knowledge at once would pull in too many owners, systems, permissions, and content types for a first attempt. A narrower, bounded domain makes for a far more useful test.
A marketing or BD team could pick one contained workflow, such as drafting first versions of content for a single practice group. The supporting knowledge might include the approved practice description, selected lawyer biographies, public representative matters, voice guidance, audience profiles, and published insights.
The background material describes a process called semantic concept partitioning, which splits long documents at natural heading breaks and adds structured metadata like descriptions, tags, and timestamps. That process turns one giant file into smaller, labeled knowledge pieces that an AI system can pull from with far more precision.
A pilot bundle could follow five stages:
The sample structure in the background material uses a root folder, an index file, and separate folders for resources like guides and processes. The exact labels would shift depending on the workflow and the technical environment a firm chooses.
Testing should compare the AI-assisted output against the team's current process. Useful things to measure might include factual accuracy, how well the voice matches the firm's style, whether sources can be traced back to their origin, how much editing time gets saved, and how many unsupported statements show up.
The supplied research doesn't offer a benchmark for expected gains, so each firm would need to build its own baseline from scratch.
Structured context can make knowledge far easier to access, but it also creates new governance demands. Law firms need clear boundaries around what enters a bundle, who gets to use it, and how a reviewer confirms its status before it goes anywhere near a client.
Google's OKF v0.2 update adds fields for provenance, trust, freshness, lifecycle, and attestation. It can also mark whether a piece of content was reviewed by a named human, confirmed only by a machine, or left completely unverified.
These fields could support several useful controls for legal marketing. Provenance can point back to the original CRM record, approved biography, or website page.
Freshness can flag a matter description that's gone stale. Lifecycle information can label content as draft, active, or retired. Human verification can show clearly whether a qualified person actually reviewed the entry.
Even so, these fields remain advisory only. Google states plainly that they don't function as access controls. A firm would still need its own permissions, security policies, confidentiality rules, retention requirements, and technical safeguards wrapped around both the repository and whatever AI platform connects to it.
The format also remains genuinely early-stage. TinyCommand reports that OKF still lacks an established base of tools and platforms actually using it, and the available research contains no public example of a law firm marketing team putting it to work.
Cross-platform compatibility, ongoing maintenance effort, data quality, and measurable business impact all still need real-world testing before anyone can speak to them with confidence.
That uncertainty calls for a limited scope and close coordination among marketing, BD, knowledge management, IT, security, and risk teams. A pilot should stick to approved material with a defined owner and a clear review process from the start.
Generic AI output almost always traces back to fragmented context. Law firms already hold rich knowledge about their clients, matters, lawyers, markets, and voice, yet that information sits scattered across systems that were each built to do a different job.
OKF offers one way to organize a chosen slice of that knowledge into readable, portable files complete with metadata and trust signals. The format carries credible technical backing alongside limited proof from inside the legal industry.
Both of those facts deserve equal weight when a firm decides how much to invest in testing it.
I would start with one knowledge domain where the source material is already approved, the workflow repeats often enough to matter, and the current burden can actually be measured before and after. A focused test like that can reveal whether structured context genuinely improves source traceability, voice alignment, and editing time within that one domain.
The bigger question will only get answered through real use: can law firms turn scattered institutional knowledge into governed context that travels cleanly across different AI tools? OKF provides a concrete format for exploring that question.
The legal marketing playbook for actually doing it still has to be written.