Building the agentic family law firm: why the fact layer matters
Family law firms are adding more AI tools, not fewer. The ones getting it right have one thing in common: a single, verified fact layer underneath all of them.
In the past two weeks, both Harvey and Anthropic announced integrations with Mary. The reason for these integrations is a shift in the winning AI models for law firms. Firms are no longer asking which AI tool to pick. They are running several at once, across teams and matters.
That creates a new problem. Every tool can draft. Every tool can summarize. But if each one is working from its own reading of the file, you get five versions of the facts and no way to know which one survives cross-examination.
The agentic law firm needs a fact layer. Here is what that looks like in practice.
Book a Demo
Why family law is at the front line
Family law is where fact chaos is most expensive, and where AI tools disagree with each other most visibly.
The money moves. Accounts are opened and closed, funds are commingled, separate property gets traced through years of statements. One missed transfer can change a characterization argument.
The story shifts. Parenting positions evolve between the petition, the declaration and the deposition. Spotting the inconsistency depends on someone holding every prior version in their head.
The record arrives late and all at once. Supplemental disclosures, subpoenaed bank records and text message exports land weeks before trial. The chronology has to be rebuilt every time.
The client is paying for every hour. Work that cannot be repeated the same way twice is hard to justify on a bill, and harder to defend in court.
Put a different AI tool on each of these problems and you multiply the versions of the truth. Put a verified fact layer underneath them and every tool starts from the same record.
What the most advanced firms are doing
The clearest model so far comes from a U.S. litigation firm heading into trial on a 100,000-page production. They are using several AI and eDiscovery tools at once. What made it work is a clean division of labour: Mary is the fact engine. Claude is the presentation layer.
Mary builds the complete record. Claude makes it usable. Mary produces the exhaustive, cited fact base. Claude turns it into a ranked, length-limited view a lawyer can act on today. Lawyers get the 10 entries that matter without losing the 140 that sit behind them.
Bespoke dashboards, built with our legal engineers. Working with Mary's legal engineers, the firm built a dashboard for each deliverable its deadlines demand: production deficiencies, deposition "landmines", a court-ordered response and an overall matter view. Each one is linked live to Mary.
Every row keeps its source trail. Each line on a dashboard links back to its Mary entry and to the native file in the firm's eDiscovery platform. Nothing in the presentation layer is unsourced.
Mary reasons, not just retrieves. Search tools find and tag documents. Mary works out what was actually requested across requests, responses and meet-and-confer letters, then checks what was produced against it. That is the analysis no one has time to do by hand on 100,000 pages.
The facts persist across tools. Because every fact is verified and maintained in Mary, it does not change when an attorney moves from Claude to Harvey. The same record can be questioned in either tool and still holds up.
What this means for family law
The same pattern maps directly onto a contested divorce:
- Production deficiency → disclosure deficiency. Which statements, returns and account records were requested, which were produced, and which are referenced but missing.
- Deposition landmines → shifting positions. Every prior statement across petitions, declarations and correspondence, flagged where the story changes.
- Court-ordered deliverables → financial schedules. Tracing and characterization tables where every line is cited to a page.
As the firm brings on more AI tools, the record underneath stays the same, and every tool starts from facts that have already been interrogated and cross-examined.
Four firms, four stages
There is no single right setup. Here is how four firms at different stages are using Mary today.
The more tools a firm runs, the more the fact layer does. The Arizona firm uses Mary for almost everything. The New York firm uses it as the one thing that does not change.
Which setup is right for your firm
The right answer depends on three things: where you want to be as a firm, what matters most to automate, and how you want the costs to appear.
If discovery and financials are the bottleneck: Mary on its own. When auditability and exhaustiveness are the point, and you want to charge the work back to the client, Mary as a stand-alone platform covers the matter from first disclosure to trial.
If you are a small, agentic firm: Claude or Copilot as the interface. Attorneys work where they already are. Mary sits underneath for complex tracing and the audit trail that has to hold up in court.
If you are building your own stack: Mary as the data layer. The most advanced firms want to capture each jump in model capability as it happens. That means plugging in new tools, or giving different teams different tools, without rebuilding the record each time. A shared fact layer is what makes that safe.
The tools will keep changing. The facts should not.
As firms scale agentic work, the interface matters less and the record underneath matters more. Mary now works inside Harvey and Claude, so your verified facts follow the matter wherever your attorneys work.
See how it would work for your firm.
Book a Demo · 30 minutes. No obligation.