Your firm's filing system is the most valuable part of its AI

In the last post I explained the AI problem most firms haven't hit yet: once AI is working across every matter, everything piles into one memory, and it starts grabbing the wrong client's details. The fix comes from a very old idea: a well-kept file room.

Here's the payoff first. When one client changes its settlement authority, you tell the AI once, and no other client's matters are touched. When one county updates its default attorney-fee schedule, only that county's defaults change. The rest of this post is how that works.

What does the research say?

In March 2026, Jake Van Clief and David McDermott published Interpretable Context Methodology: Folder Structure as Agentic Architecture.¹ Their central idea is that the way you organize folders can be the thing that directs an AI through a job.

Each folder holds one stage or one area of work, plus a short plain-English file that says what's in the folder, what the AI should read, and what it should produce. The AI opens only the folders the current task needs. In the paper's examples, each step reads a few thousand words of focused material instead of tens of thousands of words of everything.¹

Interpretable means a person can read all of it. Every instruction and every draft along the way is a file someone at the firm can open, check and correct, not something locked where only a vendor can see it.

How should a law firm organize its files for AI?

Here's a simplified layout for a firm that handles collections and unlawful detainers. Each INSTRUCTIONS file is a page or two of plain English that a paralegal could have written.

firm/
  INSTRUCTIONS.md              how this firm works, who approves what
  practice/
    collections/
      demand-letter/
        INSTRUCTIONS.md
      default-judgment/
        INSTRUCTIONS.md        steps, templates, checks before filing
    unlawful-detainer/
      notices/
        INSTRUCTIONS.md        which notice, when, and why
  courts/
    san-diego-superior.md      local rules, default attorney-fee schedule
    riverside-superior.md
  clients/
    example-lending/
      INSTRUCTIONS.md          settlement authority, letter preferences
      matters/
        2026-0412/             documents, matter log, drafts

When a paralegal asks for the default judgment packet on matter 2026-0412, the AI is given the firm's instructions, the default judgment procedure, that court's requirements and that one client's file. It isn't given the firm's other matters, so it has nothing to borrow from. Making sure of that is part of the setup, not something folder names do on their own.

The same layout works for family law and trust & estate matters, each with its own procedures and forms.

What happens when a client's instructions change?

This is where the setup pays off day to day. Say a paralegal tells the AI, "Example Lending lowered settlement authority to 65 percent." Because each client's instructions have their own home, that note goes into Example Lending's file and nowhere else. The next Example Lending matter uses it. Every other client is untouched.

Or: "Riverside updated its default attorney-fee schedule, here's the new one." California lets each court set that schedule by local rule,² so it belongs in that court's file. San Diego defaults keep using San Diego's schedule.

The same works for a courtroom ("this judge wants courtesy copies of every motion"), a document type, or the whole firm. One sentence, stored once, used every time it applies.

To be precise about the source: the paper covers step-by-step work with a person reviewing each stage. Routing everyday corrections into the right file is how I apply the same principle to a firm's ongoing work.

Who owns the files, the firm or the vendor?

The AI program that reads these folders can be replaced, and so can the model behind it. The folders, instructions and procedures are the part that belongs to your firm. They're plain text files in your own storage that you can back up, copy and move. A new AI tool still has to be connected and checked, but the firm's know-how doesn't start from zero.

Compare that with putting your firm's know-how inside a vendor's platform, where it lives on their servers and under their contract.

Is this hard for staff to use?

Simple to use. Careful to set up.

Day to day, staff don't learn anything technical. They ask for what they need in plain English. The careful part is the setup: mapping how the firm actually works, deciding where each kind of knowledge lives, and writing the first set of instructions. That has to be right at the start, because a messy setup gets messier with every use.

That setup is the work I do for firms. In the next post I show what it looks like from a paralegal's desk.