Every CIO I talk to knows the same uncomfortable fact: a large share of the firm's documents cannot be tied to the client and matter they belong to. Some arrive from clients by email with nothing attached to them. Some sit in folders named after a partner or a deal code nobody remembers. Some were filed before the matter was opened. The number varies from firm to firm, and in my experience few firms have actually measured it.
For years that gap cost lawyers some search time and the occasional frustration. AI raises the stakes. A system that answers from the firm's documents has to know which client each document belongs to. Without that, it can't apply ethical walls, can't respect who's allowed to see what, and can't tell a partner which matter an answer came from. A document without a client and matter is either invisible to the system or a risk inside it.
Why attribution matters more for AI than it did for search
Search returns a list, and a lawyer decides what to open. An AI answer blends material from several documents into one response. If one of those documents belongs to another client, the lawyer reading the answer may never see where a sentence came from. That's the incident every general counsel worries about: one client's information surfacing in work for another client.
Client and matter numbers are what the firm's controls are built on. Ethical walls, conflicts screens, access rules, retention schedules and legal holds all refer to them. When a document carries a valid client and matter, every one of those controls applies to it automatically. When it doesn't, the controls have nothing to hold on to.
Where the gaps come from
Most of the gap comes from ordinary work. Clients send documents to a lawyer's inbox, and the lawyer saves them wherever is fastest. Metadata fields in the document management system are often optional, and busy lawyers skip optional fields. Work starts before the new matter is opened in the system. Lateral hires, mergers and system migrations bring in older records with their own numbering. None of this is anyone's failure. It's what happens when attribution depends on people remembering an extra step.
Pick one source of truth
Every firm already records client and matter numbers somewhere authoritative. At some firms that's the document management system. At others it's the matter management system, or the time and billing system where every hour is recorded against a matter. The first job is to find out which one the firm actually runs on and connect to it.
Often two systems disagree. The document management system says a file belongs to one matter, and the billing system shows the lawyers recorded their time to a different one. Don't let the AI system settle that on its own. Choose the system the firm already relies on for conflicts and billing as the authority, and send every mismatch to a person as an exception. The first few weeks of mismatches usually tell the records team more about the firm's filing habits than any audit has.
Once that's settled, every record that flows into an AI system should be tagged from that source as it arrives. The last thing a firm wants is a large body of data sitting in an AI system with no link to any source of truth in its own environment. Nobody can govern that data, and sooner or later someone asks a question it answers wrongly.
Make it a rule the system enforces
The rule we use is short. Every record carries a client ID and a matter ID, or it doesn't get in. A record never carries two clients. When something arrives that breaks the rule, the system refuses it and records the reason, so the gap is visible and someone can fix it.
The enforcement belongs in the software. People forget optional steps on busy days, and a check at intake doesn't. It also gives IT and the general counsel something concrete to show when a client asks how its information is kept apart from everyone else's.
What to do with documents that have no matter
Refusing unattributed documents doesn't make them go away. The practical answer is to let the system propose a match and let a person decide. The signals are usually there: who sent the document, which parties it names, which email thread it arrived on, which lawyers were recording time to which matter that week. The system suggests the likely client and matter, someone on the records or knowledge team approves or corrects it, and the document goes in with the right tags.
What a firm shouldn't do is guess silently. A wrong match does more damage than a missing one, because it puts one client's document behind another client's walls.
Measure it by practice group
Attribution is measurable, and it deserves a place in the same reports as the firm's other operating numbers. Track the share of documents with a valid client and matter in each practice group, the size of the review queue, and how quickly the queue clears. Start AI work in the practice groups where coverage is already high, and use the other groups' numbers to plan and staff the cleanup.
The number also changes the conversation with partners. Telling a practice group leader that the group's AI results depend on attribution, and showing the group's coverage next to its peers, gets filing habits fixed faster than another memo about metadata.
You don't need a data lake to start
Some firms respond to this problem by planning a data lake first. I've written about why most law firms don't need a data lake before AI. The work described here is smaller: tie every record to a client and matter from the system the firm already trusts, enforce it at intake, and run a clear process for the exceptions. It pays off in every AI project that follows, and it makes the general counsel's questions much easier to answer.
How Reframe handles it
Reframe applies this rule at intake. Records without a client and matter don't get in, proposed matches go to a review queue where a person approves them, and a coverage view shows which practice groups are ready for AI and which need work first.
Drafted with Reframe Editorial Agent, an AI employee at Reframe. 