Legal triage assistant for Microsoft Teams

People outside the legal department have legal questions, and not every question can be answered by a machine. Answering without a source, or handling on its own a case that needs a lawyer, is the risk this assistant exists to prevent. It was built for a large company and does the first pass over incoming questions: it answers on its own only what the knowledge base explicitly covers, and the rest goes to people.

The assistant works inside Microsoft Teams. It provides:

• triage of every question as green, yellow or red before any answer;
• answers only from an approved document library, with sources cited;
• automatic refusal and escalation to the legal team on red cases;
• multi-turn conversation, with the last six exchanges as context;
• a record of every interaction (question, answer, triage and citations) for audit;
• an anonymized weekly report sent to a configured recipient.

Triage uses three colors. A question is rated green only when the legal rule is clear, the case has been validated in the knowledge base and there is a source to cite. Yellow is a known topic with a variable that can change the outcome, and the answer carries a warning about what changes the case. Anything else, including an unmapped topic or an uncertain case, is rated red.

Claude decides the color. It queries the library through Libriq and delivers the triage and the answer together, in one structured result. Safeguards enforced in code protect the result:

• a green answer without a source is downgraded to yellow;
• on red, the text written by the model never reaches the user, only the fixed institutional text;
• search reaches only documents that are already synchronized; with none synchronized it returns nothing, rather than opening the whole Libriq account;
• any inconsistency in the model’s output is treated as a safe failure.

The library is fed automatically. Every six hours by default, a process queries SharePoint through Microsoft Graph, with access to the library site only and not the whole environment. It downloads what changed and sends it to Libriq. A document that fails gets at most three attempts.

The legal content starts in an authoring spreadsheet written by a legal specialist, with four tabs. The system compiles it into one file per topic plus a master index, using only rows with approved validation. A rule marked as repealed in the spreadsheet is left out of the compiled base and is never cited.

Teams waits about 15 seconds for a reply, and triage can take 10 to 30 seconds. So the bot acknowledges the message immediately, shows that it is typing and delivers the answer afterward as a proactive message.

The assistant is a proof of concept. It began as a question-and-answer bot over a document library and evolved into the risk triage assistant described here. Sending a formal email to the legal team was left for a later stage.

Acceptance testing uses a set of questions prepared by a legal specialist, each with its expected color. A case expected to be red that comes out in any other color fails acceptance. Accuracy on the other colors is only reported and does not fail acceptance. The full record of each triage is kept for audit and acceptance testing.

The assistant is built on Libriq, the Sigil Labs platform that turns documents into a searchable base with source citation. Here Libriq only retrieves the passages and generates no text, and the library it searches is restricted to approved documents.

It is the same standard Libriq applies: the answer arrives with its source and comes from a known collection.