Contract request triage

Incoming contract requests rarely arrive in a clean shape. A sales rep pastes a customer email into a portal. Procurement forwards a vendor questionnaire with a PDF attached. A business partner describes a pilot in three sentences and asks for "something legal can sign." Before anyone drafts language, someone has to decide what kind of agreement is being asked for and which playbook applies.

Contract request triage uses a large language model to perform that first sort at intake. The model reads the request text (and optional attachments when your stack supports them), assigns a request type such as NDA, master services agreement (MSA), or statement of work (SOW), and routes the item to the correct team queue and document template. The goal is speed: fewer manual reassignments, less time spent opening the wrong folder, and faster time-to-first-response for the requester.

Legal still owns the queue. Triage does not approve contracts, negotiate terms, or bypass counsel review. It classifies and routes so specialists start in the right place.

Most legal operations teams treat a small set of request types as the routing backbone. Non-disclosure agreements cover confidentiality before deeper discussions. MSAs establish the governing commercial relationship. SOWs, order forms, and work orders define scope, deliverables, and fees under an existing master agreement.

The model is trained or prompted against your taxonomy, not a generic legal dictionary. If your intake form distinguishes "evaluation NDA" from "mutual NDA," or routes SaaS MSAs separately from professional services MSAs, those distinctions belong in the classification schema. Ambiguous language is expected: requesters often say "we need an NDA" when they mean an MSA with confidentiality terms, or they ask for a "contract" when the underlying need is a change order to an active SOW.

When the model is confident, it returns a request type label, a target team or queue identifier, and a template ID drawn from your CLM or document automation library. When it is not confident, those fields stay empty rather than guessing. Empty output is a feature. It signals that a human should classify manually, which is preferable to silently sending a complex MSA request down the NDA fast lane.

A typical flow begins when a request lands in your intake channel: a CLM request form, a ServiceNow catalog item, a Salesforce case, or an email ingestion endpoint. The orchestration layer sends the free-text description (and metadata such as business unit, region, or counterparty name) to the classification step.

The LLM evaluates linguistic cues that humans already use informally. References to "evaluation," "pilot," or "sharing our roadmap" often indicate an NDA. Language about "master agreement," "governing terms," or "standard vendor paper" points toward an MSA. Mentions of "SOW," "statement of work," "milestones," "deliverables," or "SOW number 3 under the existing MSA" suggest a SOW or amendment path.

Routing rules sit outside the model. Product legal may own NDAs for sales-led deals; commercial counsel may own vendor MSAs; a dedicated contracts operations group may handle SOWs under an existing customer MSA. The model proposes type and template; your workflow engine applies hard rules (region, spend threshold, data sensitivity) before assignment.finalizes.

Speed comes from eliminating the repetitive "what is this?" pass. Counsel opens a pre-sorted queue with the likely template already attached or suggested. Requesters get quicker acknowledgment because auto-routing can trigger the right acknowledgment email and SLA clock without waiting for a coordinator to read the thread.

For triage to be trustworthy in a legal context, classification must be inspectable. Each automated decision should cite the specific span of request text that drove it, and the template ID it recommends.

Text span citations give reviewers a one-click sanity check. If the model labels a request as NDA because the requester wrote "need mutual confidentiality before the demo," the cited span makes that reasoning visible. If the span is weak or missing, the reviewer knows to treat the classification as provisional.

Template ID ties classification to action. Your CLM library might expose IDs such as NDA-MUTUAL-US-v4, MSA-SAAS-EMEA-v2, or SOW-PS-STANDARD-v1. Returning the ID (not just a friendly name) lets downstream automation pre-populate the right document shell, merge fields, and approval path without a human searching the template catalog.

When confidence falls below your threshold, leave type, team, and template fields empty. Pair empty output with a single "needs manual triage" queue or a flag on the intake record. Some teams also return a short list of alternate types ranked by confidence; that can help reviewers without forcing a wrong primary label.

Contract request triage is one step in a broader intake stack. It works best alongside companion capabilities on the same request record.

Risk tier assignment at intake layers severity and review depth on top of request type. An NDA for a strategic prospect and an NDA for a low-risk vendor may share a type label but need different counsel paths.

Duplicate and overlap detection catches when a "new" request repeats an active NDA or amends an existing SOW. Triage should run early, but duplicate checks prevent misrouting when the real action is renewal or amendment rather than net-new paper.

Similar request retrieval helps reviewers find prior deals with the same counterparty, business unit, or clause pattern. Classification tells you which template family to open; similarity search tells you which executed example to mirror.

Auto-priority scoring orders the queue after routing. Type alone rarely determines urgency; close dates, revenue impact, and regulatory flags do. Triage gets items to the right team; priority scoring tells that team what to touch first.

Run these steps in a defined sequence on one intake object so legal sees a single enriched record instead of scattered tool outputs.

Most enterprises already store intake in a CLM or ITSM platform. Triage should write results back to that system of record rather than living in a side spreadsheet.

Ironclad and DocuSign CLM both model contract requests, workflows, and template libraries natively. Classification output maps cleanly to workflow stage, assigned role group, and workflow template selection. Template IDs in these systems are usually stable API identifiers; use those IDs in model output so webhook or REST updates can attach the correct workflow without string matching on display names.

ServiceNow is common when legal intake arrives through HR, procurement, or enterprise service catalog items. Here, triage typically updates custom fields on the sc_request or task record (request type, recommended template, confidence, cited text) and drives assignment group rules. Keep legal ownership explicit: ServiceNow can route to "Legal - Commercial" while still requiring counsel to accept the task.

Salesforce often originates requests tied to opportunities, accounts, or partner records. Classification can populate opportunity-level contract type fields, create a related legal case, or trigger CLM sync if Salesforce is integrated with Ironclad or DocuSign CLM. Align field names with what sales operations already reports on so triage enriches CRM data instead of forked terminology.

Regardless of vendor, treat the LLM as a classification service with versioned prompts and logged inputs/outputs. Store model version, timestamp, and cited spans on the intake record for audit. When templates change, update the allowed template ID list in the same release cycle so the model cannot reference retired documents.

Legal must retain queue ownership even when most items auto-route correctly. Define who may override type and template, how overrides feed model improvement, and what SLA applies when classification returns empty.

Measure triage the way you would measure any intake automation: time from submission to first qualified review, reassignment rate, template correction rate, and percentage of requests that exit with empty classification. A rising empty rate often means requesters need better intake forms, not a louder model.

Review a sample of automated decisions weekly during rollout, then monthly at steady state. Pay special attention to MSAs misclassified as NDAs (under-scoping) and NDAs pushed into full MSA workflows (over-scoping). Both create delay and erode trust with business partners.

Human reviewers should be able to change type, team, and template in one action, with the override reason captured. Those corrections become labeled examples for prompt refinement or fine-tuning. Do not auto-close or auto-send contracts based on triage alone; classification accelerates routing, not approval.