Build and own a Workflow Recommendation Specialist Model.
Given this case or request, which workflow or process path should be applied, selecting among existing defined workflows based on case attributes?
Overview
Routes each incoming case to the correct existing workflow (e.g., refund, replacement, technical support, sales, CRM update, billing, returns, order lookup, password reset, or escalation to a human) based on case attributes rather than manual triage.
Who owns this decision
VP Operations, Head of Process Automation
Why owned AI may fit
The catalog of eligible workflows, their eligibility rules, and case-attribute definitions are specific to one organization's operations, not something a general-purpose model knows without being grounded in that catalog.
What this Specialist Model actually does
Decision: Given this case or request, which workflow or process path should be applied, selecting among existing defined workflows based on case attributes?
This is a larger system with 1 specialist model candidate and 5 supporting or external components.
Who uses this Specialist Model
- VP Operations
- Head of Process Automation
- COO
Where the specialist model sits
This is a larger system with 1 specialist model candidate and 5 supporting or external components.
Case classification/matching model fine-tuned on historical case-to-workflow assignments
Specialist Model
Workflow catalog and eligibility-rule retrieval layer grounding recommendations in current workflow definitions
Supporting component
Confidence scoring and fallback-to-human logic
Supporting component
Confidence scoring and fallback-to-human logic for cases that do not cleanly match a workflow
How Oumi connects to your stack →
Override capture feeding corrections back as labeled training data
Supporting component
How Oumi connects to your stack →
Monitoring layer tracking workflow-selection drift as case mix or workflow catalog changes
Supporting component
How Oumi connects to your stack →
Integration layer handing off the selected workflow
Supporting component
Integration layer handing off the selected workflow to the execution/orchestration system
How Oumi connects to your stack →
How to evaluate and improve it
Evaluation
- Workflow selection accuracy against human-confirmed correct workflow
- Override rate (how often a human changes the recommended workflow before execution)
- Downstream resolution rate for cases routed by the recommended workflow
Feedback signals
- Human agent override of the recommended workflow before execution
- Workflow execution outcome (resolved, escalated, reopened) compared against expectation for the assigned workflow
- Cases that fell back to human handling because no workflow matched, later mapped to a new or existing workflow
- Customer complaints or re-contacts following a workflow assignment, signaling a possible misroute
- Drift in which workflows are selected for a given case type over time (signals a need to update workflow definitions or eligibility criteria)
Model candidates to evaluate
- Qwen3.5-4B (Apache 2.0)
- ModernBERT-large (Apache 2.0)
- NeoBERT (MIT)
These are candidates, not performance claims. Run them against your evaluation set.


