AI-Powered Request Routing
A confidence-gated, human-in-the-loop AI platform that classifies and routes student requests across eight FRCC departments. Built for CCCS, Colorado’s largest higher-ed system: 13 colleges and 130,000+ students a year.
Requests arrive from students at any hour. The gate scores each one against the 85% line and routes it: green auto-routes with an audit trail, amber goes to a person for review.
AI refines and flags each request, scores and gates it at the 85% line, drafts the reply for staff to approve, and rolls every case into the leadership analytics view.
Three problems, stacked on top of each other.
Students could not find the right office. The AI assistant FRCC had shipped could not be trusted on what mattered. And staff absorbed the gap manually, every week.
With no clear owner, everything defaulted to a single inbox, regardless of which department the request belonged to.
A fragmented contact system
Students faced a complex contact page with 33+ department options and no clear path. Unable to determine the correct recipient, most students defaulted to sending all requests to E-Wolf, regardless of which department actually owned the issue.
An AI assistant no one could trust
The existing chatbot returned wrong or incomplete answers on the things that mattered most, deadlines, payments, financial aid. Nothing it said was verified by a person, so a misinformation risk sat directly on FERPA-governed data. It failed operationally too: students kept calling to confirm what the bot told them, so staff carried the same load it was meant to remove.
Routing done manually, every week
E-Wolf staff spent 5 to 6 hours a week manually categorizing and routing requests. A review of 1,000+ support submissions showed frequent misrouting, unclear categories, and heavy manual triage. That same dataset became the core training source for the routing model and the empirical base for the testing results that later won cabinet approval.
Two fronts of discovery, one problem space.
Discovery ran on two fronts: a 14-week institutional research program that defined the problem space, and product-level analysis of real conversations, workflows, and request data that shaped the routing design.
Discovery working sessions
I served on the Institutional Excellence work group through a 14-week discovery program: twelve working sessions on a weekly cadence, with cross-divisional membership spanning advising, admissions, the registrar, IT, and institutional research. The sessions put me in the room with the people who own every step of the request journey E-Wolf routes.
Consensus mapping and validation
Future-state survey responses were coded question by question into agreement, divergence, and named gaps, 94 points in total, and every theme was checked with subject matter experts, with 18 of 19 confirming. I used the consensus map to separate problems the institution agreed on from problems still contested, and designed for the agreed ground first.
Field evidence and benchmarking
A multi-campus site visit produced 14 listening and working sessions, including four student listening sessions, and six peer institutions were benchmarked across five focus areas. The field evidence confirmed that the routing problems were structural rather than anecdotal, and showed how peer colleges assign ownership and escalation.
Synthesis and problem definition
Current-state inventory rows were synthesized into four themes and eight problem statements in need-and-because format, approved in the Foundation Report. Two of those themes, decentralization and operational process gaps, define the exact problems that E-Wolf's confidence-gated routing addresses.
Student listening sessions
Four sessions across Larimer, Westminster, Boulder County, and virtual put student voice directly into the evidence base: how they ask for help, where requests stall, and what conflicting answers cost them. Their language, not departmental language, set the vocabulary for E-Wolf's request intake.
Three findings reframed the product.
Product-level analysis of 1,242 student conversations, October 2025 to February 2026, showed where the existing assistant broke down, and set the direction for confidence-gated routing.
Students were not browsing.
They were trying to finish urgent tasks. Demand concentrated on operational work, not general information: account and system access drew 789 high-intent conversations, registration 352, and finance 282. The student portal was the most-requested system by a wide margin, at 440 mentions.
The breakdown was at routing, not answering.
Across the window, 193 conversations were deferred to a person, and 129 were account-specific. The assistant guided general questions reasonably well, but it lost the student exactly where it mattered, on the account and access tasks that have to be completed first.
Confidence without verification was the real risk.
The assistant answered in a confident tone even when it was wrong, with no way for a student to check. In an environment governed by student-records law, that is a trust failure, not a minor flaw.
The existing assistant provided outdated financial aid deadlines during testing, gave inconsistent answers to the same question asked differently, and used a confident tone even when wrong, with no way for students to verify accuracy. At an institutional level, AI errors on critical information are unacceptable.
A collaborative relationship between AI and people.
The solution had to create a collaborative relationship between AI and humans, where AI handles pattern recognition and initial categorization while staff provide the nuance, empathy, and contextual understanding that students need.
With insights in hand, I sketched an updated information structure and user flows. The structure was organized around two axes: user-facing portals and AI-powered backend processes.
The logic was reviewable before any pixels.
Wireframing the two surfaces
Before visual design, I wireframed the two surfaces the system needed: a student request form and a staff review console. The low-fidelity set fixed the structure first. The form leads with the student message, surfaces a single AI-suggested department with a confidence read and a change control, and keeps submission to one action. The console splits into a review queue, a request detail with the message and the AI draft, and a respond panel with reassignment and send. Each region carries a numbered annotation that states its intent and its behavior, so the logic was reviewable before any pixels were committed.
Mapping the task flows
I mapped two task flows in standard flowchart grammar so the routing logic was legible to engineering and to the cabinet. The student submission flow runs from request entry through field validation, AI wording refinement, and the confidence decision: at or above 85 percent the request auto-routes, below it the student confirms or corrects the department before submitting. The staff flow runs from the review queue through the request detail, the AI draft, a department check, and send. The confidence decision is the single branch that governs the whole system, so I drew it as the pivot of both flows.
A confidence gate that surfaces uncertainty.
Every AI classification is scored 0 to 100. An 85 percent threshold determines what happens next: tickets above it auto-route to the correct department with a full audit trail; tickets below it land in a staff review queue for one-click reassignment. Every override becomes labeled training data.
The design was deliberate. Higher-ed AI tools fail in a recognizable pattern: confident on general questions, wrong on specific ones, with no way for users to verify the answer. The confidence gate inverts that pattern. Rather than hiding uncertainty, the system surfaces it, which is what makes the model safe to deploy in a FERPA-governed environment.
Four siloed systems, one routing layer.
The hardest problem at FRCC was not routing speed. It was data fragmentation. Wolf Central, Advising, Financial Aid, and the Registrar each held their own ticket data in their own systems, with no shared queue, no shared metrics, and no visibility into where a request stalled once it crossed a team boundary.
Six layers, end to end.
A student portal and a staff console sit on top. Beneath them, orchestration scores and gates every request, a grounded model classifies it, a data layer holds the knowledge, guardrails enforce the threshold and the audit log, and the routing layer delivers to the eight departments.
I designed this to be safe in a FERPA-governed environment: on the low-confidence path the AI never sends without a human, and every decision is logged.
Augmentation by design.
The system augments staff rather than replacing them. The model handles pattern recognition and first-pass categorization. Staff hold the judgment, the empathy, and the final send. The two improve each other: staff corrections become training data, and staff inherit a model that processes routing patterns at scale. Every outbound message passes a human.
Categorization, routing, and response drafting
Every incoming request is assigned a predicted department with a confidence score, which removes most manual triage and improves routing consistency. For each request, the system generates a draft response aligned with institutional policy and tone. Staff can edit, refine, or regenerate the message before sending, maintaining 100 percent human oversight while accelerating turnaround.
Smart request submission
On the student submission side, AI analyzes the issue description and suggests the relevant department. This lowers cognitive load for students who may not know which office to contact, while giving staff clean, pre-structured tickets for faster handling. Students can still override suggestions, ensuring transparency and control.
Staff dashboard and core features
The staff-facing dashboard centralizes all student requests and layers AI assistance on top of existing workflows. The interface provides real-time visibility into workload, routing precision, and departmental trends, turning a previously fragmented support process into a unified, data-informed operations hub.
Unified request management
Staff can view all pending, in-progress, and completed requests in one place, supported by search, filtering, and clear status indicators. A dedicated AI Accuracy metric tracks model performance in real time, so staff can see how reliably the system classifies requests and where oversight is needed. The analytics view shows request load, distribution, and throughput across departments. Donut charts and stacked bars make bottlenecks, spike periods, and under-resourced areas easy to spot, which supports data-driven staffing decisions and helps leadership anticipate student needs.
One system, read at a glance.
I built the interface on a single design system, so the product reads as one platform and scales past the screens shown here. The confidence and AI tokens are deliberate: a staff member can read the state of any request at a glance.
Validated in testing, funded for production.
Testing methodology
The investment case that won production funding
I presented the formal investment case to FRCC leadership in February 2026. Cabinet approved $300K in Year 1 production funding, and Gecko Engage was scoped as production lead with a roadmap through Q4 2026.
The project executes on the SEM Institutional Excellence Work Group charter, which is the strategic alignment the cabinet funds against. The full deck covers that alignment, the confidence-gated architecture in detail, the testing-phase outcomes, and the production roadmap.
Scaling impact by simplifying complexity.
The most meaningful insight from this work was seeing how much cognitive load and operational inefficiency came from unclear routing and inconsistent messaging. By simplifying the flow, tightening classifications, and designing AI-assisted tools that actually support staff, I saw how UX can unlock organizational efficiency, not by adding more technology, but by ensuring it is used thoughtfully.
The best UX outcomes happen when technology, process, and human judgment work together, not in isolation.