Product
AI Support Pipeline
End-to-end system: ticket classification, duplicate detection, intelligent routing to the right team and person, auto-generated FAQ, and a conversational query interface.
Pilot project · AI Support Operations
Dispatch
Built and validated an AI-powered support-routing prototype for a US fintech after mapping cross-team bottlenecks. The pilot scenario covered classification, deduplication, team routing, and workload balancing across 2000+ monthly tickets.
Fractional
product leadership
2025
Overview
Validated AI routing in one week
Client & context
US Fintech
Fast-growing company processing 2000+ support tickets per month, with a 10-person support team spread across multiple specialized squads.
My Role
Fractional product leadership
One-week engagement: identified the bottleneck, designed the system, built a working prototype with OpenAI API, and validated a pilot scenario with historical ticket data.
The problem
2000 tickets a month, all routed by hand
Manual triage
One person sorting everything
Every ticket had to be manually categorized, then assigned to the right team, then balanced across agents. One person spent a significant chunk of their time just routing, not solving customer problems.
Downstream waste
Duplicates, imbalance, no self-serve
Duplicate tickets clogged queues. Some agents were overloaded while others sat idle. Many questions already had answers buried somewhere, but there was no FAQ, no knowledge base, nothing.
The system
Four layers of automation, one pipeline
Each ticket passes through a multi-step AI pipeline: classify, deduplicate, route, and balance, then the same data feeds a live FAQ and a conversational query layer.
1
Classify & deduplicate
OpenAI API scans each incoming ticket, assigns a category from the company's taxonomy, detects near-duplicates, and flags tickets that already have known answers.
2
Human-in-the-loop fallback
If the model's confidence score for classification or deduplication drops below 95%, the system holds the ticket in a designated verification queue for human review, ensuring deterministic quality.
3
Route & balance workload
High-confidence tickets are distributed evenly across the correct specialized squad based on current load, preventing one person from drowning while others wait.
4
Generate FAQ & enable queries
Recurring patterns are surfaced to auto-generate a structured FAQ. A conversational chatbot lets the team query the ticket base directly, e.g. "How many refund requests last month?"
Execution
One week from audit to validated prototype
Product thinking
Map the real bottleneck first
Started by observing the support workflow end-to-end. The problem wasn't response quality, it was everything before the response: sorting, routing, finding existing answers. That became the scope.
Quality Metrics
Evaluation-driven iteration
Didn't just "guess" if the prompt worked. Built an evaluation dataset of 200 past tickets and ran testing scripts to measure hallucination rates and categorization accuracy across different prompt iterations.
Trust & Transparency
Designing AI UX patterns
To get support agents to trust the new tool, routing decisions included citations (showing exactly what part of the ticket triggered the classification) and easy human override buttons if the AI was wrong.
Co-construction
Build with the operators, not for them
Involved the support team from day one, understanding their workflow, getting feedback on routing rules, adapting to how they actually work. Internal tools need the same product rigor as customer-facing ones: without buy-in, nothing ships.
Delivery
Validated prototype in 5 days
Delivered a functional prototype and validated the workflow on historical tickets in an internal pilot scenario. By prototyping standalone first, we de-risked logic and human-in-the-loop workflows before any production integration.
Measured impact
A validated pilot with clear time savings
15h/week (est.)
Manual triage eliminated
Estimated from operator workflow mapping and manual time-per-ticket comparison. Tickets were routed in seconds instead of roughly 3 minutes of manual sorting each.
92%
Classification accuracy
Automated category assignment matched the company's existing taxonomy in simulation runs on 10,000 historical tickets.
30%
Duplicate tickets flagged
Near-duplicate detection surfaced 30% of tickets as repeats or variations in the same 10,000-ticket simulation sample, reducing queue clutter and enabling a self-serve FAQ.
Takeaway
What I learned
Prototype standalone, integrate later
Don't wire into a live system on day one. Build a working prototype on the side, validate it with real data, prove the logic works, then connect. This approach is faster, less risky, and makes stakeholder buy-in much easier because you can show results before asking anyone to change their workflow.
Treat internal tools as real products
The biggest risk with AI automation isn't technical, it's adoption. If you build something and drop it on a team, they'll resist it. Co-constructing with operators from the start creates something that actually fits their workflow, and people who feel ownership don't push back on change.