Three acquisitions, three ticketing systems, and nobody had the full picture
Most AI service desk projects stall before production. This one didn't.
A post-acquisition IT services firm had tickets bouncing between teams with zero shared context. KeyDelta built an AI platform that pre-solves tickets before a human touches them, cutting resolution time 62% and escalations 45%. The operating model came first: defined ownership, a daily quality cadence, and model abstraction that survived an acquisition without a rebuild.
−62%
Resolution Time
Faster, pre-hydrated
−45%
Escalation Rate
Right team, first time
+78 pts
Hydration Rate
Tickets arrive pre-solved
−38%
AI Cost / Ticket
Model switching optimization
+47%
Tech Satisfaction
Solving, not routing
+28%
Customer CSAT
Faster, more accurate
- Engagement type
- KeyDelta engagement
- Company profile
- PE-backed IT services firm, three acquisitions in 18 months
- Timeframe
- 6 months
- KeyDelta role
- AI delivery lead and operator advisor
- Client confidentiality
- Name withheld
The Situation
Support running on heroics, not systems.
A multi-location IT services company, formed through three acquisitions in 18 months, was drowning in service tickets. Each legacy entity had its own ticketing system, its own triage process, and its own tribal knowledge. Reps were spending more time figuring out where a ticket should go than actually solving it.
Three acquired teams with three ticketing systems, three triage processes, and zero shared knowledge
Reps spending 40%+ of their time on ticket routing and context gathering, not problem-solving
Escalation rates climbing, tickets bounced between teams because nobody had full context
No model flexibility, locked into a single AI vendor with rising costs and inconsistent quality
ServiceNow and Jira running in parallel with no integration, duplicate work everywhere
The Approach
Decision rights first. Agent swarm second.
Before any model shipped, KeyDelta fixed the operating model the AI would run on: who owns hydration quality, a daily ticket-quality cadence, and a single resolution playbook across the three acquired teams. Only then did we build the AI agent platform that intercepts every incoming ticket, assembles context from multiple systems, and proposes a solution before routing to the right human:
Platform Architecture
Built on Python Django with LiteLLM proxy for model abstraction. Hot-swap between GPT-4, Claude, and other models without code changes, optimizing for cost, speed, or quality depending on ticket complexity.
System Integration Layer
MCP (Model Context Protocol) integration connected the agent swarm to internal knowledge bases, customer records, and configuration data. ServiceNow and Jira integrations unified the fragmented ticketing landscape into a single workflow.
Agent Swarm & Hydration Engine
Specialized agents executed in a parallel fan-out pattern: context agent, knowledge agent, and similarity agent ran concurrently, then a resolver agent synthesized their outputs into a proposed solution. A confidence score determined auto-route vs. human review.
Scaling & Optimization
AWS Lambda functions handled compute-intensive analysis tasks serverlessly. Continuous learning from every resolution improved hydration accuracy over time, the platform got measurably smarter each month.
The Results, 6 Months
Support scales with software, not headcount.
Resolution Time
Faster, pre-hydrated
Escalation Rate
Right team, first time
Hydration Rate
Tickets arrive pre-solved
AI Cost / Ticket
Model switching optimization
Tech Satisfaction
Solving, not routing
Customer CSAT
Faster, more accurate
Framework
Why it worked, the VOOCS lens
Vision
Every ticket arrives at a human already diagnosed, contextualized, and half-solved, reps close issues, not chase them.
Outcomes
Resolution time, escalation rate, and hydration accuracy measured from day one. Model costs tracked per ticket to prove ROI in real time.
Ownership
Service delivery owned hydration quality. Each agent in the swarm had a defined scope, no overlap, no gaps, no committees.
Cadence
Daily ticket quality reviews and weekly model performance analysis. Bad hydrations caught in hours, not weeks.
Systems
New agents deploy in days, not months. Model abstraction means no vendor lock-in. The platform survived an acquisition without re-architecting.
"We turned our service desk from a routing nightmare into an AI-powered resolution engine. Tickets that used to bounce between three teams now arrive pre-solved. Customer satisfaction jumped from 3.2 to 4.1. Our techs went from frustrated investigators to confident closers. And AI cost per ticket dropped 38% because we could swap models based on complexity instead of paying premium prices for password resets."
KeyDelta operating lead
Related: AI Transformation and the VOOCS operating system.
Considering AI for your support org?
Operations first. AI second. Value compounding. Tell us where execution is stuck and we'll tell you where AI actually pays back.
Find Your Highest-ROI AI Workflows