Resolution time from 4.2 hours to 1.6 hours in 6 months.
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. Operator-Built AI, built on a fixed operating model, not bolted onto chaos.
Situation
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
Approach
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.
- 01
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.
- 02
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.
- 03
Agent swarm and 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.
- 04
Scaling and 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.
Stack
Outcome
Over six months, resolution time fell from 4.2 hours to 1.6 and the escalation rate from 34% to 19%, with 78% of tickets arriving pre-solved. AI cost per ticket dropped from $2.50 to $1.50 and customer CSAT rose from 3.2 to 4.1. Builds like this now ship under the KeyDelta Method and stay current under AAMP.
| Measure | Before | After | Change |
|---|---|---|---|
| Resolution time. Faster, pre-hydrated | 4.2 hrs | 1.6 hrs | −62% |
| Escalation rate. Right team, first time | 34% | 19% | −45% |
| Hydration rate. Tickets arrive pre-solved | 0% | 78% | +78 pts |
| AI cost per ticket. Model switching optimization | $2.50 | $1.50 | −38% |
| Tech satisfaction. Solving, not routing | 3.0 | 4.4 / 5 | +47% |
| Customer CSAT. Faster, more accurate | 3.2 | 4.1 / 5 | +28% |
“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.”