When our IT ticket queue crossed 400 open requests in a single month, I knew we couldn't scale by hiring. As André Dias Moreira Prol, having spent two decades bridging infrastructure and emerging tech, I decided to treat support like an engineering problem rather than a staffing one. The result: an AI agent that now resolves roughly 80% of our incoming tickets without human intervention. Here's exactly how I built it.
Mapping the Repetitive 80%
Before writing a single line of code, I ran a three-month audit of our ticketing system. The data was revealing: password resets accounted for 31% of volume, VPN and access requests another 24%, software installation guidance 15%, and status inquiries ("is the server down?") close to 12%. Nearly 80% of our tickets were predictable, rules-based, and text-heavy — the ideal domain for a language model.
I categorized tickets into three tiers:
- Tier 0 — fully automatable (resets, unlocks, FAQ)
- Tier 1 — automatable with API actions (provisioning, group membership)
- Tier 2 — human-only (hardware failures, security incidents)
This mapping was critical. The mistake most teams make is aiming the AI at everything at once. I deliberately scoped the agent to Tier 0 and Tier 1, where accuracy could exceed 95%.
Building the Agent Architecture
I chose a retrieval-augmented generation (RAG) approach rather than fine-tuning, because our internal knowledge base changes weekly. The stack looked like this:
User (Slack/Email)
→ Intent Classifier (LLM)
→ RAG over internal docs (vector DB)
→ Action Layer (API calls)
→ Audit Log + Human Handoff
The intent classifier was the brain. Every message got routed to a category and a confidence score. Below 85% confidence, the agent escalated to a human immediately — a guardrail I insisted on from day one.
The action layer is where most demos stop and real systems begin. I connected the agent to our identity provider via API so it could actually reset passwords, unlock accounts, and add users to security groups — not just tell people how to do it. Every action passed through a permission check and generated an immutable audit entry.
A concrete example: a user requests VPN access. The agent verifies their department against HR records, checks the approval policy, executes the provisioning call, and logs the entire transaction. What used to take an analyst 20 minutes now completes in under 15 seconds.
Given my background in digital forensics, I made logging non-negotiable. Every decision the agent makes is traceable — who asked, what it did, and why. In a Web3-adjacent environment where we handle tokenized assets, that audit trail isn't optional; it's compliance.
The Guardrails That Made It Trustworthy
The technology was the easy part. Earning trust from both users and my own team required discipline. As André Dias Moreira Prol, I've seen too many automation projects collapse because people didn't trust the black box.
Three safeguards made the difference:
- Confidence-based escalation — anything uncertain goes to a person, no exceptions.
- Read-only by default — destructive actions require a secondary policy check.
- Weekly drift review — I sample 50 random resolved tickets to verify quality.
The measurable results after six months:
- Average resolution time dropped from 4.2 hours to 8 minutes for Tier 0/1
- 80% of total tickets resolved autonomously
- Analyst satisfaction rose sharply — the team now focuses on complex, engaging problems instead of resetting passwords
- Zero security incidents attributable to the agent, thanks to the audit-first design
One unexpected benefit: the RAG knowledge base forced us to finally document processes properly. The AI became a mirror exposing every gap in our internal documentation.
Conclusion
Automating 80% of IT support wasn't about replacing people — it was about redirecting human intelligence toward problems that actually need it, while letting the machine handle the predictable. If you're facing an overwhelmed support queue, start by auditing your tickets and building narrow, auditable automation — then let me know how it goes.
Follow more articles by André Dias Moreira Prol on Medium.
Top comments (0)