Discovery earns the build
If the discovery does not change what gets built, it was not a discovery — it was a demo with stakeholders in the room.
Support Operations & AI Automation Leader
Agentic AI that survives contact with a live support floor.
Marcio Prado shipped LLM-powered virtual agents into live global support workflows and measured 32% ticket deflection. He came up through that same support organization first: major incidents, escalations, root-cause analysis, endpoint security across enterprise and MSP environments worldwide.
- 32% ticket deflection across global support volume
- 2 promotions in under 3 years, same support organization
- 28+ years in enterprise technology operations
Career
The order he learned it in.
Nothing here started with AI. It started with a room of people who had never used a computer, and it moved through governance under aviation regulators, sole ownership of a product, and years on a security support queue before an agent was ever scoped. The instincts are decades old. Only the vocabulary changed.
-
Computer Instructor & Technical Support Specialist
Future Informática São Paulo, Brazil
Taught people who had never touched a computer, then repaired the machines when they broke. Homes and small businesses, no abstraction layer between the work and the person paying for it. Every role since is that job at larger scale.
-
IT Supervisor (IT Office Manager)
EJ Escola de Aeronáutica / Flight Training Center São Paulo, Brazil
Ran IT for a regulated aviation-training organization: federated governance, multi-year delivery through Scrum, Kanban and SAFe, and the procurement lifecycle for flight simulators. Internal audit, enterprise risk management, SoX controls and civil-aviation compliance all wanted evidence. He produced it.
-
IT Consultant / E-Learning Platform Manager
Independent · Clear Sky Aviation English São Paulo, Brazil (remote)
Sole product and technology owner of Clear Sky Aviation English, concept to production: discovery, implementation, vendors, reliability, cost.
- Clear Sky Aviation EnglishAviation-English e-learning taken from first conversation to running service.
- Germinar RH · OBPIDigital workflow for the psychological assessments used in commercial-pilot evaluation.
-
Technical Support Specialist
Kaseya Orlando, Florida
Advanced support for Datto EDR, Datto AV and RocketCyber: deployments, alerting, endpoint protection and RMM integration failures for MSP partners. The queue itself, worked ticket by ticket. Promoted into support leadership within 16 months.
-
Manager, Security Technical Support
Kaseya Orlando, Florida
Led security technical support for Datto EDR/AV across enterprise and MSP environments worldwide. Coached L2/L3 specialists, took the senior escalation seat on high-impact situations, and owned major incident response and root-cause analysis end to end with product and engineering.
-
Manager of Artificial Intelligence
Kaseya Orlando, Florida
Owns agentic AI across global support and billing operations, discovery through evaluation. Shipped the virtual agents that reached 32% ticket deflection, built the MCP integrations that read their conversations, and wrote the security and compliance case behind the rollout.
32%
Ticket deflection
Measured across global support volume, from virtual agents running inside live Zendesk workflows. Not a controlled pilot.
16 months
Queue to leadership
Time from working the frontline queue as a technical support specialist to being promoted into support leadership.
2
Promotions in under 3 years
Technical support specialist, then security technical support leadership, then artificial intelligence. Same organization, same customers.
28+ years
In technology operations
Enterprise technology operations, cybersecurity support, service governance and digital transformation. The AI title is the newest part of it; the operations are not.
Every number on this site traces to a dated role. Nothing here is modelled, projected or rounded up.
Signature case · Agentic support operations
32% ticket deflection, in a live global support operation.
Repeat questions and hard escalations compete for the same queue. Every support organization eventually pays for that.
The answerable tickets absorb the capacity the difficult ones need, and the people best equipped to handle a major incident spend their day on password resets and deployment questions.
The decision that mattered was sequence. Discovery ran first, with support, product, engineering, security and business leaders — the people who own the workflow and the people who work it. Agent scope was drawn against real ticket patterns, with explicit handoff points where a human takes the conversation back. The model came last.
What got built: LLM-powered virtual agents using Ada and Mosaic AI, wired into Zendesk-based workflows across global support and billing operations. Then the part most programs skip. MCP server integrations connect the Mosaic API and Ada to read what the agents actually said, surface failure patterns, and feed the fixes back into the workflow.
It held because the security and compliance case was written for the rollout rather than after it. That document is what made enterprise adoption possible, and the guardrails on customer-facing use came out of the same work.
Method
Six positions he will defend.
None of these are model problems. They are the reasons agent programs stall between a working demo and a working queue.
Guardrails ship with it
Security review is part of the build, not the last gate before launch; anything approved after the fact was approved without evidence.
Deflection is a proxy
An agent that closes a ticket the customer needed a human for has not deflected anything, it has deferred it — which is why the number is only trustworthy next to what the conversations show.
Instrument or it decays
Agent quality drifts the moment the product changes, and without conversation-level evaluation you find out from the escalation queue instead of from your own reporting.
The floor has a veto
If the specialists working the queue would not use it, the rollout is already dead, whatever the launch dashboard says.
Legible beats clever
Automation has to be explainable by the specialist defending it on a live escalation, so a workflow anyone can read beats a model nobody can account for.
What he is hired for
Six areas, and what comes out of each.
Engagements end with artifacts a team can keep using. The list under each area is what actually gets handed over.
-
Agentic AI in customer-facing workflows
Virtual agents designed, shipped and optimized inside the tooling support already runs on, owned from stakeholder discovery through rollout.
Workflow maps · Agent scope and containment rules · Human handoff design · Rollout plan and cutover sequence · Deflection reporting
-
AI evaluation and continuous improvement
The plumbing that reads what agents actually said, names the failure patterns and turns them into a change queue rather than an opinion.
MCP server integrations · Conversation analysis pipeline · Failure-pattern taxonomy · Evaluation sets · Improvement backlog with owners
-
Support operations and major incidents
Escalation paths, incident command and root-cause analysis run end to end, with product and engineering pulled in before the customer asks twice.
Escalation practices · Major-incident response and communication cadence · Root-cause write-ups · L2/L3 coaching plans · Knowledge-base content
-
Security and compliance for AI adoption
The written case that gets customer-facing AI approved, plus the guardrails it has to run inside once it is live.
Security and compliance case · Guardrail and data-handling boundaries · Responsible-use rules for customer-facing agents · Review-ready documentation
-
Service governance and IT portfolio
Decision rights, prioritization and supplier discipline for technology estates that regulators and auditors will eventually read.
Federated governance model · Prioritization and accountability practices · Vendor and procurement lifecycle · Audit-support documentation
-
Knowledge and enablement
Content built so both the customer and the agent can answer the question, because a deflection program is a knowledge program wearing new clothes.
Knowledge-base architecture · Article standards · Self-service content sets · Specialist enablement · Delivery in English, Portuguese and Spanish
Selected work
Three that explain the rest.
Twenty-eight years compresses into three projects: one in production now, one built alone end to end, and one that ran under auditors.
Agentic support operations
LLM-powered virtual agents shipped into live Zendesk workflows, then instrumented so the result holds after launch week.
32%
ticket deflection
Clear Sky Aviation English
An aviation-English e-learning platform taken from first conversation to running service by a single owner.
Concept
to production
IT governance for a regulated flight school
Governance, delivery and procurement inside a regulated flight-training organization, under internal audit and civil-aviation compliance.
9 years
of governance
Beyond the work
Aviation taught the method before support did.
Nearly a decade inside a flight-training organization leaves 3 habits behind. Each one shows up in how the AI work gets run.
Checklists beat memory
Aviation does not trust recall under pressure, and neither should an AI rollout. Guardrails are written down before go-live because a competent operator remembering the edge case is not a control.
Simulators before passengers
A simulator exists so failure costs nothing. Evaluation sets do the same job for agents: break the workflow in a room where no customer is watching, then decide what ships.
Every incident gets a report
Aviation improved by investigating what went wrong and publishing it. Root-cause analysis is that habit with worse paperwork, and the pattern across incidents is always worth more than the story of any one of them.
Words
No testimonials yet, and none invented.
This site is self-attested today. The work is verifiable through the people who were in the room, but the quotes are not written, and putting words in someone’s mouth would cost more credibility than it buys.
Named references
Worked with him on a rollout, an escalation or a governance review? A short quote and your name is worth more here than any adjective he could write about himself. Reach him at the contact page.
Speaking
Available for talks and workshops on shipping agentic AI into live support operations: discovery, guardrails, rollout, and the evaluation loop that keeps deflection honest. In English, Portuguese or Spanish.
Writing and press
Writing on agentic AI in support operations is published here first. For interviews or commentary, the fastest route is email.
Contact
The pilot works. The rollout is the hard part.
Most agentic AI programs stall in the same 3 places: a workflow nobody validated with the people who work the queue, a security review nobody wrote the case for, and an evaluation loop nobody built, so quality drifts unnoticed. None of those are model problems. They are operations problems, and they are the ones worth a conversation.
- You have a virtual agent that performs in demos and stalls on the way to production.
- Security review is where your AI proposals die, and nobody has written the case that would get one approved.
- Deflection looked good at launch and has been drifting since, and no one can tell you which conversations are failing.
Describe the stall in plain language and you will get a straight answer about whether this is the right help — including when it is not. Orlando, Florida. Originally São Paulo, Brazil. Working in English, Portuguese and Spanish.