About

28 years of operations. The AI came last.

Support operations and AI automation leader. Based in Orlando, Florida, originally from São Paulo. The current work is agentic AI inside live customer-facing workflows, designed, shipped and measured at 32% ticket deflection across global support volume. What makes that repeatable is not the model. It is the 28 years underneath it, and the order they happened in.

Marcio Prado in a black-and-white portrait, lit against a dark background

2003, São Paulo. Teaching people to use a computer.

The first job in technology was half teaching. From 2003 to 2008 at Future Informática in São Paulo: practical computer instruction and hands-on technical support for residential clients and small businesses. Instruction is a hard school. The student does not care what the system does, only whether they can do the thing they came to do. A system nobody will use is not a system, it is an invoice. Every rollout since has started there.

Then the auditors and the regulator showed up.

Flight training is a regulated business, and in December 2009 that became my problem. For 9 years at EJ Escola de Aeronáutica, the remit was IT governance, portfolio delivery, supplier management, financial management and service operations. The work was rebuilding how technology got prioritized and delivered: Scrum, Kanban and SAFe, under a federated IT governance model. Alongside it, internal audit support, enterprise risk management, SoX controls and civil-aviation regulatory compliance. Plus the procurement lifecycle for flight simulators, where the vendor relationship is the uptime. Under a regulator, the control is in the design or the thing does not fly.

Then a product, alone, with nobody to escalate to.

Running an IT organization for 9 years leaves one gap: portfolios are not products. From March 2019 to August 2022, self-employed out of São Paulo, that gap closed. Clear Sky Aviation English went from concept to production with me as sole product and technology owner: discovery, implementation, vendor management, then service operations. The same period produced Germinar RH · OBPI, a digital workflow for the psychological-assessment experience used in commercial-pilot evaluation. Working alone removes the distance between a decision and its consequence. The bug, the invoice, the support request: one desk.

September 2022, Orlando. Back into the queue on purpose.

The seat in Orlando was technical support specialist at Kaseya. Advanced support for Datto EDR, Datto AV and RocketCyber: deployment, alerting, endpoint protection and RMM integration for MSP partners. After 9 years running an IT organization and 3 years owning a product, that reads on paper like a step backward. It bought the one thing a senior hire never gets: the queue itself, before anybody had tidied it into a slide. That seat is where I learned how the product actually fails, and how a partner describes it when it does. Promoted into support leadership within 16 months.

Leadership, at the escalation end.

From January 2024 to August 2025 the title was Manager, Security Technical Support. Datto EDR and AV across enterprise and MSP environments worldwide, coaching L2 and L3 specialists, senior escalation point on high-impact customer situations. Major incidents, escalations and root-cause analysis owned end to end, coordinated with Product and Engineering, stakeholders aligned while it was on fire. Endpoint policy work alongside it: Datto EDR/AV, Microsoft Defender, Datto RMM, Azure Security Center. The quieter half matters more: knowledge-base content, repeatable escalation practice, automation workflows, frontline issue patterns turned into product fixes. That is automation by hand. It works. It does not scale. An agent should be doing it.

Deflection is a workflow question, a knowledge question, a handoff question and a security question. The model is the easy part.

August 2025. Owning the AI.

Manager of Artificial Intelligence is the newest line here, and the shortest. Three roles, two promotions, under 3 years, all inside the same organization. The remit is agentic AI strategy and delivery across global support and billing operations, owned from stakeholder discovery to continuous improvement. In practice: LLM-powered virtual agents built with Ada and Mosaic AI inside Zendesk-based workflows, designed, shipped and optimized to 32% ticket deflection across global support volume. Then the instrumentation. MCP server integrations connecting the Mosaic API and Ada to read what the agents actually said, find the failure patterns and close the loop. Then the step most programs never reach: the security and compliance case that made enterprise adoption possible, guardrails included, written by me.

What I actually do

AI delivery ownership, end to end

Owning an agentic AI program from stakeholder discovery through workflow design, rollout, evaluation and continuous improvement, with one person accountable for the number at the end.

First 30 days: read the tickets and the transcripts, name the workflows where deflection is genuinely available, and write the evaluation criteria before anyone writes a prompt.

Agent program rescue

For the pilot that performs in the demo and stalls in production. The model is usually fine. The workflow around it is not.

First 30 days: pull the conversations, sort the failures into retrieval, workflow, handoff and policy, and say plainly which are engineering problems and which are operations problems.

Getting it past security review

Responsible-AI enablement, written for the people who approve it: what the agent may touch, what it may say, where it must hand off.

First 30 days: map the data the agent can reach, then draft the guardrails and the handoff rules. The output is a security and compliance case, in a form reviewers can mark up.

Technical leadership for support operations

Support operations leadership: escalation practice, major incident response, root-cause analysis, L2 and L3 coaching, and the knowledge base every future agent will be reading.

First 30 days: shadow live escalations, rebuild the escalation path, and audit the knowledge base, because an agent cannot be better than what it stands on.

Credentials

Certifications: Microsoft Azure AI-900, AI Fundamentals · Microsoft SC-900, Security, Compliance & Identity Fundamentals
Languages: English, Portuguese and Spanish. Fluent in all 3.
Based: Orlando, Florida, USA
From: São Paulo, Brazil
Domains: Agentic AI delivery · LLM evaluation · support operations · cybersecurity support and endpoint detection and response · incident response and root-cause analysis · service governance · knowledge management · vendor and partner management
Tools: Zendesk · Ada · Mosaic AI · Model Context Protocol · Datto EDR/AV · Datto RMM · RocketCyber · Microsoft Defender · Azure Security Center
Practices: ITIL · Scrum · Kanban · SAFe · SoX controls
Experience: 28+ years across enterprise technology operations, cybersecurity support, service governance and digital delivery

Written in a personal capacity. Nothing here is an official position of any employer, past or present.

Speaking and writing

Topics, formats and press kit

Why aviation keeps showing up

Aviation is the thread I never planned. The record reads: 9 years of IT inside a flight school, simulators on the procurement ledger, an English platform for pilots, an assessment workflow used in commercial-pilot evaluation. What aviation hands you is not romance, it is an attitude toward failure. Incidents get investigated, not blamed. Procedure gets rewritten in the language of what went wrong. Checklists exist because competent people forget things under load. Every guardrail I write traces back to that. An agent in a customer-facing workflow is a system under load, in front of somebody with a real problem. Design for the bad day.

If this sounds like your problem

A pilot that works in the demo and stalls in production. A support organization told to adopt AI and given no reason to trust it. Both are operations problems in a model costume. Send the shape of yours and you get a straight answer, including the unwelcome one: that it is not the problem you think it is.

Start a conversation See how the rollout worked

Contact page linkedin.com/in/marciolprado