Podcast Summary: Using AI at Work, Episode 66: Using AI Agents at Work—Governance, Security, and Real-World Wins with Oren Michaels

Podcast: Using AI at Work: AI in the Workplace & Generative AI for Business Leaders
Host: Chris Daigle
Guest: Oren Michaels (Founder, Barn Door AI)
Air Date: August 25, 2025
Episode: 66

Overview

In this episode, Chris Daigle explores the next frontier of workplace AI with Oren Michaels, founder of Barn Door AI. The discussion dives into the emerging world of AI agents—AI systems that act rather than suggest—and tackles the major business concerns around agent deployment, notably governance and security. Daigle and Michaels provide practical frameworks, real-world examples, and cautionary tales for business leaders ready to move beyond conversational AI toward action-oriented agents that fundamentally shift productivity and operations.

Key Discussion Points & Insights

Oren Michaels’ Background and Perspective

• Oren Michaels was trained as an engineer, became a serial entrepreneur, and founded Mashery (API management) before moving into AI (02:00).
• Observed parallels between early API adoption and current AI challenges—both required a business-first, workflow-centric approach to deliver value.

What Makes an AI Agent?

• Definition: Agents don’t just suggest actions; they take them—writing, deleting, or updating data and executing tasks (08:33).
  • “An agent takes action rather than merely suggesting action.” — Oren (08:33)
• Analogy: Agents resemble “enthusiastic interns”—intelligent but lacking full context and must be given responsibility incrementally (04:13).

State of AI Agents in 2025

• Still very early (nascent) in practical deployment (10:37).
• Marketplace lacks a universal agent; instead, niche/purpose-built agents are more effective (10:37).
• Most breakthroughs will come from business problem-solvers, not just average employees experimenting (12:55).

Fears and Realities: Job Loss and Transformation

• ChatGPT and similar tools aren’t “taking jobs” on their own. Instead, AI agents change how work is done, particularly removing repetitive “soul-crushing” tasks (13:27).
• Agents carry risk due to their power—they can make dramatic changes very quickly, sometimes outside intended boundaries (04:13, 15:29).
  • “Agents don’t show up with a conscience... or a fear of being fired. If you have employees who have neither, problems will happen.” — Oren (15:29)

Governance & Security: The Central Challenge

Human vs. Agent Oversight

• Humans still make judgment calls before invoking AI (15:09).
• With agents, business leaders lose that “last check”; agents act autonomously within their permission sets (15:29).
• Strong governance frameworks are essential, modeled after—but more restrictive than—human role-based access (19:39).

Elements of an Agentic Governance Framework

• Map permissions: Mirror human role/access, but restrict agent rights further (19:46).
• Observe and adapt: Start with limited agent permissions, monitor behavior, then expand access as confidence grows (33:33, 36:45).
• “Start small. Start by saying no. See what [the agent] tries to do, and then gradually give more and more access.” — Oren (33:33)

The Model Context Protocol (MCP)

• MCP is an emerging protocol enabling AIs/agents to interact programmatically with other systems (25:21).
• MCP itself lacks native security (“there is no S in MCP”), making additional governance critical (25:21, 27:29).
  • “If you’re going to make use of MCP... you need governance or else bad things are going to happen.” — Oren (25:21)

Security Risks with Agents

• Risks most often come from well-intentioned employees using agents in unforeseen ways—not just “bad actors” (29:42).
• "Shadow usage" is rampant—where individuals experiment with agents outside formal controls, increasing risks of data leakage or prompt injection (48:55).

Where to Safely Begin with AI Agents?

• Start with low-risk, repetitive, and easily spot-checked tasks:
  • Marketing automation (campaign management, investor outreach) (20:31)
  • Repetitive finance/reporting tasks (expecting shift from “30 or 45 days to 30 or 45 minutes... then seconds” for reporting) (20:31)
  • Contract redlining (legal review)—accelerates cycles but is low-risk (20:31)
• “Look at the things that are kind of needing to get done repetitively... but that are not requiring the brain power of the executive doing it.” — Oren (20:31)

The Buying and Implementation Process

• Common buyers: Chief AI Officers (CAO), CIO, CISO, HR, and Line-of-Business leaders; group conversations split between business need and security (29:42).
• Deployment: Straightforward with established protocols (like MCP); “We drop our proxy in... and you have a customer.” (32:07)
• Enterprises (hundreds+ employees) are typical clients, since they use access management systems (Okta, Entra) suitable for agentic governance (35:41).

Real-World Wins

• Fast ROI starts with automating financial, marketing, and contract processes (41:11).
• Agents are “not here to take your jobs—they’re here to do jobs that we don’t know how to do anyway.” (41:11)

Notable Quotes & Memorable Moments

• Agents as interns:
  “We like to say that AI agents show up to work as a really enthusiastic intern.” — Oren (04:13)
• Agent discretion:
  “Agents don’t show up with a conscience... or a fear of being fired.” — Oren (15:29)
• On governance:
  “Start by saying no. See what they try to do, and then gradually give more and more access.” — Oren (33:33)
• Growth of shadow AI:
  “There is a lot of... unsanctioned MCP usage... the risk is high.” — Oren (48:55)

Timestamps for Key Segments

| Timestamp | Segment/Topic | |-----------|----------------| | 02:00 | Oren’s background and genesis of Barn Door AI | | 04:13 | Agent definition—intern analogy, risks of agent autonomy | | 08:33 | Detailed breakdown: What is an AI agent? | | 10:37 | Why we’re still early in agent adoption; purpose-built vs. universal agents | | 13:27 | Agents, automation, and the job-loss narrative | | 15:29 | Governance challenges: agent discretion, oversight, and “blast radius” | | 19:39 | How to map and limit agent permissions—governance framework | | 20:31 | Areas to safely deploy agents first (marketing, finance, contract review) | | 25:21 | Model Context Protocol (MCP): Opportunity and security void | | 29:42 | Who’s buying? What are their roles and motivations? | | 32:07 | Technical ease of deployment for governance/proxy layers | | 33:33 | Incremental governance: “Start small, start by saying no” approach | | 35:41 | Typical customer profile (size, systems, governance needs) | | 41:11 | Where are real-world ROI & wins? Financial, marketing, legal—repetitive, scalable tasks | | 48:55 | Shadow usage of agents—“unsanctioned MCP” as a real risk |

Podcast Tone & Style

The conversation is practical, direct, and business leader–focused, with the host Chris Daigle often referencing the real anxieties and learning curve of non-technical executives. Oren Michaels is congenial but candid, offering analogies and straightforward advice—emphasizing "starting small" and focusing on governance before deploying powerful new tech.

Final Takeaways & Action Items

• Embrace agents, but prioritize governance and visibility from the start.
• Start with low-risk, repetitive processes before giving agents broader authority.
• Map out permissions based on human roles—but restrict agent power further.
• Stay vigilant for “shadow AI” and MCP-based integrations happening outside official policy.
• Adopt an incremental, feedback-driven approach: trust, validate, and expand cautiously.

For more: Contact Oren Michaels at Barn Door AI or via email (oren@barndoor.ai).

“Understand the difference between your biological workforce and your agentic workforce. You manage them both—but the agentic workforce comes without a conscience.” — Oren Michaels (37:30)