AI Engineer Summit 2025

Privacy First Enterprise AI: Building AI Agents that Never Leave Your Security Boundary

Loading slides…

Read the talk

Privacy First Enterprise AI: Building AI Agents that Never Leave Your Security Boundary

Enterprise agents can use familiar channels and established controls. Steven Moon’s architecture starts with private compute, managed identities and email rather than another portal.

From a talk by Steven Moon

Another agent, another portal

How do you deploy an AI agent without making it another external system to manage? For Steven Moon, founder of Aech AI, the starting point is an agent that operates entirely within an enterprise’s existing security boundary. Model capability and prompt quality are only part of that problem: deployment must also respect security, compliance and workflows that organizations have spent years establishing.

Language models introduce a useful possibility: software can understand requests and context, then interact through the same interfaces people use. Yet a new enterprise agent often arrives with another portal, another set of credentials and another security review. That adds a barrier between employees and the capability the agent provides, even when employees already have a suitable place to request the work.

Slide listing external systems, portals, credentials, and barriers, alongside the proposal to build applications that work the way humans do.Open full source frame
The AI Engineering Challenge: more external systems, portals, credentials, and barriers.

Moon connects this mismatch to Satya Nadella’s forecast about business software, paraphrasing it as the death of SaaS. His interpretation concerns traditional SaaS interfaces: if agents become the primary way people interact with business systems, building a fresh dashboard for every agent reproduces the interface pattern those agents could replace. This is a forecast about how work reaches software, rather than evidence that business systems themselves have disappeared.

0:040:22
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:04 · section reference included

Give agents employee-like boundaries

An agent should operate under the same kinds of constraints as an employee. It should follow security policies, use approved systems, stay within data boundaries, access only what its work requires, and remain subject to monitoring and audit. Treating the agent as a digital worker makes those requirements part of its deployment, rather than features to add after it becomes useful.

The infrastructure for those constraints already exists in many enterprises. Moon identifies secure compute, identity management, data governance, compliance frameworks and audit capabilities, then proposes running agents in private clouds while keeping data within tenants.

Existing foundationRole in agent deployment
Secure compute and private cloudPlace agent execution inside the boundary
Identity managementEstablish who the agent is
Data governance and security policiesConstrain what it can access
Audit and monitoringRetain oversight of its activity

Moon says most enterprises have a private cloud suitable for this approach. The architectural requirement is to identify the organization’s actual approved execution environment and preserve its controls and workflows.

This changes the first interface-design question. Before creating another portal, ask whether the capability can be delivered through a system users already know and trust. If software can understand a request directly, an additional interface may be an unnecessary translation step between the employee’s intent and the work.

2:012:14
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

2:01 · section reference included

Provision agents through familiar IT tools

Aech AI’s chosen foundation is Microsoft 365 and Azure. Their value here goes beyond email delivery and cloud compute: enterprises have already integrated these platforms into security and compliance processes. Building there lets an agent reuse established infrastructure and administrative relationships. It does not automatically make the resulting application compliant; customers still own decisions about data, identities, configuration and access. That distinction is explicit in Microsoft’s current shared-responsibility guidance.

Moon invokes Jensen Huang’s CES discussion of agentic AI, describing IT as the future HR department for agents. The practical version is a familiar administrative sequence:

  1. Create an agent account using the organization’s existing Active Directory tools.
  2. Apply standard security policies.
  3. Assign the permissions needed for its work through familiar administration interfaces.
  4. Use existing audit and monitoring tools to supervise its activity.

The employee analogy describes how IT manages the agent; it does not determine its effective authorization. For a Microsoft 365 implementation, the distinction between delegated and application access matters: delegated access is bounded by the signed-in user’s access, whereas application permissions operate without that user. The current Microsoft Graph permissions model documents this distinction; Moon does not identify Graph as Aech AI’s implementation. These current security details clarify the deployment requirements without establishing the configuration used in the talk.

The resulting lifecycle is recognizable to IT: onboarding, access assignment, permission management and ongoing monitoring. Moon’s proposal is to manage the AI workforce through those established practices, avoiding a separate administrative system for every new agent.

3:203:33
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:20 · section reference included

Use email for agent-to-agent coordination

Email can carry more than a human’s request to an agent. Agents can email one another to share information and coordinate work, using the same communication infrastructure that employees use to collaborate. The channel therefore becomes part of the coordination architecture, with messages available for inspection through established enterprise systems.

Moon describes every interaction as logged and auditable, with permissions enforced automatically and data flows kept transparent and controllable. Email integration alone does not establish complete agent tracing. Microsoft’s mailbox-auditing documentation describes coverage for selected actions, with configuration, bypass and licensing considerations. The useful engineering distinction is between retaining evidence of mailbox activity and recording the agent’s complete execution. Existing permission systems likewise enforce the access actually configured for the agent.

Aech AI chose Microsoft’s ecosystem, but Moon proposes the same pattern for Google Workspace and other enterprise platforms. The reusable architectural ingredients are built-in identity management, established security controls, compliance frameworks and enterprise APIs. Porting the pattern means finding those foundations in the target platform and configuring them to support observable, controllable collaboration.

4:404:48
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:40 · section reference included

Find the work before building the interface

Reusing enterprise infrastructure leaves more engineering effort for new capabilities and the problems they solve. Aech AI chose email because it is universal and trusted, but email is only one entry point. The broader approach is to enhance systems the organization has already spent years securing and refining.

Document management systems, internal messaging platforms and workflow tools are other candidates. Once an agent can understand human intent, each can become a gateway to an AI capability. The design exercise begins with where work already happens: which systems are available, what employees do there, and how an agent could enhance that work. Only then does the question of building a new system arise.

The result may be quiet intelligence inside the tools customers already use every day. Moon closes with the prospect of direct understanding reducing the need for mandatory translation layers between humans and machines. The architectural opportunity is to let that understanding reach employees through trusted systems, while keeping the agent’s execution, access and oversight within the enterprise’s established boundaries.

Slide titled The Path Forward with four colored questions: What systems do we have? Where is work happening? How can we enhance it? Why build something new?Open full source frame
The Path Forward: examine existing systems, locate work, and ask how to enhance it.
5:385:50
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:38 · section reference included

Resources

From the talk

  • Agent AechDocumentation0:04

    Aech AI's current product overview describes agent workflows through email, shared inboxes and Microsoft Teams.

Updates since the talk

  • Current Azure guidance explains which security responsibilities remain with customers across cloud deployment models.

Transcript download
Full transcript

Loading transcript…