Trusting the Agent: When AI Becomes an Autonomous Coworker

Subtitle: A Practical Microsoft AI Framework for Identity, Autonomy, Human Oversight and Enterprise Trust

Session Abstract:

Enterprise AI is crossing an important boundary: from systems that answer questions and generate recommendations to AI agents capable of reasoning, making decisions and taking actions across business systems.

That transition creates a fundamentally different trust problem.

What happens when an AI agent can access sensitive enterprise knowledge, initiate workflows, interact with operational systems or make decisions that affect people and processes? And how do we determine what an agent should be allowed to do autonomously—and when a human must remain in the loop?

In this session, Srishti Jha introduces the Agent Trust Stack™, a practical framework for designing and deploying trustworthy enterprise AI agents across eight layers: Identity, Context, Permission, Reasoning, Action, Oversight, Auditability and Human Trust.

Using an industrial operations scenario, the session will demonstrate how these principles can be applied when designing an enterprise agent using the Microsoft AI ecosystem, including Microsoft Foundry, Azure AI services and enterprise identity, security and governance patterns.

Attendees will explore how to move beyond the AI prototype and address the harder questions that emerge when agents enter real operational environments:

How should an enterprise agent establish identity and permissions?
What data and organizational context should it be allowed to access?
Which decisions can safely be delegated to an agent?
Where should human approval gates exist?
How do we observe and audit agent actions?
How do organizations create employee trust without encouraging blind trust in AI?

The session connects agent architecture, AI governance, operational risk and human adoption to provide a practical approach for moving enterprise AI agents from experimentation toward trusted production.

Attendees will leave with an Agent Trust Stack™ reference framework they can use to evaluate and design agentic AI solutions within their own organizations.

What attendees will learn
How agentic AI changes the enterprise trust model — and why governance designed for copilots and chatbots is insufficient when AI systems can take actions.
How to apply the Agent Trust Stack™ across identity, context, permissions, reasoning, action, human oversight, auditability and organizational trust.
How to architect human-in-the-loop controls based on the consequence and reversibility of an agent's actions.
How Microsoft AI technologies can support trustworthy agent architectures, from enterprise knowledge grounding to identity, security, monitoring and governance.
How to evaluate whether an AI agent is ready to move from prototype to production.
Demonstration Scenario

The session uses an EV fleet/battery operations agent as a practical example.

The agent receives battery-health and operational information and identifies an abnormal degradation pattern.

Rather than simply generating an answer, the agent must determine:

Observe → Investigate → Retrieve Context → Reason → Recommend → Request Approval → Act → Audit

The demonstration examines what happens at each stage and where identity, authorization, grounding, human oversight and auditability need to be introduced.

The scenario illustrates a broader principle:

The more authority we give an AI agent, the stronger its trust architecture must become.

Intended Audience

AI developers, architects, technical leaders, product managers, technology executives, governance and security professionals, and organizations moving Microsoft AI solutions from experimentation toward enterprise production.

Level: Intermediate

Format: 45-minute conference session

Primary Topics: Microsoft Foundry / Azure AI • Agentic AI • Responsible AI • Governance & Security • Enterprise AI

Speaker Positioning

Srishti Jha is a technical program and transformation leader whose career spans EV engineering, battery R&D, large-scale manufacturing programs, battery intelligence and enterprise AI adoption across North America, Europe and Asia.

Her work sits at the intersection of complex operational systems, emerging AI technologies and organizational transformation. Drawing on experience leading high-value industrial programs and AI-enabled initiatives, she focuses on a central challenge facing enterprises adopting agentic AI:

How do we give AI systems greater autonomy without giving up accountability?

Her current work and community contributions explore trusted AI adoption, agentic AI governance and the transition from AI experimentation to operational deployment.

Security That Ships: Cloud-Native Security Lessons from Real AI & Industrial Deployments

Session Type

20-minute Lightning Talk

Description

Cloud-native security conversations often focus on tools and policies. In production environments supporting AI platforms, industrial data pipelines, and distributed operational systems, security becomes an architectural and delivery challenge that must coexist with uptime, scale, and cross-team execution.

This session shares practical implementation lessons from deploying and governing containerized workloads in environments where system reliability and data integrity directly impact operational outcomes. Drawing on real deployment experiences and case-informed scenarios, the talk highlights how security decisions were embedded into architecture rather than added as afterthoughts.

Examples explored in the session include:

Securing containerized analytics workloads processing operational manufacturing data
Designing access boundaries, secret handling, and pipeline controls to prevent data leakage while maintaining continuous delivery velocity.

Governance safeguards around AI-enabled decision-support services
Implementing runtime visibility and auditability when model-driven services interact with production workflows and enterprise data sources.

Identity and privilege segmentation across distributed teams and environments
Managing access across platform, engineering, and vendor stakeholders without introducing delivery bottlenecks.

Observability-led detection of abnormal service behavior
Using telemetry and monitoring signals to identify configuration drift or unexpected runtime patterns within microservice ecosystems.

Integrating security checks into CI/CD pipelines
Enforcing baseline controls while avoiding developer friction and deployment slowdowns.

Rather than presenting tooling comparisons, this session focuses on architectural patterns, tradeoff decisions, and organizational realities encountered when implementing cloud-native security practices in complex operational settings.

Attendees will leave with actionable insights on how to:

Translate governance requirements into enforceable runtime controls

Align security with delivery velocity

Build visibility into dynamic container environments

Navigate cross-functional adoption challenges

Treat cloud-native security as a systems design discipline

This talk is designed for practitioners and technical leaders responsible for deploying real systems under real constraints seeking to bridge the gap between security theory and operational implementation.

Trusting the Agent: From AI Assistant to Autonomous Coworker

A Practical Microsoft AI Framework for Identity, Autonomy, Human Oversight and Enterprise Trust

Session Abstract

Enterprise AI is crossing an important boundary: from systems that answer questions and generate recommendations to AI agents capable of reasoning, making decisions and taking actions across business systems.

That transition creates a fundamentally different trust problem.

What happens when an AI agent can access sensitive enterprise knowledge, initiate workflows, interact with operational systems or make decisions that affect people and processes? And how do we determine what an agent should be allowed to do autonomously—and when a human must remain in the loop?

In this session, Srishti Jha introduces the Agent Trust Stack™, a practical framework for designing and deploying trustworthy enterprise AI agents across eight layers: Identity, Context, Permission, Reasoning, Action, Oversight, Auditability and Human Trust.

Using an industrial operations scenario, the session will demonstrate how these principles can be applied when designing an enterprise agent using the Microsoft AI ecosystem, including Microsoft Foundry, Azure AI services and enterprise identity, security and governance patterns.

Attendees will explore how to move beyond the AI prototype and address the harder questions that emerge when agents enter real operational environments

Intended Audience

AI developers, architects, technical leaders, product managers, technology executives, governance and security professionals, and organizations moving Microsoft AI solutions from experimentation toward enterprise production.

Level: Intermediate

Format: 45-minute conference session

Primary Topics: Microsoft Foundry / Azure AI • Agentic AI • Responsible AI • Governance & Security • Enterprise AI