IT Ops Query: AI pilot to (Microsoft) Copilot: the state of enterprise AI

### Overview In this episode of *IT Ops Query*, host Beth Pariso speaks with Aaron Rich, Chief Technology and Innovation Officer at Avanade, about the current state of enterprise generative AI adoption. While consumer tools have achieved rapid adoption, enterprise adoption remains largely at an inflection point dominated by pilots and proof of concepts. Rich discusses the cultural, operational, and technical challenges businesses encounter as they transition from individual productivity tools to deep business process reinvention.

### Key Topics - **Current Adoption State:** Most enterprises are still running experiments and pilots. While "AI on the business" (productivity aids like Microsoft Copilot) is broadly in production, "AI in the business" (core process transformation) is ramping up much more slowly. - **Human Factors and Culture:** Within organizations—including tech-forward consulting firms like Avanade—employees fall into three categories: enthusiastic early adopters, an anxious middle trying to balance new tools with daily delivery pressures, and a reluctant group adhering to legacy workflows. Cognitive fatigue and the mental overhead of navigating approved tools and data policies contribute to hesitation. - **Enterprise AI vs. Consumer Pace:** Consumer tools iterate rapidly, but enterprises require rigorous evaluation of data privacy, administrative controls, and compliance before operationalizing new technologies. - **Workforce Impact and Digital Labor:** Rather than a direct one-to-one replacement of human workers, enterprise AI acts as supplemental "digital labor" (such as always-on assistants for project management). While labor displacement and role evolution will occur, engineering and core business functions are expected to adapt rather than disappear entirely.

### Technical Details - **Data Governance and Security:** Enterprises face operational hurdles regarding data loss protection (DLP), access boundaries, and privilege escalation. When users build or publish autonomous agents, systems must ensure recipients execute within their own permission levels rather than inheriting the creator's elevated privileges. - **IT vs. Non-IT Developed Agents:** Organizations are grappling with governance models for citizen-built (non-IT) agents versus centralized IT implementations. - **Multi-Model Orchestration:** Platforms like Microsoft Copilot Cowork are exploring configurations where models (such as OpenAI and Anthropic's Claude) interact prior to serving responses to users to improve output quality and mitigate hallucinations. - **Agent Orchestration and Cost Control:** Key emerging architectural best practices include establishing catalogs to promote agent and component reuse, preventing redundant builds, and setting up monitoring and throttling mechanisms to manage infrastructure consumption and model inference costs.

### Takeaways - **Interdisciplinary Governance:** Effective AI rollout requires breaking down operational silos by running joint governance workshops involving data, security, and legal teams to adapt existing policies rather than attempting to rewrite them from scratch. - **Targeting High-Impact Workflows:** Successful transformations focus on bottleneck processes that drive growth rather than just marginal efficiency. For instance, redesigning an end-to-end quote generation workflow by combining human oversight, traditional predictive machine learning models, and generative AI agents. - **Outlook:** Meaningful transformation will take time; adoption is predicted to remain mostly in pilot phases through 2026 (reaching perhaps 35% in production) before seeing larger-scale enterprise production swings around 2027. Major risks that could slow progress include potential enterprise security breaches and behavioral friction.