Aaron believes that organizations succeeding with AI today aren't relying solely on a single model or vendor; instead, they are developing a diverse portfolio. While a single-platform approach may seem efficient (requiring only one contract, integration, and a team to train), it becomes a vulnerability if the top-performing model shifts, new regulations limit the use of vendor infrastructure, or the vendor's development roadmap no longer aligns with your needs.

Aaron's experience spans both perspectives: as a CTO making long-term infrastructure choices and as an open-source maintainer witnessing how quickly the leading tools in a category evolve. He advocates for designing flexibility from the start by standardizing interfaces and governance layers rather than specific models, enabling easy provider switches as configuration changes rather than complete re-architectures.

He notes that the organizations that succeed treat their AI infrastructure like a prudent investor manages a portfolio: diversified, regularly rebalanced, and not over-committed to a single option, ,and helps tech leaders in regulated industries build AI strategies that stay flexible and resilient.

Most enterprises attempt to achieve perfect AI governance by waiting for foolproof policies, complete visibility, and zero risk before any team begins using the technology. 

For IT leaders in regulated organizations, Aaron suggests that this isn't cautious but potentially destructive to the organization. While one team spends two quarters developing a detailed 40-page AI usage policy, a competitor's team is already three iterations into a workflow that saves actual hours and collects behavioral data, which could inform the first team's policy.

Although it's understandable to restrict everything up front in regulated industries, it reverses the natural order: you can't govern behaviors you haven't observed yet, and you can't observe them until you start, which is delayed by waiting for approval.

Aaron explains what "shipping with the controls you have" entails: focusing on narrow scopes, maintaining quick feedback cycles, and adapting governance based on real usage data, rather than attempting to pre-empt every failure mode from the start.

Most organizations in regulated industries face a common challenge today:  Most AI observability mainly produces dashboards that improve visibility but still lack actionable steps. Teams may notice increased usage, flagged prompts, or drifted model outputs from last week, but what next?

Aaron highlights the gap between monitoring and governance: monitoring shows what happened, while governance influences future actions. For CTOs and IT leaders, this requires systems that do more than log behavior; they must react automatically: addressing drift before it impacts users, auto-flagging policy breaches instead of waiting for weekly reports, and creating audit trails that regulators require without manual reconstruction. 

This distinction is critical in regulated settings, where "we have a dashboard" isn’t enough during an audit. What truly matters is proof of real-time detection and correction, not just the ability to spot issues in retrospect.

Many blame failed enterprise AI initiatives on weak models, but Aaron believes most AI architectures fail because they were never built to scale. For CTOs and IT leaders, this challenge is even greater, where security, compliance, and governance cannot be compromised.

Many organizations start with a pilot, a chatbot, or a single use case that works well in isolation. But once AI adoption expands across teams, systems, and workflows, the cracks start to show. Many rely on single-model setups, disconnected point solutions, or manual workflow builders that cannot handle the real complexity of enterprise environments, especially in highly regulated environments.

Aaron thinks the biggest issue is that enterprise environments are inherently messy. Teams operate across CRMs, Slack, Google Drive, SharePoint, ERPs, and countless other systems, and AI lives across all these environments. When organizations try to force AI into rigid workflows or rely on a single model provider, complexity grows quickly. This is why so many AI initiatives stall after the pilot stage. What looks simple in a controlled environment becomes difficult to secure, govern, and manage in production.

For Aaron, the solution is not to add more tools or give teams more AI building blocks. It’s a better architecture. He argues that enterprise AI needs to be built for flexibility from day one. That means model-agnostic orchestration, strong governance, secure data handling, and systems that reduce complexity rather than expose it. The goal is to help IT leaders deploy AI safely at scale while maintaining control, compliance, and performance.

A clear example of this is Liminal’s Behavioral Agent Automation Platform (BAAP). Aaron notes that instead of forcing teams to manually build and maintain workflows, the platform observes how work actually happens across systems and learns from real behavior. It routes tasks to the right models, securely connects to enterprise data, and reduces the burden on IT teams.

Aaron believes one of the biggest flaws in enterprise AI today is the assumption that workflows must be manually designed before automation can happen. Most organizations still rely on IT teams, consultants, or low-code tools to map processes, define logic, and build workflows from scratch. Aaron thinks this approach is too slow, too expensive, and too difficult to scale.

Aaron notes that the real problem is that work rarely follows a perfectly documented process. Enterprise workflows are messy, dynamic, and constantly changing across systems like CRMs, Slack, Google Drive, and ERPs. He feels this is why so many automation projects stall. By the time a workflow is mapped and built, the business has already changed.

For Aaron, the solution is moving beyond static workflow design. He believes AI systems should be able to observe how people actually work, identify patterns, and learn where automation can create value. Instead of asking teams to manually build every workflow, the system should gradually understand behavior and adapt over time.

Aaron helps CTOs and IT leaders move beyond rigid workflow design by rethinking how automation is built. He believes the future lies in systems that learn from real human-AI interactions, uncover friction points, and automatically identify repeatable tasks. Aaron believes this is how regulated organizations can scale automation faster without overwhelming IT teams.