Enterprise AI is being pulled in two directions at once: open, interoperable standards on one side, and deep, proprietary agent platforms on the other. Neither side is winning outright, which means the architecture decision — not the model choice — is where the real risk sits.
Open Standards: Betting Against Lock-In
Efforts like the Linux Foundation's Open Platform for Enterprise AI (OPEA) push back against vendor lock-in by standardizing how models, retrieval systems and agent frameworks talk to each other. As enterprises move from single chatbots to multi-step agentic workflows, the lack of interoperability between models becomes a real bottleneck — every proprietary integration is another thing that breaks when you want to swap a model or a vendor.
For Irish CTOs, this is an architectural decision point: are you building on a closed API or an open standard? Leaning toward open frameworks now can save months of refactoring later, particularly once regulatory requirements demand a level of model transparency that a fully proprietary "black box" can't easily provide.
Vendor Agent Platforms: Depth Over Openness
At the other end, platforms like Salesforce's Agentforce go deep instead of open — autonomous agents wired directly into a specific vendor's data model, effectively arguing that the "copilot" era (human-in-the-loop, one suggestion at a time) is giving way to an "agent" era (human-on-the-loop, reviewing outcomes instead of every step).
That's a thesis worth taking seriously: real value increasingly lies in execution, not just conversation. But a vendor's agent platform is also a walled garden for that vendor's own data. A pragmatic middle path is to use a vendor's native agent layer for tasks that live entirely inside that vendor's data (CRM-specific automation, for example), and orchestrate anything that spans multiple systems — ERP, legacy databases, other SaaS tools — through your own vendor-agnostic control plane.
Key Insight
Don't lock your entire business logic into a SaaS provider's AI agent. Keep your "agentic core" — the part that decides what happens next — independent of any single vendor, so you keep the option to change platforms later.
The Governance Gap Nobody's Closing
Industry surveys (KPMG among them) have repeatedly flagged a growing governance gap: teams rush AI features to production and treat governance as a follow-up task rather than a launch requirement. That gap doesn't close itself, and the regulatory runway that made it feel low-priority is shorter than it looks — see our EU AI Act update for the current compliance timeline.
Governance can't be a paper exercise. It has to be implemented at the code level — semantic routers and guardrails that programmatically enforce policy before an agent ever executes a tool call, not a review meeting that happens after it already has.
Conclusion
Whether you're adopting open standards or building on a vendor's agent platform, the discipline required is the same: keep the decision-making core of your system portable, and treat governance as part of the architecture, not a compliance checkbox added at the end.