主なポイント
- Data and integration teams still largely operate as separate groups — an approach that breaks down once agentic AI needs fast, cross-functional collaboration.
- Part Two of the O’Reilly report Beyond Pipelines: The New Data Operating Model for AI at Scale, sponsored by Boomi, looks at how roles, accountability, and governance need to evolve to support autonomous agents.
- “Guardrails, not roadblocks” is the report’s core governance principle for letting agents operate autonomously within approved limits.
We’ve already talked about the fact that agentic AI can’t run on fragmented data, and that fixing it requires an AI-ready data foundation built on centralization, live state, and embedded intelligence. But a great data foundation is only half the story. Plenty of organizations will finish that work, deploy their first few agents, and then hit a second wall: a team structure and approval processes built for a world without autonomous systems.
A pilot agent may work beautifully in a demo, then stall in production because no single team is accountable when it misfires. Or an agent is technically capable of resolving an issue from end to end, but three separate approval steps mean a human (or three) ends up doing the work anyway, it just takes longer. The data foundation didn’t fail — the operating model around it did.
Part two of “Beyond Pipelines: The New Data Operating Model for AI at Scale,” written by data strategist Adam Morton and sponsored by Boomi, tackles how roles need to shift, and how governance can become an enabler instead of a bottleneck. If you’re rethinking how your teams collaborate around AI, or wondering how much autonomy to hand your agents, it’s worth a read.
Governance as an Enabler
Guardrails shouldn’t be roadblocks. The idea is to create consistent boundaries that let agents act freely within approved limits, instead of routing every decision through a queue that requires manual approval. Defining what agents can access, what they’re allowed to do, and what they should optimize for can be done in layers, so autonomy can expand without risk expanding along with it. The report shares examples of what that looks like in practice, including ways to govern AI agents at scale without slowing them down.
The shift is subtle but important: instead of asking a person to approve every refund, every re-route, and/or every price change, you define the limits once — refunds under a certain amount, re-routes within a cost threshold — and let the agent operate freely inside those limits. Anything that doesn’t fit within the boundaries still escalates to a person.
That’s a very different governance posture than most organizations have today, and it’s usually the harder part of an agentic AI rollout to get right, because it means business stakeholders and technical teams have to agree on where the limits sit well before the agent goes live.
Guardrails: Where to Start
One of the biggest promises of AI is the ability to move faster. So if you already have agents deployed, or are getting ready to deploy them, we aren’t suggesting you have to stop everything and redesign your organizational structure or governance rules. A more realistic starting point looks like this:
- Pick one live (or planned) agent and name a single team accountable for its outcome, not just its underlying components.
- Write down the approval steps that agent currently has to go through and determine which steps are protecting against real risk.
- Turn one of those approval steps into a guardrail — a defined limit the agent can operate within on its own — and measure what changes.
This is a small step compared to the full operating model the report describes, but it may be enough to show stakeholders what’s possible and open the door for a broader organizational conversation on guardrail design.
Ready to build the team and governance model agentic AI needs? Check out the two sneak preview sections of “Beyond Pipelines: The New Data Operating Model for AI at Scale” — “From System Failure to System Future” and “From Silos to Synergy.”