Picture an agentic workflow that quietly declines the wrong customer because of a fault somewhere upstream. Now picture the meeting where your organization decides whose fault that was. Unless you did some specific, unglamorous work in advance, the answer is going to be you. That uncomfortable scenario is the honest starting point for talking about agentic AI, and almost nobody wants to lead with it.

Integration is where the risk actually lives

There is a tidy assumption that the risk in AI sits inside the model. In practice it sits in the plumbing. An agent is only as good as the data it can reach, and data only becomes useful once it is integrated, cleaned, and given context. The moment you wire a model into your real systems through something like an automation platform or an MCP server, you have handed it the ability to act, and the old problem of tangled integrations becomes a new problem of tangled agent behaviour. Same spaghetti, higher stakes, because now it makes decisions instead of just moving records.

The question that has no satisfying answer

How do you constrain an agent once it can call tools on its own? The honest answer is the one that frustrates everyone: it depends. It depends on what the agent is allowed to touch, what it is allowed to change without a human, and how a person can see what it did and why. None of that is exciting, and all of it is the actual job. High data quality and real governance are not the boring prelude to the AI work. On an agentic system, they are the work, because a confident model acting on bad or context-free data does the wrong thing faster and more consistently than any human ever could.

Accountability does not disappear, it moves

The dangerous fantasy is that handing a decision to an agent also hands off the responsibility for it. It does the opposite. When the workflow declines someone it should not have, the customer does not care that a model made the call, and neither will your regulator. Responsibility flows back to whoever designed the loop and decided what it could do unsupervised. So the design question is not just can the agent do this, it is who answers when it does this wrong, and can they see enough to explain it. If the honest answer is that no one can reconstruct why the agent acted, the system is not ready to run, however good the demo looked.

How we build for it

This is why every system we ship keeps a human on the decision that carries consequence, logs what the agent did in a form a person can audit, and escalates anything it is unsure about with the full context attached. It is the same standard we hold across our solutions and the reason we treat the unhappy paths, not the demo, as the real test. Agentic AI is a genuine capability multiplier, and the teams that invest in the controls will pull ahead of the ones that cut them. But the controls are the price of running it at all, not an optional upgrade, and pretending otherwise just decides in advance who takes the blame. It connects directly to why integration is the fastest-growing paid skill: the model is easy, the governance around it is not.

Prompted by Robert Hogg’s talk “How ethical is your agentic AI integration platform and how would you even know” (Black Marble, Integrate 2026). Credit to the speaker; the framing here is ours.