As businesses move from experimenting with AI to putting it into real workflows, the conversation around human oversight needs to evolve.
We hear a lot about keeping a “human-in-the-loop.” Most often, that means a person reviews what AI produces before a decision is made or an action is taken. For many organizations, that’s an appropriate place to start. When the technology, data, or process is still being tested, human review provides an important safeguard.
But reviewing every AI output by hand isn’t a scalable operating model.
The opportunity now is to move human judgment earlier in the process, where business context, rules, and guardrails can guide how an agent acts.
Put business expertise into the workflow
I refer to that shift as “embedded oversight”.
Instead of relying on people to catch problems after AI has produced an answer, we can embed the knowledge of domain experts into the workflow before AI ever acts.
After all, the people closest to a business process usually understand its nuances better than anyone.
A finance analyst knows which variances warrant investigation. A claims expert understands which combinations of information should trigger additional scrutiny. A tax professional knows which rules apply to a particular jurisdiction. An operations leader understands when an exception is meaningful and when it is simply normal variation.
That knowledge is incredibly valuable to AI.
The goal should be to translate it into reusable rules, workflows, thresholds, and guardrails that AI-powered systems can consistently follow. That same principle is reflected in our VURA framework for trustworthy AI. For AI to operate reliably at scale, the workflows and business logic behind it need to be visible, understandable, repeatable, and auditable.
Embedded oversight helps make those principles operational by putting human expertise into the process before an agent acts.
What embedded oversight looks like in practice
Let me give you two examples of what this looks like, because I think the idea only lands when you see it in a real workflow.
The first is from a finance team doing month-end balance sheet reconciliation. The workflow pulls in journal entries, PDFs, and extracts from the ERP system, then rolls up 30 days of daily calculations. Most of the time it runs in the background. But if it spots something off, say a large, unexplained swing against a single vendor across that 30-day window, it stops and sends an alert, either by email or as a pop-up inside the platform.
The analyst can approve the number, reject it, or correct it before the workflow continues. Nobody’s reviewing every line of the reconciliation by hand. The human attention goes exactly where it’s needed and nowhere else.
The second comes from one of our insurance partners. Their AI handles the early back-and-forth with someone reporting a car accident. Most of the time, that conversation is routine. But when something doesn’t line up, say a claimant valuing their vehicle at £1 million against a registration plate that matches a much cheaper make and model, the workflow pauses and pulls in a claims handler.
The handler doesn’t have to review every claim that comes through the door. They show up exactly when the numbers stop making sense, look at the mismatch, and make the call a person should be making anyway.
Both examples work the same way underneath. Someone decided in advance what the agent was allowed to do. The workflow ran inside those limits and stopped the moment it hit the edge of them. The human wasn’t a bottleneck at the end of the process. They were a decision made at the front of it, before the workflow ever ran.
How this impacts the analyst
I’m often asked how embedded oversight will impact the job of the analyst. My take is that analysts will become managers of agents.
They’ll give it the business logic it needs to do the job well, set the thresholds that tell it when to stop and ask rather than guess, and review the logic behind its decisions, adjusting as the business itself changes.
Those responsibilities require many of the skills that already make strong analysts valuable: business context, data literacy, critical thinking, and an understanding of how processes really work.
The next phase of human-in-the-loop
As enterprise AI matures, I don’t think success will be measured by how completely we remove people from a process but how effectively we can apply human expertise.
For some decisions, that will still mean explicit human approval. For others, it will mean monitoring AI and intervening when necessary. And increasingly, it will mean capturing human judgment at the beginning of the process through trusted data, business rules, workflows, and guardrails.
That’s where human-in-the-loop is heading.
When we shift human judgment to the front end, we can give AI more room to operate without giving up the context, control, and accountability enterprises need to trust the results.