Agentic AI is moving faster than most organizations' ability to govern it. Vendors are shipping autonomous agents that can reason, plan, and act without a person approving each step. Boards are asking for an AI strategy. And in the middle of that rush, a phrase keeps getting repeated as if it settles the matter: "we'll have a human in the loop."
But what does that actually mean in practice? According to Dr. Fern Halper, VP of Research at TDWI and founder of the AI Foundations Group, for a lot of organizations it doesn't mean much at all. It's a checkbox.
Halper joined the Lights On Data Show to talk about her new book, Data Makes the World Go 'Round: The Data, Tech, and Trust Behind AI Success, and about research she's currently running at TDWI into what human oversight of AI agents actually looks like once the pilot phase ends. Here's what stood out.
What Makes Agentic AI Different From a Chatbot
Generative AI, in Halper's framing, produces an output that a person then reviews. Agentic AI is a different animal: agents that reason and act autonomously, deciding on their own what task to perform next and which other agent to hand the output to.
That autonomy is exactly why "human in the loop" has become the default answer to the obvious follow-up question: what happens when the agent gets it wrong?
Why So Few Organizations Are Actually There Yet
Despite the hype, Halper's research suggests the industry isn't as far along as the marketing suggests. Most organizations that say they're building agentic systems are really building single-agent systems that perform one task. By her estimate, fewer than 10% of organizations are currently running true multi-agent systems where agents hand work off to each other autonomously.
That gap matters. It means most companies haven't yet hit the harder problem Halper is most concerned about: what happens to human judgment once agents are doing most of the reasoning.
The Math Problem Nobody's Solving
Here's the tension at the center of the human-in-the-loop conversation. TDWI has an active survey asking data and AI leaders whether they currently act as a human in the loop, whether they expect to in five to ten years, and whether they're worried about losing their jobs to agents.
Halper found that data and AI practitioners with two-plus decades of experience are largely unconcerned about being replaced. They assume their judgment and critical thinking will make them the ones supervising the agents.
The problem, as Halper put it, is that it's a math game: organizations don't need nearly as many supervisors as they need current employees, so even if human oversight is the plan, far fewer humans will be doing it. That's a workforce redesign question that most technology leaders haven't thought through carefully. Leaders are treating "we'll have a human in the loop" as a solved design pattern rather than a genuine organizational redesign.
Automation Bias Is Already Showing Up
Even where a human oversight role does exist, Halper pointed to a well-documented risk that predates generative AI by decades: automation bias, the tendency for people to simply accept whatever the AI produces rather than critically evaluating it.
She connected this to a broader concern she's written about separately: the erosion of original thought. As AI output pulls responses toward the statistical middle of what a model has learned, the humans reviewing that output can start defaulting to it too, even when they're not directly copying it. Some companies are now instituting deliberate countermeasures, like requiring analysts to form a hypothesis before consulting AI, or scheduling AI-free days for certain types of work.
When Human in the Loop Actually Works
Not every example Halper raised was a cautionary tale. She pointed to healthcare as a domain where human-AI collaboration is being designed thoughtfully, because clinicians are working alongside AI tools rather than being replaced by them, and organizations in that space have generally taken the time to redesign workflows around that collaboration rather than bolting AI on top of the old process.
A similar principle showed up in a call center example discussed on the episode: instead of routing customers to an AI agent and escalating only the failures to a human, some organizations are doing the reverse. The human stays on the call, and an AI system listens in and surfaces contextual guidance, essentially a real-time cheat sheet, that the human can choose to use or ignore. That's AI in the loop supporting a human, rather than a human in the loop supervising an AI.
The Real Risk of "Work Slop"
One of the more concrete failure modes Halper described: a marketing team generates content with AI, then hands it to a subject matter expert to review and approve, rather than having that expert produce the content themselves. The expert's day-to-day work shifts from creating to checking, and the material they're checking is often, in Halper's words, low-quality output that takes longer to fix than it would have taken to write from scratch. Researchers have started calling this pattern "work slop." It's a version of human-in-the-loop that technically exists but leaves people cognitively disengaged and, according to the research Halper cited, generally unsatisfied with the role.
Why Trust Gets Equal Billing With Data and Technology
The subtitle of Halper's book puts trust on the same level as data and technology, and she was direct about why: trust isn't a single checkpoint, it's a property that has to hold at every control point across the data and AI lifecycle. That includes trusting the data going into a model (increasingly unstructured data, which most organizations have governed far less rigorously than structured data), trusting the model's behavior, and trusting the output that comes out the other end.
Her research has found a consistent pattern: organizations that deploy real governance around that full lifecycle see measurably bigger returns from their AI investments than organizations that treat governance as an afterthought.
The Core Takeaway
AI doesn't fix what's broken in an organization. It magnifies it. Weak data governance, undefined roles, and ungoverned processes don't go away when you add an agent, they get amplified by one. The organizations Halper's research shows succeeding are the ones treating AI as a set of organizational capabilities to build, not a tool to install.
Do you want to learn more?
Practical Data Governance: Implementation - online course
Learn how to implement a data governance program from scratch or improve the one you have.
Where to Learn More
Dr. Fern Halper's book, Data Makes the World Go 'Round: The Data, Tech, and Trust Behind AI Success, is available through Amazon, Barnes & Noble, and Wiley, the publisher. You can also find her ongoing research on human-in-the-loop AI adoption through TDWI.
Listen to the full conversation with Dr. Fern Halper on the Lights On Data Show.
Frequently Asked Questions
What's the difference between agentic AI and generative AI?
Generative AI produces an output for a person to review. Agentic AI involves agents that reason, plan, and act autonomously, including deciding what task to do next and which other agent to hand results to.
What does "human in the loop" mean in AI governance?
It refers to a person validating, checking, or approving an AI system's output or actions. In practice, its effectiveness varies widely, from genuine oversight to what Dr. Fern Halper describes as a compliance checkbox with little real scrutiny behind it.
Why is automation bias a risk in human-in-the-loop systems?
Automation bias is the well-documented tendency for people to accept AI-generated output without critically evaluating it, which can undermine the entire premise of human oversight.
How many organizations are actually running multi-agent AI systems?
Based on Dr. Fern Halper's research at TDWI, fewer than 10% of organizations are currently running true multi-agent systems, even though many describe their AI initiatives as "agentic."
Connect with LightsOnData
- Podcast: www.anchor.fm/lightsondata
- LinkedIn: www.linkedin.com/company/lightsondata
- YouTube: www.youtube.com/c/lightsondata
- Instagram: https://instagram.com/lightsondata
- George on LinkedIn: www.linkedin.com/in/georgefirican
