In detail
Human in the loop vs. human on the loop at a glance
The European Commission’s expert group on AI names three oversight approaches in its 2019 guidelines. The table puts them side by side, using the deployment sense described below for the first two.
| In the loop | On the loop | In command | |
|---|---|---|---|
| Who acts | The AI proposes; a person approves | The AI acts; a person monitors | A person decides whether and how the system is used at all |
| When the human intervenes | Before each action | While the system runs: step in, reverse a decision or shut it down | Before and during use: set how much discretion people have, or decide not to use the system |
| Latency cost | Every action waits for a person | No wait per action; a problem is caught after it happens | No per-action wait; the decision sits upstream |
| Example | A person approves each refund before it is issued | An agent runs routine, reversible tasks while a person watches its traces | Choosing not to use an AI system for a particular decision |
The two models, defined
The clearest definitions come from the European Commission’s expert group on AI, in its 2019 ethics guidelines. Human in the loop is “the capability for human intervention in every decision cycle of the system.” Human on the loop is the capability for human intervention while the system is designed and for “monitoring the system’s operation.” The guidelines add a note that matters for agents: intervening in every decision is “in many cases neither possible nor desirable.”
Teams deploying agents use the terms more narrowly. In the loop usually means a person approves actions before they happen: every email sent, every refund issued, every code change merged. On the loop means the agent acts on its own and a person watches what it does, with the power to step in, reverse a decision or shut it down. The rest of this page uses that deployment sense.
Why “in the loop” stops scaling
Approving every action works for a pilot. It breaks down when agents act at volume and in parallel, because the person becomes the bottleneck and then a rubber stamp. Kris Lovejoy, global head of strategy at Kyndryl, questioned the whole frame in episode 62: “And so I keep pushing back and asking, is it really human in the loop or is it human on top, right?” (11:03) She argues that people think about business processes in a linear way and AI doesn’t work that way, so the design problem is setting policy constraints such that, in her words, “humans don’t become kind of that roadblock to process fulfillment.” (11:03) That isn’t a case for removing checks. In the same answer she calls for humans in supervisory roles and says she is “a big fan of agentic AI requiring secondary evaluations and approvals before it can actually complete it.” (11:03)
The rubber-stamp risk has a name. Article 14 of the EU AI Act requires that the people overseeing a high-risk system be able “to remain aware of the possible tendency of automatically relying or over-relying on the output,” which the Act calls automation bias. A person who approves 500 actions a day without reading them is technically in the loop and practically absent. The same thing happens in code review when agents write most of the code.
How teams move from in to on
Start in the loop. Jeetu Patel, Cisco’s president and chief product officer, set the bar for removing a person altogether in episode 67: you consider it only once you trust the agent’s decisions “because I’ve seen enough of them statistically to know that I feel good about it. And then, and only then, should you actually even consider taking a human out of the loop.” (23:09) He expects that point to come at different times for simple and more complicated use cases. Moving a task from in the loop to on the loop is a smaller step than taking the person out, but it rests on the same kind of evidence.
Twilio’s Vinnie Giarrusso described that starting point on Twilio’s agent platform in episode 7 at 4:00, published in December 2024. Twilio chose to frame its product around assistants, he says, because it expects these agents, at least at the beginning, to work side by side with humans and augment the human role rather than replace it, with people who may be tasked with accepting or rejecting certain actions. At 17:00 he adds that Twilio’s early customers weren’t looking to fully deploy agents and let them run on their own, and that he didn’t think anyone wanted that yet. They set a goal and decide what the assistant can access, then test it. Galileo sponsored Season 1, when Conor worked there, and Giarrusso names Galileo as Twilio’s monitoring tool at 14:00.
That evidence is what evals produce. A team that can show an agent’s error rate on a task, measured on real cases, has a reason to move that task on the loop. A team that can’t is guessing.
Keep hard lines where mistakes are expensive
Moving to on the loop doesn’t mean removing every checkpoint. Tyler Akidau, CTO of Redpanda, described a wealth-management demo from a research paper his team wrote, with the approval rule enforced outside the agent: “But any trade that has more impact than, say, $1,000 on my portfolio always needs to be reviewed by a human. And there is no way for the agent to get past that.” The agent doesn’t even know the rule is there. It makes recommendations, and the infrastructure routes anything over the limit to a person.
A sensible deployment mixes both models. Routine, reversible, low-value actions run on the loop. Actions that are expensive, irreversible or regulated stay in it. Kapil Chhabra, co-founder and chief product officer of WisdomAI, describes where agents are heading: agents that take action on data rather than just retrieve it, which removes the person who used to judge the answer before anyone acted. That makes it more important to decide on purpose where a person still checks.
Deciding, action by action
For each action an agent can take, ask four questions:
- What does a mistake cost? Money, customers, safety, legal exposure.
- Can it be undone? A draft can be deleted. A wire transfer or a sent email usually can’t.
- How often does it happen? Approving ten actions a day is a review. Approving ten thousand is a rubber stamp.
- What is the evidence? A measured error rate on real cases, or a demo that went well.
High cost, irreversible or unproven: keep a person in the loop. Low cost, reversible and proven: move it on the loop, with monitoring, alerts and a way to stop it. How much autonomy to give an agent walks the same decision up a five-step ladder, and how to govern AI agents covers the controls around it.
Hear it from the guest
“And so I keep pushing back and asking, is it really human in the loop or is it human on top, right?”
“Because I've seen enough of them statistically to know that I feel good about it. And then, and only then, should you actually even consider taking a human out of the loop.”
“But any trade that has more impact than, say, $1,000 on my portfolio always needs to be reviewed by a human. And there is no way for the agent to get past that.”
“But instead sort of augment that human role with additional capabilities”
Quotes lightly edited to remove filler words.
Go deeper
- Ethics Guidelines for Trustworthy AI The source of the in-the-loop, on-the-loop and in-command definitions quoted on this page.
- Article 14: Human oversight The EU AI Act's oversight requirements for high-risk systems, including the warning about automation bias and the stop button.
- Building effective agents How agent builders design pauses for human feedback at checkpoints and stopping conditions into the loop itself.
Common questions
- Does the EU AI Act require a human in the loop?
- Not in the sense of approving every decision. Article 14 requires high-risk AI systems to be designed so that people can effectively oversee them, with oversight proportionate to the risk and the system's level of autonomy. As appropriate and proportionate, the people overseeing must be able to monitor it, disregard, override or reverse its output, and stop it. It adds one stricter rule for remote biometric identification: no action or decision can rest on an identification until two qualified people have separately verified it, with exceptions for law enforcement, migration, border control and asylum where the law deems that disproportionate. This is a summary, not legal advice.
- What is human in command?
- The third oversight approach named in the European Commission expert group's 2019 guidelines: oversight of the AI system's overall activity and its wider impact, including the decision about whether and how to use it in a given situation at all. It is broader than either in the loop or on the loop.
- Is human on the loop safe enough?
- Only if the monitoring is real. On the loop depends on traces someone actually reviews, alerts that fire on the right signals, and a working way to stop or reverse the system. Without those, it is no human at all. Keep fixed approval points for actions that are expensive or impossible to undo.
- When should you use human on the loop instead of human in the loop?
- Move a task on the loop when a mistake is cheap, it can be undone, and you have a measured error rate on real cases. Keep a person in the loop when an action is costly, irreversible or unproven. Cisco's Jeetu Patel puts the bar for removing the person at having seen enough decisions statistically to feel good about them, and only then. Approving ten thousand actions a day is a rubber stamp, not a review.
- Can one AI agent use both human in the loop and human on the loop?
- Yes, and most deployments should. Routine, reversible, low-value actions run on the loop with monitoring. Costly or regulated actions keep an approval step. Redpanda's Tyler Akidau described a demo where any trade with more than $1,000 of impact on the portfolio always needs human review and the agent cannot get past that rule, which is enforced outside the agent (episode 60).
From the conversation
This explainer is drawn from these episodes — each carries its full transcript.
-
The First Fully Autonomous AI Attack is Coming | Kris Lovejoy, Kyndryl -
Parenting Your AI Agents for Best Results | Cisco's Jeetu Patel -
We Built Agents, Nobody Built HR | Tyler Akidau, Redpanda -
Why Context Alone Isn't Enough for Enterprise AI Agents | WisdomAI CPO Kapil Chhabra -
Beyond Chatbots: How Twilio Uses AI to Strengthen Human Connection | Vinnie Giarrusso
Concepts in this explainer
AutomationAutomation BiasHigh-Risk AI SystemHuman in the LoopAI EvaluationEU AI ActAgentic AIAI Agent