Human in the loop (HITL) is a system design in which an automated process cannot complete a consequential step until a person has reviewed it and said yes. In advertising, the phrase now most often describes an AI agent that drafts a campaign, a budget change or a deal and then pauses for sign-off before anything goes live. The arrangement exists because software can act faster and at greater scale than anyone can check after the fact, and because law and contracts still need a named person to answer for what happened.

How an approval loop works

A system runs until it reaches a step marked as requiring approval. It then halts, saves its state and presents the pending action to a person. The system resumes once that person approves, edits or rejects it.

In OpenAI's Agents SDK, a tool can be declared with needs_approval, either always or per call. When the agent calls that tool, the run stops and returns a list of interruptions, each naming the agent, the tool and the arguments it wanted to pass. A developer approves or rejects each item and runs the agent again. LangGraph, LangChain's orchestration library, does the same through an interrupt() call that saves the graph's state until a person resumes it.

The Model Context Protocol (MCP), which many ad platforms use to expose tools to AI assistants, writes the expectation into its specification. "For trust & safety and security, there SHOULD always be a human in the loop with the ability to deny tool invocations," the June 18, 2025 version of the tools specification states. In that document "should" is a recommendation, not a requirement.

Typical gates in ad tech sit at:

  • Activation - a campaign, line item or deal is created but stays paused until a person switches it on.
  • Edits to live spend - budget, bid or targeting changes are queued as proposals.
  • External communication - anything an agent would send to a counterparty or publish under a brand.
  • Thresholds - changes above a set amount or percentage escalate, while smaller ones run automatically.

Autonomy levels: in, on and out of the loop

The phrase belongs to a family. Human on the loop describes a system that acts by itself while a person monitors and can step in. Human out of the loop means no one intervenes in individual decisions at all. The German digital association BVDW's 25-page agent framework, published on January 21, 2026, adds human in command, where people set goals and boundaries, and sets out a simple rule: "High damage scenarios require human-in-the-loop."

Several bodies have turned this into scales. The BVDW uses five levels. UC Berkeley's 67-page agentic AI risk profile uses six, from L0 under direct human control to L5, where "users function as observers". Spain's data protection authority, the AEPD, maps four in a 71-page guide dated February 2026, from "agent proposes, human operates" to "agent operates, human observes". None is an industry standard, but all tie greater autonomy to stronger logging, shutdown and permission controls.

Origin and evolution

The term came from engineering. Control engineers used "man in the loop", later "human in the loop", for simulators in which a person operated real controls against a modelled system. Thomas Sheridan and William Verplank's 1978 report for the US Office of Naval Research proposed ten levels of automation, from a computer offering no help to one that acts entirely on its own and ignores the human. It underlies the scales above.

Military policy gave the in/on/out distinction its sharpest form. US Department of Defense Directive 3000.09, issued on November 21, 2012 and updated on January 25, 2023, requires autonomous and semi-autonomous weapon systems to allow "appropriate levels of human judgment over the use of force".

Machine learning took the phrase in a third direction. There, HITL means people labelling data, correcting predictions or ranking outputs to train a model. Reinforcement learning from human feedback, described by Paul Christiano and colleagues in 2017, is the best-known form.

The legal track ran in parallel. Article 22 of the General Data Protection Regulation (GDPR), applicable since May 25, 2018, restricts decisions about people "based solely on automated processing" that carry legal or similarly significant effects. The EU Artificial Intelligence Act, Regulation (EU) 2024/1689, went further with Article 14 on human oversight for high-risk systems. The timetable change agreed in May 2026 and published as Regulation (EU) 2026/1744 pushes those duties to December 2, 2027 for stand-alone Annex III systems and August 2, 2028 for AI built into regulated products.

What the law actually requires

Article 14(1) of the AI Act requires high-risk AI systems to be designed "in such a way, including with appropriate human-machine interface tools, that they can be effectively overseen by natural persons during the period in which they are in use." Article 14(4) lists what the overseer must be enabled to do: understand the system's capacities and limitations, remain aware of the tendency to rely or over-rely on its output, interpret that output correctly, decide not to use it, and interrupt the system through "a 'stop' button or a similar procedure". For remote biometric identification, Article 14(5) requires results to be separately verified by at least two people. Article 26(2) puts the burden on deployers, who "shall assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support."

The Act does not use the phrase "human in the loop". Its oversight duty can be met by monitoring and a stop button, which is closer to human on the loop.

The GDPR asks a different question: whether the human involvement is real. EU data protection authorities wrote in 2017 that it must not be "a token review". The Dutch Autoriteit Persoonsgegevens (AP) spelled out what that means in a consultation run from March 6 to April 6, 2025: assessors "must have the authority to overrule algorithmic outcomes and must actually exercise this authority". The cost of failing that test is covered in the explainer on automated decision-making.

Why it matters for advertising

Agentic buying has turned HITL into a product feature. Amazon's Ads Agent, announced on November 11, 2025, let advertisers issue commands such as "pause all campaigns with return on ad spend less than 2", but campaigns launched only after review and approval. Adform went further in the other direction: the 29 skills it released for its FLOW demand-side platform (DSP) are all read-only. "Read-only - it reports; the trafficker fixes," one skill description says.

The Czech market's first agent-to-agent ad buy, completed on June 24, 2026, was billed as fully autonomous. Yet every campaign and creative still had to be approved manually by the publisher MAFRA before going live. Automation covered discovery and negotiation; a person stood at activation. Paused-by-default creation has become the common pattern, documented in PPC Land's explainer on the sandwich check, the review step between an agent's proposal and its execution.

An agent that misreads a brief can commit budget across hundreds of line items before a weekly report shows it, and contracts still assume a human placed the order.

Limitations and disputes

The best-documented weakness is automation bias: the tendency to accept a machine's recommendation without checking it. Research by Kathleen Mosier, Linda Skitka and colleagues in the 1990s found that people working with automated aids missed problems the aid did not flag and followed wrong advice the aid did give. The AI Act names the risk in Article 14(4)(b). The AEPD says automation bias is "exacerbated when systems operate with high autonomy and insufficient transparency". Respondents to the Dutch AP's consultation described interfaces that encourage routine approval, according to the summary published on June 3, 2025.

Volume is the second problem. The economic case for agents rests on doing more than people can. If every action needs sign-off, the human becomes the bottleneck; if few do, the loop is nominal. Lisanne Bainbridge's 1983 paper "Ironies of Automation" added that operators who rarely intervene lose the skill to do so.

The third dispute is about where responsibility lands. The Dutch AP warned that controllers must not shift "the responsibility for overseeing the entire process onto the assessor". A reviewer who approves an agent's output without the time, information or authority to reject it becomes a liability shield rather than a safeguard. Approval also checks only what is shown.

Not the same as

Human on the loop. Supervision with the ability to intervene, rather than approval before each action. Most AI Act oversight designs fit here.

Meaningful human involvement. The GDPR Article 22 test for whether a decision is "solely" automated. It concerns decisions about people, not campaign actions, and can fail even where a human signs off.

Human-in-the-loop machine learning. People labelling or ranking data to train models. The human shapes the model, not individual live actions.

Recent developments

Supermetrics said on November 3, 2026 that Anthropic's Claude can now build campaigns on six ad platforms through its connector. New campaigns stay paused until a person approves them, edits to live campaigns can be reviewed first, and customers can set approval requirements by account. Chief Executive Officer Kalle Thuneberg described the aim as automating work "while keeping the human firmly in control". It did not say whether approval can be switched off.

Amazon's relaunched Ads Agent kept campaigns on hold for approval on September 29, 2026. The AI Act's human oversight rules for high-risk systems remain 13 months away from applying, and the harmonised standards meant to show how Article 14 is met had not been finalised as of November 2026.

Timeline

  • 1978 - Sheridan and Verplank publish a ten-level scale of automation for the US Office of Naval Research.
  • 1983 - Lisanne Bainbridge publishes "Ironies of Automation".
  • 1990s - Mosier, Skitka and colleagues document automation bias in automated decision aids.
  • November 21, 2012 - US DoD Directive 3000.09 requires "appropriate levels of human judgment" over autonomous weapons.
  • 2017 - Christiano and colleagues describe reinforcement learning from human feedback.
  • October 3, 2017 - Article 29 Working Party guidelines say human involvement must not be "a token review".
  • May 25, 2018 - GDPR applies, including Article 22.
  • 2021 - Robert Munro publishes "Human-in-the-Loop Machine Learning".
  • January 25, 2023 - DoD Directive 3000.09 updated.
  • August 1, 2024 - EU AI Act enters into force, with Article 14 on human oversight.
  • March 6 to April 6, 2025 - Dutch AP consults on meaningful human intervention.
  • June 18, 2025 - MCP specification revision says a human SHOULD be able to deny tool calls.
  • November 11, 2025 - Amazon announces Ads Agent with approval before launch.
  • January 21, 2026 - BVDW publishes five-level agent autonomy framework.
  • February 2026 - AEPD publishes agentic AI guide with four autonomy levels.
  • June 24, 2026 - First Czech agent-to-agent ad buy, with publisher approval at activation.
  • July 27, 2026 - Regulation (EU) 2026/1744 enters into force, delaying high-risk duties.
  • November 3, 2026 - Supermetrics adds paused-by-default campaign creation through Claude.
  • December 2, 2027 - AI Act human oversight duties apply to Annex III high-risk systems.

Summary

Who. Advertisers, agencies and platforms deploying AI agents, which configure the approval gates; the people who review proposed actions; and regulators such as the AEPD, the Dutch AP and EU market surveillance authorities, which judge whether oversight is real.

What. A design in which an automated system pauses at defined steps until a person approves, edits or rejects its proposed action. It differs from human on the loop, where a person monitors and can intervene but does not sign off each step.

When. The term comes from control engineering and simulation, was formalised in automation research from 1978, entered military policy in 2012 and AI regulation through GDPR Article 22 in 2018 and AI Act Article 14 in 2024. Oversight duties for high-risk AI apply from December 2, 2027.

Where. In agent frameworks such as OpenAI's Agents SDK and LangGraph, in the MCP specification, in ad platform tools from Amazon, Adform, Meta and Supermetrics, and in EU law.

Why. Agents can commit budget and act on people faster than anyone can audit afterwards. A human approval step keeps errors from going live and keeps a named person accountable, though automation bias and approval volume can empty it of substance.