{"id":4363,"date":"2026-11-06T18:00:46","date_gmt":"2026-11-06T18:00:46","guid":{"rendered":"https:\/\/palegoldenrod-boar-492303.hostingersite.com\/?p=4363"},"modified":"2026-11-08T18:01:26","modified_gmt":"2026-11-08T18:01:26","slug":"human-in-the-loop-ai-agents","status":"publish","type":"post","link":"https:\/\/growthnation.ai\/insights\/human-in-the-loop-ai-agents\/","title":{"rendered":"Human in the Loop AI Agents: Choose the Person Before the Approval Button"},"content":{"rendered":"<span class=\"span-reading-time rt-reading-time\" style=\"display: block;\"><span class=\"rt-label rt-prefix\">Reading Time: <\/span> <span class=\"rt-time\"> 12<\/span> <span class=\"rt-label rt-postfix\">minutes<\/span><\/span><div class=\"\">\n\t\t\t<div class=\"aioseo-toc-header\">\n\t\t\t\t<header class=\"aioseo-toc-header-area\">\n\t\t\t\t\t<div class=\"aioseo-toc-header-title aioseo-toc-header-collapsible-closed \">\n\t\t\t\t\t<div class=\"aioseo-toc-header-collapsible\">\n\t\t\t\t\t\t<svg width=\"14\" height=\"14\" viewBox=\"0 0 14 14\" fill=\"none\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\">\n\t\t  <path d=\"M6 8H0V6H6V0H8V6H14V8H8V14H6V8Z\" fill=\"#005AE0\"\/>\n\t\t<\/svg>\n\t\t\t\t\t<\/div>\n\t\t\t\t\tShow Table of Contents\n\t\t\t\t\t<\/div>\n\n\t\t\t\t\t<div class=\"aioseo-toc-header-title aioseo-toc-header-collapsible-open aioseo-toc-collapsed\">\n\t\t\t\t\t<div class=\"aioseo-toc-header-collapsible\">\n\t\t\t\t\t\t<svg width=\"14\" height=\"2\" viewBox=\"0 0 14 2\" fill=\"none\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\">\n\t\t  <path d=\"M0 2V0H14V2H0Z\" fill=\"#005AE0\"\/>\n\t\t<\/svg>\n\t\t\t\t\t<\/div>\n\t\t\t\t\tHide Table of Contents\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/header>\n\t\t\t\t<div class=\"aioseo-toc-contents aioseo-toc-collapsed\">\n\t\t\t\t\t<ul><li><a class=\"aioseo-toc-item\" href=\"#aioseo-why-approval-steps-turn-into-rubber-stamps-4\">Why Approval Steps Turn Into Rubber Stamps<\/a><\/li><li><a class=\"aioseo-toc-item\" href=\"#aioseo-a-person-who-approves-and-a-person-who-checks-12\">A Person Who Approves and a Person Who Checks<\/a><\/li><li><a class=\"aioseo-toc-item\" href=\"#aioseo-what-human-in-the-loop-ai-agents-need-from-their-person-16\">What Human in the Loop AI Agents Need From Their Person<\/a><ul><li><a class=\"aioseo-toc-item\" href=\"#aioseo-someone-who-still-knows-the-work-18\">Someone Who Still Knows the Work<\/a><\/li><li><a class=\"aioseo-toc-item\" href=\"#aioseo-time-to-check-written-into-the-job-20\">Time to Check, Written Into the Job<\/a><\/li><li><a class=\"aioseo-toc-item\" href=\"#aioseo-authority-to-say-no-without-asking-22\">Authority to Say No Without Asking<\/a><\/li><\/ul><\/li><li><a class=\"aioseo-toc-item\" href=\"#aioseo-how-to-tell-whether-anyone-is-still-checking-24\">How to Tell Whether Anyone Is Still Checking<\/a><\/li><li><a class=\"aioseo-toc-item\" href=\"#aioseo-where-the-person-sits-in-one-tour-operator-32\">Where the Person Sits in One Tour Operator<\/a><\/li><li><a class=\"aioseo-toc-item\" href=\"#aioseo-setting-up-human-in-the-loop-ai-agents-36\">Setting Up Human in the Loop AI Agents<\/a><\/li><li><a class=\"aioseo-toc-item\" href=\"#aioseo-where-growthnation-fits-44\">Where GrowthNation Fits<\/a><\/li><li><a class=\"aioseo-toc-item\" href=\"#aioseo-find-the-person-before-you-add-the-button-46\">Find the Person Before You Add the Button<\/a><\/li><li><a class=\"aioseo-toc-item\" href=\"#aioseo-frequently-asked-questions-56\">Frequently Asked Questions<\/a><ul><li><a class=\"aioseo-toc-item\" href=\"#aioseo-what-are-human-in-the-loop-ai-agents-57\">What are human in the loop AI agents?<\/a><\/li><li><a class=\"aioseo-toc-item\" href=\"#aioseo-what-is-the-difference-between-human-in-the-loop-and-human-on-the-loop-59\">What is the difference between human in the loop and human on the loop?<\/a><\/li><li><a class=\"aioseo-toc-item\" href=\"#aioseo-which-actions-should-an-ai-agent-need-a-persons-approval-for-61\">Which actions should an AI agent need a person&#039;s approval for?<\/a><\/li><li><a class=\"aioseo-toc-item\" href=\"#aioseo-how-do-you-stop-approvals-becoming-a-rubber-stamp-63\">How do you stop approvals becoming a rubber stamp?<\/a><\/li><li><a class=\"aioseo-toc-item\" href=\"#aioseo-is-human-oversight-of-ai-agents-a-legal-requirement-65\">Is human oversight of AI agents a legal requirement?<\/a><\/li><li><a class=\"aioseo-toc-item\" href=\"#aioseo-who-is-accountable-when-an-approved-ai-agent-gets-something-wrong-67\">Who is accountable when an approved AI agent gets something wrong?<\/a><\/li><\/ul><\/li><\/ul>\n\t\t\t\t<\/div>\n\t\t\t<\/div>\n\t\t<\/div>\n\n\n<p><em><strong>Quick Answer: <\/strong>Human in the loop AI agents pause before the actions that matter so a person can approve, change or stop them. That pause only protects anyone if the person still knows the work, has time set aside to check and can say no without asking permission. Record what they change, and test now and then that they would catch a planted error.<\/em><\/p>\n\n\n\n<p>More than 90 per cent of <a href=\"https:\/\/www.bankofengland.co.uk\/report\/2026\/artificial-intelligence-in-uk-financial-services-2026\">UK financial firms<\/a> answering the full Bank of England and FCA survey, published in October 2026, reported human approval among their AI controls. Ongoing monitoring and periodic review were each reported by around half. Adding a person has become routine; checking that the person still makes a difference has not. For human in the loop AI agents, that person is often all that stands between a draft and a customer, a payment or a decision about someone&#8217;s job.<\/p>\n\n\n\n<p>Most advice on keeping a person in the loop is about the gate: which <a href=\"https:\/\/growthnation.ai\/insights\/ai-acceptable-use-policy\/\">actions to pause<\/a>, what triggers a review, and whether a person approves first or watches afterwards. That matters once an agent <a href=\"https:\/\/growthnation.ai\/insights\/what-is-an-ai-agent\/\">takes actions<\/a> in real systems, and choosing work that is <a href=\"https:\/\/growthnation.ai\/insights\/ai-transformation-strategy\/\">easy to undo<\/a> comes first. The advice says little about the person at the gate, though more people now spend their day <a href=\"https:\/\/growthnation.ai\/insights\/agentic-transformation\/\">reviewing the output<\/a> rather than producing it. What follows is about that person: who they should be, what they need, and how to tell whether they are still checking.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"aioseo-why-approval-steps-turn-into-rubber-stamps-4\">Why Approval Steps Turn Into Rubber Stamps<\/h2>\n\n\n\n<p><a href=\"https:\/\/academic.oup.com\/pnasnexus\/article\/5\/6\/pgag146\/8703788\">One 2026 experiment<\/a> shows how easily a check bends. Researchers gave 1,339 teachers in Greece a piece of student work with a deliberately wrong mark, labelled as either a colleague&#8217;s or AI-generated. When the wrong mark was too harsh, the gap between the teacher&#8217;s final mark and the fair one was 22 per cent larger under the AI label, most of all among younger, highly qualified and tech-confident teachers. Each teacher marked one exercise in a survey, not live work, but the pattern fits any approval queue.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>The agent is usually right.<\/strong> After weeks of correct <a href=\"https:\/\/growthnation.ai\/insights\/ai-agents-for-customer-service\/\">refund replies<\/a>, attention drifts, and the occasional wrong one goes through with the rest.<\/li>\n\n\n\n<li><strong>An AI answer looks competent.<\/strong> A confident, tidy output gets less scrutiny than the same work from a colleague.<\/li>\n\n\n\n<li><strong>The answer arrives before the case.<\/strong> Once <a href=\"https:\/\/growthnation.ai\/insights\/ai-agents-for-finance-teams\/\">a payment run<\/a> has been prepared, the reviewer asks whether it looks plausible, not whether it is right.<\/li>\n\n\n\n<li><strong>Saying no costs more than saying yes.<\/strong> Approving is one click. Rejecting needs a reason, delays a customer and sometimes needs a manager.<\/li>\n<\/ul>\n\n\n\n<p>With human in the loop AI agents, an approval step records that someone clicked. It does not record that anyone checked.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"aioseo-a-person-who-approves-and-a-person-who-checks-12\">A Person Who Approves and a Person Who Checks<\/h2>\n\n\n\n<p>Courts have started to draw the same line. In a June 2025 <a href=\"https:\/\/caselaw.nationalarchives.gov.uk\/ewhc\/admin\/2025\/1383\">High Court ruling<\/a>, the Divisional Court said lawyers who use AI for research must check it against authoritative sources, including research someone else prepared, and that leaders must make sure everyone understands that duty. In one of the two cases before it, 18 of 45 citations did not exist. The ruling concerns lawyers, but the distinction applies to any team checking an agent.<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><td>&nbsp;<\/td><td><strong>A person who approves<\/strong><\/td><td><strong>A person who checks<\/strong><\/td><\/tr><\/thead><tbody><tr><td><strong>What they see first<\/strong><\/td><td>The agent&#8217;s answer<\/td><td>The case, then the answer<\/td><\/tr><tr><td><strong>What they know about the work<\/strong><\/td><td>Enough to spot a draft that looks wrong<\/td><td>Enough to have done it by hand recently<\/td><\/tr><tr><td><strong>Time for it<\/strong><\/td><td>Fitted around the day job<\/td><td>Counted in their hours<\/td><\/tr><tr><td><strong>Saying no<\/strong><\/td><td>Needs a reason, sometimes a manager<\/td><td>Stops the action, no permission needed<\/td><\/tr><tr><td><strong>What gets recorded<\/strong><\/td><td>That they clicked approve<\/td><td>What they changed, and why<\/td><\/tr><tr><td><strong>When it goes wrong<\/strong><\/td><td>The log shows a person agreed<\/td><td>The log shows what they checked<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>Approval still has a job. It records who agreed and who <a href=\"https:\/\/growthnation.ai\/insights\/who-owns-ai-agents-built-by-employees\/\">answers for it<\/a> when something goes wrong. It cannot be the only safeguard.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"aioseo-what-human-in-the-loop-ai-agents-need-from-their-person-16\">What Human in the Loop AI Agents Need From Their Person<\/h2>\n\n\n\n<p>When a <a href=\"https:\/\/www.sra.org.uk\/news\/news\/releases\/2025\/garfield-ai-authorised\/\">regulator approved<\/a> the first AI-driven law firm in England and Wales in May 2025, it settled who stands in the loop as well as where. The system takes a step only once the client has approved it, oversight was greater <a href=\"https:\/\/growthnation.ai\/insights\/ai-pilot-to-production\/\">during launch<\/a>, and named solicitors remain responsible for all its outputs. Those are one regulator&#8217;s terms for one firm, but they describe what human in the loop AI agents need from their person: knowledge, time and authority.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"aioseo-someone-who-still-knows-the-work-18\">Someone Who Still Knows the Work<\/h3>\n\n\n\n<p>A check is only as good as the checker&#8217;s grasp of the work. In <a href=\"https:\/\/arxiv.org\/abs\/2509.08514\">a 2025 experiment<\/a>, 2,784 US crowdworkers reviewed staged AI answers from company reports. They corrected spelling-type slips 82 per cent of the time, but caught errors that needed knowledge of the task&#8217;s rules only 31 per cent of the time. It is a preprint using non-specialists, but the pattern is clear. Skill also fades once the agent does the work. In an observational <a href=\"https:\/\/www.thelancet.com\/journals\/langas\/article\/PIIS2468-1253%2825%2900133-5\/abstract\">colonoscopy study<\/a> at four Polish centres, doctors&#8217; detection rate without AI fell from 28.4 to 22.4 per cent after AI was introduced. The best checker has done the job and still does a <a href=\"https:\/\/growthnation.ai\/insights\/do-ai-agents-replace-employees\/\">share by hand<\/a>, because what they check against was <a href=\"https:\/\/growthnation.ai\/insights\/tacit-knowledge-management\/\">learned by heart<\/a> and is easily lost when <a href=\"https:\/\/growthnation.ai\/insights\/knowledge-loss-when-employees-leave\/\">the expert retires<\/a>.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"aioseo-time-to-check-written-into-the-job-20\">Time to Check, Written Into the Job<\/h3>\n\n\n\n<p>A check fitted around other work becomes a glance. Decide how many items one person checks in a day, and give the check its own time. Count the <a href=\"https:\/\/growthnation.ai\/insights\/how-to-measure-ai-roi\/\">checking time<\/a> against the hours an agent saves, or the saving is counted twice. Record who checks which agent in <a href=\"https:\/\/growthnation.ai\/insights\/ai-center-of-excellence\/\">the agent register<\/a>, with cover for leave. Human in the loop AI agents without a named, timed check fall back on whoever is free.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"aioseo-authority-to-say-no-without-asking-22\">Authority to Say No Without Asking<\/h3>\n\n\n\n<p>The person in the loop must be able to stop an action and send it back without asking a manager. The agent should wait rather than time out and send anyway, and rejecting should take no longer than approving. Some decisions should never rest on a click: anything about <a href=\"https:\/\/growthnation.ai\/insights\/ai-agents-for-hr\/\">pay or dismissal<\/a> stays with a named person. And someone must be able to <a href=\"https:\/\/growthnation.ai\/insights\/ai-agent-security\/\">stop the agent<\/a> outright when the same mistake keeps coming back.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"aioseo-how-to-tell-whether-anyone-is-still-checking-24\">How to Tell Whether Anyone Is Still Checking<\/h2>\n\n\n\n<p>Few organisations test whether their checks work. The European Data Protection Supervisor&#8217;s <a href=\"https:\/\/www.edps.europa.eu\/data-protection\/our-work\/publications\/techdispatch\/2025-09-23-techdispatch-22025-human-oversight-automated-making_en\">2025 oversight guidance<\/a>, written for EU institutions, lists twelve assumptions about oversight that often do not hold, including that the person in the loop has authority to decide. A later checklist from the same office suggests tracking how often operators override the system and sometimes testing them without it. The practices carry over to human in the loop AI agents anywhere.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Record every change the person makes.<\/strong> Months of near-total approval is a question to ask, not a result to report.<\/li>\n\n\n\n<li><strong>Plant a known error now and then.<\/strong> If it goes through, the check has stopped working.<\/li>\n\n\n\n<li><strong>Ask for a decision before the answer.<\/strong> On a sample of cases, the person decides before seeing what the agent prepared.<\/li>\n\n\n\n<li><strong>Count the load.<\/strong> Note how many checks land on one person in a day, and move work before careful checking becomes impossible.<\/li>\n<\/ul>\n\n\n\n<p>If nobody has changed an agent&#8217;s work in three months, find out whether the agent is that good or the check has stopped.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"aioseo-where-the-person-sits-in-one-tour-operator-32\">Where the Person Sits in One Tour Operator<\/h2>\n\n\n\n<p>Consider a regional tour operator with around 350 staff across a head office, a contact centre and a small accounts team. Its five human in the loop AI agents each had an approval button, pressed by whoever was free. After a review, it named one checker per agent and decided what each looks at first.<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><td><strong>Agent<\/strong><\/td><td><strong>What it prepares<\/strong><\/td><td><strong>Who checks<\/strong><\/td><td><strong>What they look at first<\/strong><\/td><td><strong>How the check is tested<\/strong><\/td><\/tr><\/thead><tbody><tr><td>Refunds for cancelled holidays<\/td><td>The refund amount and the letter<\/td><td>A customer relations adviser with refund authority<\/td><td>Booking conditions and payments made<\/td><td>A planted case with a wrong deposit each month<\/td><\/tr><tr><td>Rebooking during disruption<\/td><td>New flights, hotels and the message<\/td><td>The duty operations manager<\/td><td>Travellers with children or assistance needs<\/td><td>A sample rebooked by hand and compared<\/td><\/tr><tr><td>Hotel invoice matching<\/td><td>Invoice matches and the payment queue<\/td><td>The accounts assistant who used to match them<\/td><td>Room nights that differ from bookings<\/td><td>One week a quarter matched by hand first<\/td><\/tr><tr><td>Complaint final responses<\/td><td>The reply, drafted from the case notes<\/td><td>The complaints team leader, who signs it<\/td><td>What the customer asked for, before the draft<\/td><td>Monthly review of changes; a month with none is queried<\/td><\/tr><tr><td>Special assistance requests<\/td><td>A summary of needs and a note to the hotel<\/td><td>The assistance coordinator, every time<\/td><td>The traveller&#8217;s own words<\/td><td>The agent never sends; the coordinator does<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>The <a href=\"https:\/\/growthnation.ai\/insights\/ai-agent-for-invoice-processing\/\">hotel invoices<\/a> check went to the assistant who used to match them by hand, not the finance manager, because the assistant knew which hotels invoiced differently.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"aioseo-setting-up-human-in-the-loop-ai-agents-36\">Setting Up Human in the Loop AI Agents<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Name the person before you place the step.<\/strong> One named checker per agent, with a stand-in.<\/li>\n\n\n\n<li><strong>Write down what they check.<\/strong> <a href=\"https:\/\/growthnation.ai\/insights\/capturing-institutional-knowledge\/\">Ask the people<\/a> who do the work which errors matter, and build the check around those errors.<\/li>\n\n\n\n<li><strong>Keep them doing some of the work by hand.<\/strong> A week a quarter, or a sample each month, keeps their judgement current.<\/li>\n\n\n\n<li><strong>Put the checking time in the plan.<\/strong> An agent that saves five hours a week and needs two hours of checking <a href=\"https:\/\/growthnation.ai\/insights\/how-to-prove-ai-value-to-leadership\/\">saves three<\/a>.<\/li>\n\n\n\n<li><strong>Let them stop it without asking.<\/strong> Then <a href=\"https:\/\/growthnation.ai\/insights\/change-management-for-ai-adoption\/\">tell the team<\/a>, so stopping an agent is seen as part of the job.<\/li>\n\n\n\n<li><strong>Read the overrides every month.<\/strong> An agent that is never corrected tends to be <a href=\"https:\/\/growthnation.ai\/insights\/why-ai-projects-fail\/\">quietly dropped<\/a> or, worse, quietly trusted.<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"aioseo-where-growthnation-fits-44\">Where GrowthNation Fits<\/h2>\n\n\n\n<p>GrowthNation&#8217;s AI agent interviews every team, by voice or text and in confidence, about the work, including which mistakes matter and who catches them. Agents built from those interviews are served centrally and belong to the organisation, so the named checker and the check step stay with the agent when people move on. Hours saved are tracked for each agent in daily use, with checking time counted, which keeps human in the loop AI agents honest about what the loop costs.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"aioseo-find-the-person-before-you-add-the-button-46\">Find the Person Before You Add the Button<\/h2>\n\n\n\n<p>Pick one agent you already run. Ask who approved its last ten outputs, how long each took and how many they changed. If the answers are whoever was free, a few seconds and none, the loop has a button in it but no check.<\/p>\n\n\n\n<p>Book a walkthrough and GrowthNation will interview your teams, then hand you three things:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Who checks what today:<\/strong> team by team, and how long each check takes<\/li>\n\n\n\n<li><strong>The errors that matter:<\/strong> for your human in the loop AI agents, in the words of the people who catch them<\/li>\n\n\n\n<li><strong>A working agent:<\/strong> with its checker named from the start<\/li>\n<\/ul>\n\n\n\n<p>Hours are measured against where you started, with checking time counted, so every saving can be checked.<\/p>\n\n\n\n<p><strong>Book a walkthrough: every team heard, one named checker, one working agent.<\/strong><\/p>\n\n\n\n<p>Organisations join in batches, so a place now means your teams are interviewed in the next round. <a href=\"https:\/\/growthnation.ai\/\">Join the GrowthNation waitlist<\/a>.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"aioseo-frequently-asked-questions-56\">Frequently Asked Questions<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"aioseo-what-are-human-in-the-loop-ai-agents-57\">What are human in the loop AI agents?<\/h3>\n\n\n\n<p>An agent with a human in the loop pauses before chosen actions, such as sending, paying or deciding, so a named person can approve, change or stop them. It waits for that person before it acts.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"aioseo-what-is-the-difference-between-human-in-the-loop-and-human-on-the-loop-59\">What is the difference between human in the loop and human on the loop?<\/h3>\n\n\n\n<p>In the loop, the agent waits for a person before it acts. On the loop, the agent acts on its own and a person monitors it, stepping in if something looks wrong. Many organisations use both, depending on how costly a mistake would be to undo.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"aioseo-which-actions-should-an-ai-agent-need-a-persons-approval-for-61\">Which actions should an AI agent need a person&#8217;s approval for?<\/h3>\n\n\n\n<p>Anything that moves money, commits the organisation to a customer or supplier, sends something outside the organisation, or decides something about a person. A human-in-the-loop step also suits new work, until the agent&#8217;s record is known.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"aioseo-how-do-you-stop-approvals-becoming-a-rubber-stamp-63\">How do you stop approvals becoming a rubber stamp?<\/h3>\n\n\n\n<p>To stop human in the loop AI agents becoming a rubber stamp, give the checker time and the authority to say no, record every change they make, and question months with no changes. Plant known errors now and then, and have checkers decide some cases before seeing the agent&#8217;s answer.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"aioseo-is-human-oversight-of-ai-agents-a-legal-requirement-65\">Is human oversight of AI agents a legal requirement?<\/h3>\n\n\n\n<p>No single UK law requires it for every agent. UK GDPR places conditions and safeguards on solely automated decisions with legal or similarly significant effects, and the ICO expects any human involvement to be meaningful rather than token. In the EU, from 2 December 2027, the AI Act requires deployers of high-risk systems listed in Annex III to assign oversight to people with the necessary competence, training and authority. This is general information, not legal advice.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"aioseo-who-is-accountable-when-an-approved-ai-agent-gets-something-wrong-67\">Who is accountable when an approved AI agent gets something wrong?<\/h3>\n\n\n\n<p>The organisation that runs the agent, and the named person who signed the work off. Approving without checking does not move responsibility to the agent or its supplier.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Quick Answer: Human in the loop AI agents pause before the actions that matter so a person can approve, change or stop them. That pause only protects anyone if the person still knows the work, has time set aside to check and can say no without asking permission. Record what they change, and test now [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":4365,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"content-type":"","_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[16],"tags":[],"class_list":["post-4363","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-thoughts"],"aioseo_notices":[],"jetpack_featured_media_url":"https:\/\/i0.wp.com\/growthnation.ai\/insights\/wp-content\/uploads\/2026\/11\/human-in-the-loop-ai-agents-featured.webp?fit=1672%2C941&ssl=1","jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/growthnation.ai\/insights\/wp-json\/wp\/v2\/posts\/4363","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/growthnation.ai\/insights\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/growthnation.ai\/insights\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/growthnation.ai\/insights\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/growthnation.ai\/insights\/wp-json\/wp\/v2\/comments?post=4363"}],"version-history":[{"count":1,"href":"https:\/\/growthnation.ai\/insights\/wp-json\/wp\/v2\/posts\/4363\/revisions"}],"predecessor-version":[{"id":4364,"href":"https:\/\/growthnation.ai\/insights\/wp-json\/wp\/v2\/posts\/4363\/revisions\/4364"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/growthnation.ai\/insights\/wp-json\/wp\/v2\/media\/4365"}],"wp:attachment":[{"href":"https:\/\/growthnation.ai\/insights\/wp-json\/wp\/v2\/media?parent=4363"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/growthnation.ai\/insights\/wp-json\/wp\/v2\/categories?post=4363"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/growthnation.ai\/insights\/wp-json\/wp\/v2\/tags?post=4363"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}