{"id":1972,"date":"2026-06-14T11:28:17","date_gmt":"2026-06-14T15:28:17","guid":{"rendered":"https:\/\/carleton.ca\/cipser\/?p=1972"},"modified":"2026-06-14T11:28:18","modified_gmt":"2026-06-14T15:28:18","slug":"the-accelerated-swivel-chair-humans-spinning-in-the-ai-loop","status":"publish","type":"post","link":"https:\/\/carleton.ca\/cipser\/2026\/the-accelerated-swivel-chair-humans-spinning-in-the-ai-loop\/","title":{"rendered":"The Accelerated Swivel Chair – Humans spinning in the AI loop"},"content":{"rendered":"\n
\n
\n\n
\n
\n
\n

\n The Accelerated Swivel Chair – Humans spinning in the AI loop\n <\/h1>\n \n \n \n\n

Daksha Bhasker | June 14, 2026<\/p>\n\n\n <\/header>\n\n <\/div>\n\n <\/div>\n\n <\/div>\n<\/section>\n\n\n\n

\"Evolution<\/figure>\n\n\n\n

The Accelerated Swivel Chair – Humans spinning in the AI loop.<\/h3>\n\n\n\n

For roughly six decades, across nearly every industry vertical, enterprise technology has attempted to eliminate swivel chair operations between disconnected systems.<\/p>\n\n\n\n

Same Chair, Different Decade<\/h4>\n\n\n\n

Swivel chair operations occur when a human becomes the integration layer between disconnected systems. Humans manually move information, context and decisions between systems that cannot communicate directly. Information is received from one system, interpreted by a person, transformed mentally or manually, and then re-entered into another system or multiple systems. The human often performs semantic normalization, syntax customization, triage, prioritization, enrichment, classification, workflow routing while maintaining contextual continuity across multiple applications. This commonly means swiveling between terminals, physically switching between devices or applications, documents, spreadsheets, dashboards and software systems. More broadly, swivel chair operations describe any workflow where humans maintain coherence across fragmented software environments.<\/p>\n\n\n\n

The human is an integral part of the workflow.<\/p>\n\n\n\n

The Wobble in the Wheel<\/h4>\n\n\n\n

Weaknesses of swivel chair operations have been understood for decades. Human error is common in manual data transfers: transcription errors, forgetting context mid-pivot, misinterpreting values between different semantic models etc. When human operators, context switch between systems, semantics and workflows, the cognitive load increases and decision quality degrades. In addition, high-volume swivel chair work produces predictable error patterns. When human operators sustain fast operational load, they increasingly rely on heuristics, rather than analytical processing producing predictable misclassification and judgement failures.<\/p>\n\n\n\n

 These are not usability problems. Nor are these problems fixed by training or process discipline. They are limitations inherent to systems where humans are integrated into system workflows.<\/p>\n\n\n\n

We Thought We Fixed This<\/h4>\n\n\n\n

The industry spun up various solutions to eliminate the human mediated integration where possible. APIs, enterprise service buses, workflow orchestration platforms, event-driven architectures, robotic process automation and modern SaaS integration ecosystems all emerged from the same underlying objective: reduce latency, improve consistency, increase auditability and remove human bottlenecks from operational workflows. Swivel chair operations are inefficient, slow and structurally fragile.<\/p>\n\n\n\n

The Loop Strikes Back<\/h4>\n\n\n\n

This is what makes the current AI landscape with Human-In-The-Loop (HITL) and Human-On-the-Loop (HOTL) particularly ironic. Modern AI systems, the most consequential technology revolution of our times, is increasingly putting humans back into workflows that the industry has spent decades trying to automate away. <\/p>\n\n\n\n

The term Human-In-The-Loop (HITL) emerged from control systems, aerospace, military systems, and simulation environments where human operators remained part of a feedback loop despite increasing automation. In AI, HITL is implemented to oversee AI outputs that could be probabilistic or potentially error prone. In HOTL, the system acts autonomously while the human supervises and can override AI output.<\/p>\n\n\n\n

HITL\/HOTL is commonly implemented in AI systems through several patterns. As a decision point, humans adjudicate high-risk actions, review low-confidence, ambiguous outputs, and provide override or kill-switch controls during anomalous behaviours or hallucinations. For exception management, anomalies, edge cases and threshold breaches are escalated for human triage. From an audit perspective, humans perform statistical sampling, spot-checking, compliance verification and drift monitoring to ensure ongoing oversight. Humans provide feedback into AI models, with human labelling, corrections and feedback loops used to retrain and recalibrate AI behaviour over time.<\/p>\n\n\n\n

Most enterprise AI systems that claim to be \u201cautonomous\u201d still implement multiple layers of HITL or HOTL controls. This is especially true, in regulated, customer-facing, financial, safety-critical or reputationally sensitive environments. In many cases, it is simply the responsible business practice given the nature of AI systems that feature model hallucination and drift over time, besides anomalous behaviours. The result is that humans once again become the operational middleware layer between systems, decisions and workflows. Only now at machine speed and machine scale.<\/p>\n\n\n\n

This matters because AI increases throughput without necessarily removing dependency on humans. Older swivel chair systems were limited by how fast people could process information manually. AI systems now generate thousands of recommendations, alerts, summaries and decisions per hour. If humans still need to supervise, validate or correct large portions of that output, then the underlying problem has not disappeared. The swivel chair has simply been accelerated with humans spinning faster than ever in the HITL\/HOTL loops.<\/p>\n\n\n\n

Keeping Humans in the Panic Loop<\/h4>\n\n\n\n

AI solutions are often justified based on speed, scale and complexity of tasks. HITL\/HOTL reintroduces human error at precisely the moments of highest consequence in AI workflows, in the escalated, ambiguous, and high-risk decisions. To make matters worse, these aren\u2019t routine tasks that swivel chair operations were typically deployed for in the past. The human in the AI workflow will be dealing with items that are time-pressured, cognitively demanding, poorly documented for the human reviewer (sometimes generated by AI black box), and outside the reviewers\u2019 domain context.

HITL\/HOTL does not mitigate AI error, hallucinations, AI inefficiencies or anomalies. It relocates them.As a result, it introduces its own compounding failure patterns. Figure-1 below illustrates how AI errors and human errors compound and circulate through the HITL\/HOTL loop, with approved outputs feeding undetected errors back into AI systems.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Figure 1:Human-in-the-loop: Loop of Compounding Errors<\/p>\n\n\n\n

AI amplifies automation bias where humans over-trust AI outputs, rubber stamping rather than genuinely reviewing all details. The HITL becomes a nominally placed human in the loop. There is context collapse where AI operates across thousands of data points, while the human reviewer only sees a summary or the condensed output. This lack of visibility and human limitation of processing vast swaths of data to validate the AI output makes verification structurally impossible, leading to errors in human decision. In addition, humans simply cannot actively supervise autonomous systems for extended periods without attention degrading. AI system designed to rely on \u201cone fatigued human at 2am\u201d is neither reliable nor truly autonomous.<\/p>\n\n\n\n

In fully automated systems, errors are systematic and traceable. When humans enter the loop at critical junctures, errors become unpredictable, hard to audit and carry the false legitimacy of human judgement. In addition to compounding errors, HITL\/HOTL introduces degradation in AI workflows. Human latency replaces machine speed. Human bottlenecks replace volume handling. And human reviewers operating under time pressure with partial context add their own errors to AI systems, workflows and outputs.<\/p>\n\n\n\n

The industry spent six decades engineering humans out of the integration workflows in systems. HITL\/HOTL engineers humans back in by design, at machine speed and scale. The swivel chair has reincarnated as Human-in-the-Loop, spinning faster than ever in AI systems.<\/p>\n\n\n\n

Escaping the AI Swivel Chair<\/h4>\n\n\n\n

HITL does not eliminate operational risk in AI. It redistributes and can compound it through human supervision, intervention and error. Human oversight also does not scale linearly with AI speed, scale or complexity. As organizations deploy increasingly \u201cautonomous\u201d AI systems, the real measure of autonomy will be the outcome of how dependent those systems remain on humans to operate safely and reliably. At the same time, it can be expected that new technologies and operating models will emerge to reduce the growing complexity and overhead introduced by HITL\/HOTL AI workflows.<\/p>\n\n\n\n

Author\u2019s note: <\/em><\/strong>Opinions expressed in this post are the author\u2019s and not necessarily those of her employer. Daksha is one of our Editors\/Reviewers for the Pulse & Praxis Journal for Critical Infrastructure Protection, Security and Resilience. <\/p>\n\n\n\n

Header Image and Figure 1 AI generated through various prompts provided by the author. <\/p>\n\n\n\n

<\/p>\n","protected":false},"excerpt":{"rendered":"

The Accelerated Swivel Chair – Humans spinning in the AI loop. For roughly six decades, across nearly every industry vertical, enterprise technology has attempted to eliminate swivel chair operations between disconnected systems. Same Chair, Different Decade Swivel chair operations occur when a human becomes the integration layer between disconnected systems. Humans manually move information, context […]<\/p>\n","protected":false},"author":340,"featured_media":1973,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":"","_links_to":"","_links_to_target":""},"categories":[24],"tags":[64,65],"class_list":["post-1972","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-nc-cipser","tag-artificial-intelligence","tag-humans-in-the-loop"],"acf":{"cu_post_thumbnail":""},"_links":{"self":[{"href":"https:\/\/carleton.ca\/cipser\/wp-json\/wp\/v2\/posts\/1972","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/carleton.ca\/cipser\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/carleton.ca\/cipser\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/carleton.ca\/cipser\/wp-json\/wp\/v2\/users\/340"}],"replies":[{"embeddable":true,"href":"https:\/\/carleton.ca\/cipser\/wp-json\/wp\/v2\/comments?post=1972"}],"version-history":[{"count":5,"href":"https:\/\/carleton.ca\/cipser\/wp-json\/wp\/v2\/posts\/1972\/revisions"}],"predecessor-version":[{"id":1979,"href":"https:\/\/carleton.ca\/cipser\/wp-json\/wp\/v2\/posts\/1972\/revisions\/1979"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/carleton.ca\/cipser\/wp-json\/wp\/v2\/media\/1973"}],"wp:attachment":[{"href":"https:\/\/carleton.ca\/cipser\/wp-json\/wp\/v2\/media?parent=1972"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/carleton.ca\/cipser\/wp-json\/wp\/v2\/categories?post=1972"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/carleton.ca\/cipser\/wp-json\/wp\/v2\/tags?post=1972"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}