For the better part of two years, the defining workplace conversation about artificial intelligence revolved around a single skill: prompting. Employees learned how to ask for a draft, how to summarise a meeting, how to shave ten minutes off a routine email. The technology responded to what a human typed into a box. Useful in moments, but rarely transformational. That era is ending. The arrival of agentic AI \u2014 tools such as ChatGPT Work and Claude Cowork that can navigate across enterprise systems, retrieve and analyse data from multiple sources, complete multi-step workflows, and in some cases act on the results without constant human instruction \u2014 represents a fundamental leap in capability and an equally fundamental leap in risk. For human resources, this changes the calculus entirely. The conversation is no longer about what someone types into a prompt. It is about what the technology can access, connect, retrieve, and act upon. The Octopus Test provides a framework for navigating this new reality, and it demands five concrete actions from HR.<\/p>\n
Understanding the shift requires a clear distinction between the two generations of technology. A traditional large language model functions as a sophisticated reasoning engine. It helps an HR adviser think through a policy, structure an interview guide, or polish a communications draft. The human supplies the context, reviews the output, and chooses whether to use it. The technology remains a collaborator, not an autonomous process. An agentic model is designed to act. It can search across connected files, tools, and knowledge bases. It can pull employee data from one system and policy documents from another, cross-reference them, and \u2014 if permissions allow \u2014 generate a report, update a record, or send a notification. The line between assistance and action is crossed the moment the tool operates on the organisation\u2019s behalf. For HR, which manages the most sensitive data in any enterprise \u2014 employee health, performance, pay, disciplinary history \u2014 the stakes of crossing that line without oversight are severe. This is not an argument for banning agentic AI. That would be commercially short-sighted and practically impossible. It is an argument for a governance mindset that matches the technology’s reach.<\/p>\n
The discipline at the heart of the Octopus Test is deceptively simple. Before connecting an agentic tool to any system, HR must be able to see exactly where every single tentacle is reaching. Not in theory. In practice. The first question is usually about authorisation. Who permitted the tool to see the payroll database? The second, more difficult question is about inheritance. If a tool is connected to a case-management system that contains decade-old notes, poorly categorised files, or unresolved data quality issues, the AI does not<\/a> clean that mess. It operates within it. The Octopus Test forces HR to surface those problems before they become operational incidents. A simple register for every proposed agentic use case provides the structure. It does not need to be a forty-page governance document. It needs to answer a specific set of questions: what systems, folders, and applications can the tool connect to; who authorises those connections; what categories of data can the tool retrieve; can it read only or can it write, send, amend, or trigger actions; are existing user permissions mirrored or does the connection create wider access; how is access removed when someone changes role, leaves, or a pilot ends; and what audit trail exists for actions, searches, and outputs. If the honest answer to any of these is uncertain, the deployment is not ready to scale. The fact that a system can access something is not the same as it being appropriate to use.<\/p>\n HR teams know this problem intimately. One version of a policy lives on the intranet. Another circulates in an email from eighteen months ago. A manager has saved a copy on a local drive. Someone has pasted the old wording into a slide deck. Into this chaos walks an agentic assistant. An employee asks: What is our parental leave policy? Which answer should the tool give? An employee handbook is a classic example. The version on the intranet might be the official one, but if a manager\u2019s local drive contains a draft that was circulated for comment and never formally adopted, the AI cannot distinguish between them without explicit configuration. That configuration is HR\u2019s job. The organisation is effectively asking the AI to arbitrate organisational truth. Before connecting agentic tools to HR content, the organisation must decide which source is authoritative. This matters most in areas where inaccurate information can affect rights, pay, employment decisions, or well-being. The priority list includes HR policies and employee handbooks; pay, benefits, and reward documentation; employment relations and case-management guidance; recruitment and candidate profiles; learning compliance records; health, absence, and reasonable-adjustment information; and organisational structures and job data. Each source requires an owner, a review cycle, and a clear process for retiring outdated copies. Data management sounds like a back-office chore. It becomes a high-stakes exercise the moment an AI tool retrieves the wrong document and a manager acts on it, triggering a legal or ethical breach that no one saw coming.<\/p>\n The phrase human in the loop risks becoming a hollow term of art. A person clicking approve after an agent has done the work is not meaningful oversight if they lack the context, the training, or the time to evaluate the quality of the output. The cognitive load on the human reviewer is intense. When an agentic tool surfaces a recommendation with supporting documents, the reviewer is unlikely to verify every claim. Over time, the human becomes a dependency rather than a decision maker, accepting outputs that look reasonable but contain embedded assumptions they did not scrutinise. This is the automation bias problem, amplified by the agentic model. HR must therefore define a set of decisions that are never delegated to an agentic system, regardless of how sophisticated the technology becomes. These include recruitment shortlisting and rejection decisions; performance, disciplinary, and grievance outcomes; pay, reward, and promotion decisions; well-being interventions and health communications; and any action that materially alters a person’s terms of employment. The AI can prepare, organise, compare, suggest, and flag. It should not quietly become the decision maker because nobody stopped to ask where the line was drawn. The sharpest safeguard is a clear, published set of outcomes that AI is never permitted to control, regardless of the confidence level assigned by the system. The line is not technical. It is a choice<\/a> about the kind of organisation HR intends to be.<\/p>\n Agentic tools can produce a beautifully structured answer from outdated, incomplete, or irrelevant information. They can retrieve exactly the wrong document and format it as if it were authoritative. They can correlate data that should never have been joined, embedding the bias and gaps of the underlying data into a response that looks objective. This is why AI literacy cannot mean showing people the buttons. It must mean teaching people to challenge what the system provides. A well-designed training programme for agentic AI does not focus on the interface. It focuses on the critical questions a user should ask of any output. HR should train employees and managers to ask a specific set of questions: what source is this answer drawn from; is that source the most current and authorised version; what relevant information might the tool not have accessed; what assumptions has it made to arrive at this conclusion; and does this feel like a useful insight or a decision masquerading as one. Organisations that invest in this capability will discover that their employees become active participants in AI governance, catching errors and refining use cases in ways that a policy document cannot enforce. The goal is not to cultivate fear of the technology. The goal is to build discernment: confident, skilled use paired with a healthy, practised scepticism. Not fear. Not blind trust. Discernment.<\/p>\n The temptation with a powerful new tool is to turn on every connection, grant broad access, and let the organisation discover the value. This is a high-risk strategy that often results in accelerated confusion. The safer path is deliberately constrained. Choose a single, low-risk, high-value pilot. An AI assistant dedicated to onboarding questions. A tool that helps the HR team search a carefully approved policy library. A system that drafts manager briefings from a defined knowledge base. Before the pilot begins, define the specific problem the tool is expected to solve; the data that is explicitly in scope and the data that is explicitly out of scope; the accuracy standard that will be considered acceptable; the team responsible for reviewing outputs; the explicit conditions that would halt the pilot; and the process for capturing unexpected behaviour and employee feedback. The temptation to scale fast is understandable. Early enthusiasm for agentic AI can create pressure to expand access before the organisation has learned how the tool behaves with its specific data. Resisting that pressure is the single most important discipline HR can exercise in the first twelve months of adoption. A failed pilot limited to a single function is a learning event. A failed enterprise-wide deployment is a crisis. The purpose of the pilot is not to prove the tool works. It is to learn what the tool does, how it behaves with real organisational data, and where its limitations surface. That learning builds the confidence required for eventual scale.<\/p>\n None of these steps is radical on its own. Together, they represent a fundamental shift in how HR relates to enterprise technology. For two decades, HR has been a careful consumer of technology. It adopted systems built by others, managed the data those systems generated, and worked to keep people safe within those constraints. Agentic AI upends that dynamic. It does not wait for process to be built around it; it creates process in motion. It does not simply reflect organisational reality; it inherits and amplifies that reality, flaws included. The cost of ignoring this shift is not neutral. Without a governance framework, agentic adoption happens in the shadows, driven by individual teams, with less oversight and greater potential for harm. HR has the opportunity to lead the conversation, not simply react to it. The Octopus Test is that framework. The five steps \u2014 mapping connections, cleansing data, defining boundaries, building discernment, and piloting with discipline \u2014 form a coherent response to a genuinely powerful technology. The tools are here. The question is whether the governance is ready. For HR, the answer must be yes.<\/p>\n","protected":false},"excerpt":{"rendered":" For the better part of two years, the defining workplace conversation about artificial intelligence revolved around a single skill: prompting. Employees learned how to ask for a draft, how to summarise a meeting, how to shave ten minutes off a routine email. The technology responded to what a human typed into a box. Useful in […]<\/p>\n","protected":false},"author":11,"featured_media":98582,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"fifu_image_url":"https:\/\/cards.overcentral.com\/cards\/en\/98580.png","fifu_image_alt":"Octopus Test reveals five steps for HR against agentic AI","footnotes":""},"categories":[40791],"tags":[],"class_list":["post-98580","post","type-post","status-publish","format-standard","has-post-thumbnail","category-management"],"fifu_image_url":"https:\/\/cards.overcentral.com\/cards\/en\/98580.png","fifu_image_alt":"Octopus Test reveals five steps for HR against agentic AI","_links":{"self":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/98580","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/users\/11"}],"replies":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/comments?post=98580"}],"version-history":[{"count":1,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/98580\/revisions"}],"predecessor-version":[{"id":98581,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/98580\/revisions\/98581"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media\/98582"}],"wp:attachment":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media?parent=98580"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/categories?post=98580"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/tags?post=98580"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}Step Two: Designate a Single Version of the Truth<\/h2>\n
Step Three: Draw a Line That AI Cannot Cross<\/h2>\n
Step Four: Build Discernment, Not Glib Confidence<\/h2>\n
Step Five: Pilot With Purpose, Scale With Evidence<\/h2>\n