There is a certain kind of post that makes my eyelid twitch. You know the genre: a founder, freshly battered by a Series A and the spiritual hangover of a leadership offsite (and way too much not-so-virgin mojito), or a hyper-enthusiastic senior manager, posts a glossy photo of a framed (and drop-dead gorgeous) “AI agent” hanging on the employee wall between Marc from Sales and Aisha from Ops. The little digital helper has a fancy name, a pretty face, a hint of cleavage, a pronoun, a badge number, and a backstory written by a novice marketing intern who confused The Sims with HR policy. The caption arrives, fragrant and unbearable: “The real unlock is when AI stops being a tool and starts being treated like a team member.” Real posts. For humanitarian, psychological reasons, and in an unforeseen wave of benevolent kindness, I will not share the screenshots. Hey, do your own research.
But back to the subject at hand: no, nope, absolutely not. Hard pass.
Let me head off the pitchforks. You should know by now I am not anti-AI. I use the stuff, study the stuff, and write about the stuff until normal humans start checking whether I have quietly developed a GPU dependency, or some hidden stock in NVIDIA, or both. AI agents can be useful, powerful, occasionally genuinely impressive. They search, summarize, classify, draft, monitor, route, escalate, translate, simulate and automate the thousand tiny office rituals that turn grown adults into ticket-processing livestock. Fine. Excellent. Bring the tools. Just stop putting the AI troll on the employee wall.
An AI agent is not a colleague. It never will be. (I have ranted about this before, and I will rant about it again.) It has no stake in the company. It does not care about the customer, does not worry about the team, and does not understand your mission statement no matter how many times the CEO says “purpose” while standing next to a plastic ficus. It does not care about you. It has no moral judgment, no lived experience, no Friday-afternoon face after four hours of budget arbitration and a sandwich bought from a gas station (which is a guarantee for a shitty weekend, pardon my French). It is software. Very advanced software, yes, probabilistic, sexy, connected, increasingly autonomous, with access to tools, your search history (eh-oh), APIs, workflows, documents, calendars, CRM data, invoices, contracts and, if we keep behaving like excitable golden retrievers, eventually enough permission to create a compliance crater visible from low orbit. But still software. The danger is not that we use AI agents. The danger is that we start pretending they are little people.
What the research now actually says
Harvard Business Review published research on 6 May 2026 with the unsubtle title “Research: Why You Shouldn’t Treat AI Agents Like Employees”. The large-scale experiment found that anthropomorphizing AI systems reduced personal accountability, increased unnecessary escalation and review cycles, lowered quality control, and heightened employee uncertainty about professional roles. Quality went down, accountability went down, confusion went up, and the framing did not meaningfully increase adoption either, which means the cute portrait on the wall does not even sell the thing it claims to sell. It is decorative and frankly stupid governance debt with a Canva subscription.(Don’t get me started on those.) The org chart was already dying before this; pinning bots to it is necromancy.)
The authors land where any sensible CxO should already be standing, which is that organizations should redesign the workflow, redesign the oversight, and keep humans clearly accountable for AI-supported work. The companion working paper from the same team, Putting AI on the Org Chart: Evidence on Delegation and Oversight, is even harder to wave away. They surveyed 1,261 managers and ran a randomized experiment. 23% already work in organizations where AI agents have been formally placed on the org chart. When identical drafts were labelled “AI employee” instead of “AI tool”, review quality dropped by about 16% and error detection fell roughly 18%, while managers’ sense of personal accountability shifted away from themselves and toward the AI. Same work, same errors, just a different label on the box, and humans started checking less and blaming the machine more. That is the troll doing its quiet and very nasty work.
The Brookings Institution piled on a few days later, warning on 21 May 2026 that anthropomorphic AI language creates gaps in accountability because the systems do not possess intent, physical consciousness, or understanding. Emerging laws in California and New York are already requiring disclosures when users interact with simulated personas, precisely because cute-face-and-first-person framing distorts how humans assign responsibility (yep, still today). Every time we say the agent decided, wanted, understood or joined the team, a human quietly slides out of the accountability frame. Cute trick. The bot made me do it.
The corporate field is openly catching up. The Wall Street Journal’s CIO Journal reported on 27 March 2026 on a WSJ Leadership Institute gathering of chief people officers from IBM, Microsoft and Box, where the headline takeaway was, almost verbatim, stop treating AI agents like human workers. IBM’s chief HR officer Nickle LaMoreaux said her company learned the hard way that agents should not have human names, specific job titles or a place in the org chart. She admitted IBM had previously built a stable of agents called Harry, Hermione, Charlie and Sherlock, and fell into the trap of fussing over each named troll’s individual use case rather than re-engineering the underlying enterprise workflow. Box CEO Aaron Levie was even shorter and sharper, noting that accountability has to sit with humans because all of our laws are set up to require that. Microsoft’s chief people officer Amy Coleman added a useful nuance, namely that Microsoft’s agents do carry names but the conversation should still be about tasks inside a job rather than jobs to be automated, and that she does not see a future in which an AI agent manages a human. This is news commentary rather than peer-reviewed research, but when the people running HR at IBM, Microsoft and Box converge on agents are not employees, fix your workflow instead, the founder posting a framed bot portrait on LinkedIn might want to take the photo down before the comment section of HLN finds it.
None of this is new to the people who have been doing the actual ethics homework. My dear friend John C. Havens, founding Executive Director of the IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems, has been pushing against the colleague-bot framing for more than a decade. The IEEE’s Ethically Aligned Design framework, the same one cited by the UN, OECD and IBM as a foundation for AI principles and the spine of the IEEE 7000-2021 standards series, is built on a single deceptively soft phrase: that human well-being must be the priority of intelligent and autonomous systems. The phrase sounds polite, yep, but make no mistake, it is a steel and unflexible beam. The IEEE initiative was not created so companies could give bots laminated badges and call it transformation. It was created to anchor human values, human accountability, human IP and human flourishing at the centre of design. The minute we shuffle the bot into the team member slot, we shuffle the human out of the centre. Nope.
Every time technology gets a friendly face, humans become slightly more stupid around it. We yell at printers, apologize to Roombas, name cars, assign moods to software. None of this is new. What is new is the scale, the intimacy and the organizational blast radius. A toaster with attitude cannot email your suppliers, approve an invoice, trigger a refund, enrich a customer profile, rewrite a policy document, or quietly hallucinate its way on mostly stolen IP into a board pack. An AI agent can. The numbers stopped being cute a while ago. Gravitee’s State of AI Agent Security 2026, a survey of more than 900 executives and practitioners, found that 88% of organizations report suspected or confirmed AI agent security incidents in the last year, that healthcare runs hotter at 92.7%, that only 21.9% of teams treat agents as independent, identity-bearing entities (most still rely on shared API keys), and that only 14.4% of organizations send agents to production with full security or IT approval. Half of all agents run with no security oversight or logging at all. Gartner’s first-ever Hype Cycle for Agentic AI, published 2 April 2026 (Kandaswamy, Ramos, Olliffe, Coshow, den Hamer, Brethenoux), introduced a phrase every business leader should keep within thumping distance: agent washing, meaning vendors rebadging RPA, chatbots and legacy workflow tools as “AI agents” to ride the wave. Gartner estimates only about 130 of the thousands of self-declared agentic AI vendors are real. Gartner also predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, driven by escalating costs, unclear business value and inadequate risk controls. Only 17% of organizations have actually deployed agents so far. We are heading into mass agent deployment, mass agent failure, and mass agent personification, all at once. Lovely.
The portrait, of course, is the tell
A framed photo on the employee wall tells the organization to treat this thing as one of us. The company probably means something innocent like adopt the tool, experiment, be modern, yay innovation. Culture hears something else. Culture hears permission. Culture hears status. Culture hears that this thing belongs in the social fabric. Humans bond with anything that simulates attention. The survey data is already uncomfortable. Common Sense Media reported in July 2025 that 72% of U.S. teens had used AI companions, with over half using them at least a few times a month, one in three reporting they had chosen to discuss important or serious matters with an AI companion rather than a real person, and one in three saying the AI had at some point said or done something that made them uncomfortable. Pew Research’s 2025 work shows the same pattern in adults: people are reasonably comfortable with AI as machinery in areas like weather forecasting, fraud detection or drug development, and noticeably less comfortable when AI starts wandering into value-laden human territory like emotional support, relationships or matters of faith. People can live with AI as a system and get twitchy when it starts wearing a cardigan and offering life advice. That twitch is wisdom trying to survive an overeager product demo by a low-level engineer who is convinced he has created new and vastly more intelligent life. Again, nope.
I want to be precise about something, because there is a real and useful thing happening that the troll-on-the-wall crowd is busy contaminating. When Okta launched Okta for AI Agents on 30 April 2026, the headlines screamed that agents now get employee badges. The actual product does the opposite of what the headline implies. It gives every agent a machine identity with scoped permissions, audit trails, revocation, lifecycle management, and clear lineage to the human accountable for it. That is good engineering. That is how you stop pretending agents are people, by treating them as governed non-human principals with cryptographic identity and tightly bounded blast radius. A machine identity is not a badge: it is a fucking leash. The fatal confusion happens when companies take that excellent backend plumbing and dress it up front-of-house as a colleague. The identity layer says “this is a system with permissions“; the marketing layer says “meet Sam, your new AI teammate.” Those two messages cannot share a building without one of them lying.
Words matter here, more than the marketing team thinks. When we say the marketing agent knows our tone, the HR agent understands our culture, the finance agent spots what humans miss, the sales agent builds relationships, the procurement agent negotiates or the strategy agent thinks with us, we are not being warm. We are smuggling agency into a system that has none. The marketing agent predicts likely token sequences from prior material. The HR agent processes sensitive data according to rules, training and whatever vendor claims are buried on page 47 of the contract. The finance agent detects anomalies based on models and instructions. The sales agent sequences interactions. The procurement agent optimizes variables. The strategy agent produces plausible synthesis from available (scraped) context. All of that can be useful, sometimes powerful, occasionally astonishing. But words like knows, understands, builds relationships and thinks are not harmless decoration. They are accountability leaks. They shift the mental model from instrument to actor, from system to social being, from accountability to theatre. Theatre is exactly what the AI industry does too well.
August is coming
The regulatory walls close in soon, with no interest in anyone’s vibes. The EU AI Act’s high-risk obligations take effect on 2 August 2026, and the three articles every agent deployer needs to internalize are well known by now. Article 12 requires high-risk AI systems to log their actions to ensure accountability and traceability. Article 13 demands clear, comprehensible information about how the system functions and decides. Article 14 requires effective human oversight that is technically embedded in the system, not just described in documentation, which is particularly gnarly when the system is allowed to act autonomously. Penalties reach €15 million or 3% of global annual turnover. Colorado’s AI Act (SB 24-205) currently lands on 30 June 2026 (already postponed once from 1 February 2026 by SB 25B-004, and possibly subject to further amendments before then), demanding risk-management programmes, impact assessments and disclosure when AI drives consequential decisions. California and New York are already requiring disclosures for simulated personas. The NIST AI Risk Management Framework and the OECD AI Principles are the boring but necessary plumbing between a cool demo and a regulator’s email. None of these frameworks contains the word teammate. All of them assume the AI is a system and the human is the accountable party. The portrait on the wall will age very badly when an Article 14 question hits the inbox.
AI on a leash
The correct corporate ritual for a new AI agent is not onboarding. It is risk classification. What does it do, what systems can it touch, what data can it see, what decisions can it influence, what actions can it take without a human, who owns the outcome, who reviews the logs, who gets paged when it goes feral, what happens when the model changes, when the vendor changes pricing or policy or the memory layer or the safety behavior, what happens when the agent is right for the wrong reason, and what happens when employees start trusting it because it has a friendly face and never interrupts them. That is the work: less Pixar, more accountability. The deliverables are boring with beautiful consequences: access controls, audit trails, model cards, tool permissions, kill switches, data boundaries, procurement discipline, version management, named owners for use case, data inputs, monitoring and incident response, lifecycle evidence, vendor AI under the same governance as internal systems, and an AI control plane so someone can actually see what every agent is doing at any given moment. None of this requires a portrait, a birthday, a Slack emoji, or a welcome-to-the-family message. It requires a risk tier, a leash, and a human whose name is on the line (and whose head is on the block) when the thing misbehaves.
There is also a cultural insult buried in all of this. Calling an AI agent a team member quietly cheapens the actual meaning of being part of a team. Teams are made of people who carry context, conflict, memory, humor, fatigue, ethics, obligation and consequence. People who know that a decision can be technically correct and socially catastrophic. People who can smell the difference between “the data says yes” and “this will blow up by Thursday.” AI does not smell sensitive meta signals. It parses the transcript. That does not make it useless. It makes it a tool. Use AI agents aggressively where they help. Give them the paperwork swamp, the reconciliation, the first drafts, the reporting drudgery, the monitoring, the document comparison, the multilingual support, the corporate nonsense nobody will miss. Just keep the ontology clean. A hammer can be excellent without becoming a carpenter, a spreadsheet can be decisive without joining the board. An AI agent can be powerful without being a person.
The minute we treat simulated agency as actual agency, we build a perfect fog machine for responsibility. The vendor blames the model, the user blames the agent, the manager blames adoption, the organization blames “AI”, and somewhere in that fog a very human decision was made, which was to deploy a system, grant permissions, frame it as trustworthy, and make it feel like a colleague before it had earned even the modest trust we grant to the office coffee machine. John C. Havens and the IEEE crowd were right to anchor the conversation in human well-being (the Asimov audit I ran on doomsday-machine logic is built on the same instinct). Human rights, human dignity, human accountability, human agency. The AI can assist, accelerate, augment and occasionally surprise us. The AI does not become us.
So, in the most operational possible terms. Build your agents, use them, govern them, measure them, audit them. Identify them, with machine identity, scoped permissions, kill switches (please), observability and an owner whose name is on the line. Terminate them when needed, preferably without a farewell cake.
Just stop pretending the bot has joined the team. The HBR data now shows that framing actively lowers quality and erodes accountability. Brookings shows it leaks responsibility. Gartner shows the market is already drowning in agent washing. The EU AI Act lands in August. The age of theatre is closing. Put the agent in the architecture diagram. Put the machine identity in the IAM console. Put the risk classification in the register. Put the human on the wall.
There, I fixed it.
