Treating AI agents like employees doesn't work — here's why

While Stuart Lauchlan recently blew a gasket over the tendency to anthropomorphize AI, it has become an increasingly popular trend among senior executives lately to consider treating agents as employees or teammates rather than just another tech tool. According to research published by the Harvard Business Review (HBR), leaders assume that anthropomorphizing the technology will make it feel less foreign to employees and will signal their AI ambitions more effectively to stakeholders. But the HBR report entitled 'Why you shouldn't treat AI agents like employees' found there were various downsides to this approach for both the workforce and the wider organisation. For instance, it revealed that going down this route can make staff feel personally less accountable. This, in turn, leads to a reduction in the quality of their review activity, costs more due to the need for additional review cycles, and erodes both professional identity and trust in the organization. It also fails to make much difference to employee intent to adopt the technology and integrate it into their workflows So, why does the notion of anthropomorphizing AI appear to be gaining so much traction and what are the wider implications of adopting this stance? Emily Rose McRae is a Senior Director Analyst at Gartner. In her view, the notion of treating agents like employees is a concept pushed by vendors in order to "access to headcount budget": > Headcount is usually one of the largest organizational expenses. Typically, there's a cap on how much is spent on technology, but if vendors can expand that into headcount budget, they can make more money. So, they'll tell you to treat agents like employees - only agents can't be put on a performance improvement plan. Another consideration here is that a return on investment on AI agent technology is "harder to find than you might think", McRae says. This means it makes sense from a supplier perspective to correlate numbers of agents directly with numbers of employees as doing so can make them appear a cheaper option. But the downside for customers in buying into this stance is that: > You still have to monitor the tool to see if it's performing and you may need to work on workflows and outputs etc. So, it isn't as simple as saying an agent will do X. It's all a vendor construct, but it feeds into how people talk about these tools too as most of the discourse is on cost saving, and headcount is the most obvious cost. To anthropomorphize or not to anthropomorphize? Tina Shah Paikeday is General Manager at AI talent management platform, Findem. Another problem with anthropomorphizing the technology, she believes, is that it "collapses categories that need to stay distinct". These are task execution, judgement and accountability: > AI agents can execute at scale, but they can't be held accountable in the way a person can. The research shows that when we borrow the language and structure of employment, such as names, org chart placement, and managers, people quietly reassign responsibility to the AI rather than to themselves. That is the opposite of what we want in a transition period where trust is already fragile. The problem is that people start saying things like 'Kevin made a mistake'. This means they treat the system as a "social actor with its own agency" rather than software that a human has deployed and is accountable for. In other words, the space in which it makes sense to anthropomorphize AI agents is narrower than most leaders might assume. As Paikeday says: > It works for well bounded, high volume, rules-based workflows, where the question of who is accountable is already answered by policy, for example first pass screening with mandatory human sign off...Giving something a name and a role can also lower the psychological barrier to adoption, especially for less technical staff, and it can be a useful shorthand internally and externally for signalling ambition. That's about it. Rick Rider is Senior Vice President of AI Innovation at cloud-based enterprise software provider, Infor. He agrees that giving an agent a name or persona can help lower the psychological barriers for people who are nervous about the technology. This is worth considering, he points out, as change management is the "hardest part of AI rollout". In every other instance though, AI should be framed and governed as infrastructure with a named human owner rather than a colleague. This is because, Paikeday says: > The framing question is not cosmetic. It's a governance decision that determines who reviews, who escalates, and who answers for outcomes. Rider agrees: > It dilutes accountability and slows down proper scrutiny. Once people start treating an agent's output the way they'd treat a trusted colleague's work, they stop double-checking it the way they should. That's dangerous because, unlike human colleagues, an agent doesn't yet carry the same weight of consequences when it gets something wrong. Keeping governance unsentimental As a result, Paikeday believes that it makes more sense to frame AI as a path to more valuable work rather than as a colleague: > Employees don't need AI to feel human to trust it. They need to know what decisions still require their judgement, what is changing about their role, and that the organization has a plan rather than a slogan. My view on the optimal approach is to treat AI as a capability we're deploying, be explicit about the human accountability wrapped around it, and spend our 'make it feel less foreign' energy on skills and role clarity instead of personification. Rider takes a similar stance: > The optimal approach is to separate the emotional layer from the operational layer. Make the tools approachable and non-threatening but keep the governance layer completely unsentimental. Internally we call this GRC for AI: governance, risk, and compliance specific to the scope any given agent is allowed to operate in. You can be warm and encouraging about experimentation while still being strict and deterministic about what the agent is actually permitted to do. But he also points out that it is still very much early days in terms of agentic AI system deployment anyway: > Most organizations aren't even close to giving agents that kind of deep integration into core business processes yet. Most of what people are doing today is still fairly surface-level: research, drafting, and decision support, where a human is still very much making the final call. The risk HBR is describing shows up when companies architecturally wire agents into decision-making without rethinking the human accountability structure round it. So, really it comes down to knowing exactly where the line of accountability sits and being careful not to blur that just because the tool feels conversational. This means it makes more sense to treat the technology as infrastructure "with judgement built in" rather than as a colleague per se, Rider believes. The upshot is that: > You govern it the way you'd govern any high-leverage system with clear scope, clear guardrails, and clear ownership sitting with a human. When re-design becomes non-negotiable If the organization does choose to treat their agents as employees though, the HBR report argues that governance, which is already important in an AI context, becomes even more critical. This means how it works will need revamping, alongside job roles, operating models, and workforce capabilities. As Paikeday points out: > The re-design is necessary regardless of whether we anthropomorphize. But anthropomorphizing raises the stakes because it actively degrades the very accountability structure we are relying on to catch errors. If we're going to use employee-like framing anywhere, we can't skip the redesign step. Skipping it compounds a governance gap with a psychological one. Rider agrees: > Once you add that emotional layer, the redesign becomes non-negotiable because you're manging how humans relate to and trust that tool.