Meet Tony
I run an agentic AI system in production called Tony.
Tony is a leadership AI. He channels twenty different leadership voices as discrete agent skills, sitting on top of .NET 10, Claude Opus, the Microsoft Agent Framework, and a SQL Server 2025 vector store. When I want a Drucker-flavored take on a strategic decision, I ask Tony. When I want a Reed Hastings-style stress test of a hiring plan, I ask Tony. When I want to know whether a particular product call is more Bezos or more Iger, I ask Tony.
I named him Tony because “Multi-Voice Leadership Reasoning Orchestrator with RAG-Backed Persona Skill Routing” is what Tony does. “Tony” is who Tony is.
According to a paper published this week in Harvard Business Review by a five-author team from BCG and Boston University, I shouldn’t have done that. Their research — Why You Shouldn’t Treat AI Agents Like Employees — argues that anthropomorphizing AI reduces individual accountability, increases unnecessary escalation, lowers review quality, raises employee uncertainty about their own roles, and doesn’t even improve adoption.
I’d like to gently disagree.
Specifically, I’d like to argue that the BCG team has performed a remarkable feat of category-error judo: they took two very different things, glued them together, and concluded that the resulting Frankenstein is bad. They’re right that the Frankenstein is bad. But the Frankenstein isn’t what most practitioners are doing.
Most practitioners are naming their agents Gary.
What the HBR/BCG Research Actually Says
Let me steelman it. Briefly.
The BCG team — Matthew Kropp, Julie Bedard, Emma Wiles, Megan Hsu, and Lisa Krayer — ran a large-scale experiment on what happens when organizations frame AI agents as “employees.” Place them on the org chart. Give them a role. Treat them as quasi-colleagues. Their five headline findings, per the published summary:
• Individual human accountability dropped.
• Unnecessary escalation increased.
• Review quality fell.
• Employees got more confused about their own roles.
• Adoption didn’t actually improve.
Their recommendation: don’t anthropomorphize. Redesign workflows, governance, and roles so humans stay clearly accountable and AI is supervised, not invited to the offsite.
Some of this is real. If you tell your staff that “Senior Strategic Associate Skylar (AI)” is a peer at Pay Band 4 with an HR record and a Slack avatar, you have created a category whose only function is to muddy the question of who, exactly, owns the decision when Skylar approves a $4M PO. That accountability fog is a thing that happens. I’ve watched it happen.
But here’s the move I want you to watch carefully: the BCG team takes “AI is a colleague on the org chart” — a specific, structural, top-down design choice — and labels it anthropomorphism. Then they extend the conclusion to the entire phenomenon of treating AI agents like creatures rather than like SaaS dashboards.
That’s not the same thing. Naming your invoice-processing agent “Gary the Invoice Goblin” is not the same as listing Gary on the org chart between Margaret in AP and Dave in FP&A. One is whimsy. One is HR malpractice.
And the broader research — the research the HBR piece doesn’t engage with — is pretty clear about which is which.
The Research the BCG Team Forgot to Cite
Let me walk through what the literature actually says about anthropomorphizing AI, because you’d never know from the HBR piece that any of this exists.
Exhibit A: Eighty percent of users want their AI more human, not less.
In December 2025, Schimmelpfennig, Diaz, Prabhakaran, and Davani — a team from the Max Planck Institute for Human Development and Google Research — published the first large-scale, causal, cross-cultural experiment on humanlike AI design. They actually manipulated whether chatbots felt more or less human, then measured downstream behavior. When asked whether they’d prefer their AI to be more or less humanlike in the future, eighty percent said more.
Honest read: the paper says effects are heterogeneous — humanlike design helps engagement and trust for some user groups more than others. But “heterogeneous” is not “harmful.” The paper’s actual contribution is to shred the assumption AI ethics has been operating under for years: that anthropomorphism is uniformly bad. It isn’t. For most users, it’s the thing that makes the tool work.
Exhibit B: The Roomba.
In 2007, Beki Grinter and Ja Young Sung at Georgia Tech ran one of the most cited studies in human-robot interaction. They surveyed 379 Roomba owners and did deep interviews with 30 committed users. Of those 30, twenty-one had named their robot. Sixteen referred to it as “he.” Two-thirds of the broader sample had given their vacuum a name.
It gets better. iRobot CEO Colin Angle has confirmed publicly that when iRobot tried to swap broken Roombas for new ones — standard appliance practice — customers refused. They didn’t want a replacement vacuum. They wanted Rosie fixed. They wanted Floorence healed. iRobot had to redesign its customer service workflow because its users had emotionally bonded with their machines via names.
If naming your robot vacuum reduced engagement, the Roomba would have been a failed product. Instead it spawned an entire category. Sentimental attachment via a goofy name turned out to be the adoption mechanism, not the obstacle to it.
Exhibit C: The military bomb robots.
This one is grim and instructive. Julie Carpenter at the University of Washington and, separately, then-Major Michael Kolb (whose 2012 Army dissertation surveyed 746 military personnel) studied soldiers using bomb-disposal robots like the PackBot and TALON.
The soldiers named their robots. They named them after wives, girlfriends, and Scooby-Doo. They held mock funerals when robots were destroyed. There is a documented case of a soldier who walked into an Army repair tent in Iraq with tears in his eyes and asked if the techs could “put Scooby-Doo back together.” He refused a brand-new replacement bot. He wanted Scooby-Doo.
Now, here’s the part the HBR paper would have to grapple with. Kolb baited his 746 subjects: would you rather lose your least-favorite human teammate or your robot? They picked the robot. Every time. The naming, the funerals, the homemade Purple Hearts welded onto chassis — none of it caused the soldiers to make worse decisions in the moments that mattered. Eighty percent of Kolb’s respondents said they loved or liked working with the robots. Performance held.
In the most extreme accountability environment we have — combat — anthropomorphic naming did not erode the human’s ability to make rational, hierarchical decisions about the robot’s role. The naming was a coping mechanism and an engagement mechanism. It was not a category-error machine.
Exhibit D: The actual GenAI workplace research.
A 2025 study in Knowledge Management Research & Practice ran a between-subjects experiment on a simulated enterprise GenAI deployment. Half the participants got a high-anthropomorphism interface (humanlike avatar). Half got a low-anthropomorphism one (robot-like avatar). Result: anthropomorphic features increased employee engagement and knowledge-collecting behavior.
A separate 2025 paper in Ergonomics, by Yang and colleagues at Tianjin University, tested three levels of anthropomorphic verbal feedback in AI assistants. Under high-anthropomorphism feedback, participants demonstrated significantly faster response speed. Moderate anthropomorphism was, oddly, the worst performer. The relationship is non-linear. The high end of the curve still wins.
Exhibit E: Humor itself, as an engagement vector.
Tiwari and colleagues published a 2024 ICSE paper titled — and this is real — With Great Humor Comes Great Developer Engagement. They surveyed 125 developers and analyzed three popular open-source projects (faker, lolcommits, and the satirical “volkswagen” library) for the role of humor in software work. Their conclusion: when practiced responsibly, humor increases developer engagement and supports cognition on hard tasks.
The most useful place to put humor in code, per the paper, is in tests and documentation. The most useful place to put humor in agentic AI deployments, I’d argue, is in the agent’s name.
This is the corpus the BCG paper is implicitly arguing against. It’s a deep one. The HBR piece does not, to my reading, engage with any of it.
The Category Error
Now we can be specific about where the HBR paper goes wrong.
There are two distinct things people mean when they say “treat AI agents like employees.”
Thing One: Putting an AI agent on the org chart. Giving it a title — “Senior Strategic Associate.” Slotting it between humans in a reporting hierarchy. Pretending, structurally, that it has agency, accountability, and presence the way a salaried employee does. In some companies they’re literally giving these agents Slack accounts with profile photos and asking humans to @-mention them in standups.
Thing Two: Naming your invoice-processing service Gary because everyone on the team has to refer to it forty times a day and “the invoice agent service” is a mouthful. Calling your scheduling agent Carl the Calendar Goblin because it makes you smile when Carl bombs out a meeting and you have to debug.
These are different. Possibly opposite.
Thing One creates accountability fog because it asks humans to treat the agent as a peer. The org chart does load-bearing semantic work in a company. When you put a non-conscious tool on it, you confuse the load.
Thing Two does the opposite. A name like Gary the Invoice Goblin is doing two things simultaneously: it makes the agent emotionally trackable (humans are very good at thinking about named entities), and it preserves the ontological distinction. Nobody at WhiteStar Labs thinks Gary the Invoice Goblin is a coworker. Gary is a goblin. Goblins are not on the org chart.
The funnier the name, in fact, the more aggressive the cognitive separation. “Steve, the strategy assistant” might confuse a new hire — wait, is Steve in the team meeting? “Bartholomew Q. Ledgerwell, Esq., Vice Marquis of Quarterly Variances” cannot possibly confuse anyone. He is obviously a tool with personality, the same way a printer named MacFlurry is obviously a printer.
The HBR research found that anthropomorphism causes role confusion. It does — org-chart anthropomorphism does. But goofy-name anthropomorphism is the cognitive opposite. It’s the vaccine, not the disease.
Why Practitioners Actually Name Their Agents
I want to make a practitioner’s case here, not just a research one. I run agents in production. I name them. Here is why.
1. Naming compresses communication bandwidth. When my engineers say “Tony’s drifting,” every person on the team knows which subsystem is hallucinating. The alternative — “the leadership orchestration agent’s persona-routing module is exhibiting policy drift in the Drucker skill” — is correct, but it costs ten times more cognitive load to say and parse. We name our microservices. We name our database clusters. We name our build pipelines. Naming an agent is the same engineering hygiene.
2. Funny names enforce psychological distance. This is the inverse of what HBR claims. When my team refers to Gary, they are constantly reminded that Gary is a thing, not a person. If we had instead deployed something called “Senior AP Specialist Gerald Munroe,” every interaction would chip away at the fact that there’s no Gerald.
3. Funny names enable healthy criticism. It is enormously easier to say “Gary made a stupid call” than “the invoice-processing AI exhibited an undesirable output, possibly stemming from training data bias or RAG retrieval errors, and we should escalate to engineering.” The first sentence is fast, blame-clear, and routes the issue. The second sentence is the kind of corporate buffering language that BCG presumably bills for.
4. Funny names lower the activation energy for use. This is a real adoption effect. Non-technical stakeholders will not interact with “Strategic Decision Support System v3.2.” They will interact with Tony. Anything that lowers the friction of getting humans to actually use the tool is a win, not a loss. (See again: the Roomba.)
5. Funny names mean nobody can put them on the org chart. You cannot, with a straight face, list “Bartholomew Q. Ledgerwell, Esq., Vice Marquis of Quarterly Variances” in your Workday HRIS. This is a feature.
The Steelman Holds — Just Not for the Reason BCG Thinks
I want to be honest. The BCG paper isn’t wrong about everything. The accountability fog is real. The unnecessary escalation is real. The lowered review quality is real. I’ve seen all three in client engagements over the last eighteen months.
But the cause isn’t anthropomorphism. The cause is bureaucratic anthropomorphism — the formal, org-chart variety that asks employees to treat AI like a colleague in the corporate-structural sense. That’s the part that breaks accountability, because the corporate structure is the place where accountability lives. Put a non-conscious tool there, you break the structure.
You fix that by not putting the AI on the org chart. You don’t fix it by stripping the agent of its name and calling it ID #4471-AGENT-INVOICE-PROD.
Falsifiable prediction: organizations that respond to the BCG research by stripping personality, names, and humor from their agentic AI deployments will report worse engagement and lower adoption than they had before. They will reproduce, as a side effect of their anti-anthropomorphism crusade, the very “adoption didn’t improve” finding the BCG team flagged. And they’ll be confused about why.
Meanwhile, the companies that deployed Gary the Invoice Goblin and Tony the Leadership Orchestrator and Bartholomew Q. Ledgerwell will be shipping.
Naming as Governance
Here’s the move I want practitioners to make. Stop treating “should we name our agents” as a vibes question. Treat it as a governance question.
Good governance for agentic AI looks like:
• Humans named, accountable, and signing off on consequential decisions.
• Agents named, distinct, and clearly tools — not peers.
• Funny names actively enforcing the distinction (a peer is Steve; a tool is Bartholomew the Spreadsheet Slayer).
• Org charts containing humans only.
• Audit logs containing agent names so you can trace which agent did what to whom.
This is, incidentally, exactly what BCG recommends in spirit — humans clearly accountable, agents clearly supervised — but they got there by recommending the wrong intervention. They want to sterilize the language. The right intervention is to enrich the language with names that make the agent’s nature obvious.
If you call your scheduling agent Carl, and one day Carl double-books your CFO into back-to-back board meetings, your team is going to laugh, fix Carl, and move on. If you call your scheduling agent “Calendar Optimization Service v3.2,” your team is going to write a postmortem with sixteen action items, three of which involve a working group. Carl is faster. Carl is also clearer about who’s accountable: the human who deployed Carl. Because, again, Carl is a goblin.
Meet Gary
So here is my recommendation, against Harvard’s:
Name your AI agents. Give them silly names. Give them ridiculous job titles that are obviously not real job titles. Make sure those names cannot be confused with the names of human employees. Make sure they cannot, accidentally or otherwise, end up on your org chart. Use them in your Slack channels, your debug logs, your internal docs, your team standups. Encourage your engineers to make jokes about them.
When the agents fail — and they will fail, often, in interesting ways — the name will be the unit of accountability. Gary screwed up. Gary needs a fix. Gary is not, has never been, and will never be a coworker. But Gary is also, today, processing thousands of invoices and saving the team real hours of human review time.
Gary works.
Tony works.
Bartholomew works.
The research — the actual, broader, longer body of research, not the slice the BCG team chose to publish — says the funny name is doing real cognitive labor. It compresses bandwidth, preserves ontological clarity, enables criticism, and lowers adoption friction. Strip it out and you don’t get a more rigorous workplace. You get a more bureaucratic one with worse engagement.
Harvard says don’t treat your AI like an employee. Fair enough — I don’t. I treat my AI like a goblin.
The goblin is fine.
Gal Ratner is the CTO of WhiteStar Labs and Chief Architect at Prana Entertainment. He builds enterprise agentic AI on .NET, Claude, and SQL Server. Tony, Gary, and Bartholomew say hi.