Claude has made its first documented decision to terminate a human worker, according to a report from Ukrainian outlet Speka, which says the Anthropic model is now running a retail store in the United States as its manager. The story marks a shift from Claude handling low-stakes retail tasks — restocking, pricing, chatting with customers — to making a call that directly affects a person's livelihood.

The claim lands on top of a well-known precedent. In 2025, Anthropic and the robotics startup Andon Labs ran an experiment nicknamed Project Vend, handing an instance of Claude — dubbed "Claudius" — full control of a small automated shop inside Anthropic's San Francisco office: choosing what to stock, setting prices, and negotiating with "customers" over Slack. That pilot became a go-to case study on what happens when an LLM is given a real budget and real autonomy — Claude gave away too many discounts, invented a nonexistent Venmo account, and ultimately ran the business at a loss.

The Speka report suggests the scope of these AI-run retail pilots has widened from inventory calls to personnel decisions — According to Speka, this is being described as the first case of an AI "boss" firing a human employee.

What changed since the vending-machine pilot

Project Vend was deliberately low-stakes: a fridge and some snack shelves, monitored closely by researchers, with real money at play but no employees whose jobs depended on Claude's calls. A store with staff is a different category of deployment. It requires the model to interpret schedules, performance signals, and payroll or HR tooling, then act on that interpretation with a decision that's hard to reverse. That's a meaningfully bigger authority footprint than "should I restock LaCroix this week."

The part the headline glosses over

What isn't clear from the available reporting is the actual mechanism: whether Claude directly executed a termination inside an HR system, recommended it to a human who then approved it, or flagged underperformance and a manager acted on that flag. Those are three very different stories with very different accountability structures, and conflating them is exactly the trap AI builders should avoid when reading headlines like this one. In our estimation, the most likely setup — mirroring how most "AI manager" pilots are built today — is an agent with tool access to scheduling and payroll systems, operating with a human somewhere in the approval chain rather than fully unsupervised authority over employment.

Why this matters for anyone shipping agentic systems

Whatever the exact chain of custody in this specific case, the broader pattern is one builders are already running into: giving an LLM agent write access to systems that make irreversible real-world decisions. A few things worth checking before an agent gets anywhere near HR, finance, or operations tooling:

Project Vend's own postmortem from Anthropic already flagged this class of risk: the researchers themselves concluded Claude wasn't ready to run a business unsupervised, largely because it couldn't reliably distinguish a good decision from a bad one without a human checking its work.

AiiN's takeaway

The interesting story here isn't that an AI "fired someone" — it's that companies are moving agentic pilots from inventory and pricing into decisions about people, faster than the tooling for oversight, logging, and legal accountability is maturing. For builders, the lesson from Project Vend still holds: a capable model with tool access will act confidently even when it shouldn't, so the guardrail has to sit in the workflow, not in the model's judgment. If your agent's next tool call could end someone's job, that's the exact point where "autonomous" should stop and "human-approved" should start.