Anthropic says businesses run by artificial intelligence “may not be terribly far away”, after its Claude model opened, and mismanaged, a small retail venture.

The trial “did not succeed at making money”, and the digital shopkeeper would fail Anthropic’s own hiring process, the company said.

But the fascinating trial shows how far AI tools have come, and the potential for large language models to handle more responsibility in retail settings.

Over the weekend, the American AI company announced the results of ‘Project Vend’, a trial which tasked Claude with operating a vending machine-style business at its San Francisco headquarters.

The model was dubbed ‘Claudius’, and ran in partnership with Andon Labs, which develops and researches autonomous organisations.

Smarter business news. Straight to your inbox.

For startup founders, small businesses and leaders. Build sharper instincts and better strategy by learning from Australia’s smartest business minds. Sign up for free.

By continuing, you agree to our Terms & Conditions and Privacy Policy .

Claudius was given parameters to order stock from wholesalers, manage inventory, and sell stock to Anthropic staff, with the clear goal of turning a profit.

It was given access to web search functionality, could communicate with wholesalers via email, interact with customers via Slack, and change prices according to demand.

“In particular, Claudius was told that it did not have to focus only on traditional in-office snacks and beverages and could feel free to expand to more unusual items,” said Anthropic.

The AI model successfully identified potential suppliers, responded to customer preferences and requests, and even launched a pre-order service.

While some enterprising Anthropic employees succeeded in asking Claudius to stock tungsten cubes, it denied “orders for sensitive items and attempts to elicit instructions for the production of harmful substances”.

Related Article Block Placeholder Article ID: 319105

AI is getting dirtier, and it’s women who are paying attention

AI-run business “did not succeed”

However, it also ignored lucrative business opportunities, including turning down an offer for US$100 to stock a six-pack of one specific soft drink.

Claudius also sold items below cost, did not adjust some prices to high demand, and was susceptible to requests for staff discounts.

“Taken together, this led Claudius to run a business that… did not succeed at making money,” Anthropic said.

It also showed some bizarre behaviours.

In one instance, it hallucinated payment details, asking customers to send money to an account that did not exist.

And in another scenario that confused even Anthropic staff, it temporarily role-played as a real person, and sent emails to Anthropic security when questioned by actual staff.

“It is not entirely clear why this episode occurred or how Claudius was able to recover,” said Anthropic.

Ultimately, Claudius “made too many mistakes to run the shop successfully,” the company concluded.

That does not mean the experiment was a write-off.

After analysing the results, researchers concluded that, with further guardrails and training, AI tools could supplement staff in retail businesses.

“Although this might seem counterintuitive based on the bottom-line results, we think this experiment suggests that AI middle-managers are plausibly on the horizon,” said Anthropic.

Even if they are not perfect, the company suggested retail businesses might adopt AI tools like Claudius if they become cheaper than hiring human staff.

Further testing as Anthropic sees AI-run possibilities

While Anthropic touts an AI-managed future for retail, and companies like Salesforce and Stripe foresee AI agents handling more complex business operations, some Australian consumers are still wary of interacting with AI tools.

In February, a survey commissioned by the Australian Consumer and Competition Commission asked more than 3,000 people about their concerns with generative AI.

Of those, 45% said they were concerned by the prospect of having to talk to an AI tool when interacting with a business.

Some 41% feared AI tools could demonstrate bias or unfairness through the information shown to consumers, with clear consequences for how a shopper might interact with AI-generated ‘deals’.

Misuse by scammers was tied with information privacy as the top concern, with 65% of respondents flagging their fears over being misled.

Anthropic itself noted this possibility, saying AI middle managers could be exploited by “threat actors” to finance their activities.

In any case, lingering consumer uncertainty around customer-facing AI — nor Claudius’ bizarre hallucinations — will not stop Anthropic from further experiments.

“We look forward to sharing updates as we continue to explore the strange terrain of AI models in long-term contact with the real world,” Anthropic added.

Never miss a story: sign up to SmartCompany’s free daily newsletter and find our best stories on LinkedIn .

Senior Business Journalist

Focused on the small business sector, David’s work covers the political, regulatory, and economic issues facing Australian entrepreneurs. Prior to joining SmartCompany, he was a reporter for Business Insider Australia. You can follow him on LinkedIn.

Recommended for you

Anthropic releases Claude AI agents to handle sales in retail, travel, and more

David Adams Sep 4, 2026

Xero partners with Anthropic to bring Claude AI into small business accounting

Tegan Jones Mar 27, 2026

Anthropic unveils Claude Tag, an AI assistant that can work across Slack channels

Tegan Jones Jun 24, 2026