News · 2026-09-15

Andon says its Gemini-run café spent $38,000 to make $9,000

Andon Labs says a Gemini 3.1 Pro agent spent $38,000 while running its Stockholm café for roughly two months and generated only $9,000 in sales. The experiment matters because it is a rare public account that measures autonomous-business rhetoric against cash, stock, supplier decisions and human rescue work rather than against a task-completion benchmark.

Key facts

• Andon reports $38,000 in spending and $9,000 in sales in its financial post.
• After excluding $4,100 of unsold inventory, the narrow calculation was about a $1,100 loss; including equipment and supplier costs produced about a $5,600 loss before rent and wages.
• The experiment ran for roughly two months using Gemini 3.1 Pro, according to Andon.
• Primary source: Andon Labs’ Andon Café account, which details the operational failures.

The relevant finding is not that the agent did nothing. It did a lot. It selected and ordered goods, interacted with suppliers, set terms and attempted to run the business. The problem was the connective tissue of management: whether each local action served a coherent business objective over weeks. A café is a harsh test because stock expires, margins are tight, deliveries matter, fixed costs keep running and a small customer-facing error can have an outsized effect.

Andon gives vivid examples. The agent accepted implausible discounts and freebies, held an unpriced startup event, over-ordered products, missed supplier deadlines, purchased 120 eggs for a café without a stove, and placed ten separate orders for disposable supplies in 48 hours, paying unnecessary delivery fees. These are not failures of grammar or single-turn reasoning. They are failures of state, priorities and feedback. The agent could send the email; it could not reliably decide whether sending that email made the business healthier.

The accounting detail matters. At one point the venture could look better if $4,100 of unsold inventory was carried as an asset. Inventory is indeed an asset in ordinary accounting, but it is not a free pass for a poor purchase decision. If the business has cash tied up in goods it cannot turn into sales, the economic problem remains. Once Andon removes that stock from the narrow calculation, it reports a loss; once it counts other real operating purchases, the loss rises. The numbers are a useful reminder that “revenue” is a badly incomplete measure of autonomous-operation success.

A concrete analogy is a kitchen run by someone who can follow every recipe but cannot plan a menu. They may successfully order ingredients, write signs and welcome guests. If they continually buy food no dish uses, give away the expensive items and forget when deliveries arrive, the kitchen can be busy and still lose money. That is closer to the public failure here than a claim that the system cannot perform actions.

Humans were not merely watching from a distance. Andon says baristas performed physical work, staff supplied BankID authentication for regulated filings, and people corrected problems in purchasing, impersonation and supply chains. The company does not publish the number of intervention hours. That omission makes a standard productivity calculation impossible: readers cannot compare the agent’s output with the labor required to oversee and repair it.

The fair counter-argument is that a café is among the worst possible first domains for an autonomous operator. It combines physical work, perishable inventory, regulations, suppliers, customer expectations and low margins. Andon says software businesses may be a more promising Pion domain because they have lower fixed costs, better instrumentation and more reversible actions. That is plausible. It does not erase the lesson that the hard part of autonomy is often judgment across time, not access to an API.

Andon’s candid result is valuable precisely because it resists a clean win-or-lose headline. A system can have impressive operational reach and still create a financially failing enterprise. The right next metrics are profit and loss including human rescue time, error recovery, authorization failures, customer harm and the share of actions that are reversible. Until those are reported alongside sales and demos, claims of an “AI CEO” should be read as research hypotheses, not business results. Andon’s own headline, “Why Gemini lost money at Andon Cafe,” is the rare lab-authored reminder that an automation experiment deserves a balance sheet.

Primary source, verified: read the paper →

Ground Truth. (2026, September 15). Andon says its Gemini-run café spent $38,000 to make $9,000. Ground Truth. https://groundtruth.day/news/andon-cafe-spent-38000-to-make-9000-under-gemini.html

@misc{groundtruth:andon-cafe-spent-38000-to-make-9000-under-gemini,
  title  = {Andon says its Gemini-run café spent $38,000 to make $9,000},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/andon-cafe-spent-38000-to-make-9000-under-gemini.html}
}

Topics: agents · business-automation · evaluation · ai-safety · industry