FX.co ★ Breakthroughs and blunders of autonomous AI

Talk down auto‑attendant

One very popular use case for autonomous agents became fighting persistent subscriptions and call‑center robots. Users send voice AI agents to support lines of providers and insurers. An agent can stay on the line for 40 minutes, work through complex IVR menus, debate an auto‑responder, and insistently demand a human operator until a paid service is canceled or a discount is officially confirmed. Algorithms have no irritation or fatigue, turning endless hold time into the consumer’s perfect weapon.

Undo nonrefundable

One user tasked Dots to cancel a nonrefundable hotel booking and plane tickets — a task that even experienced lawyers often balk at. The AI drafted a polite letter to the hotel and, when a quick reply did not come, escalated autonomously to the chain’s general manager. After a series of persuasive messages from the agent, the penalty was fully waived. Moreover, while analyzing card statements, the agent spotted an erroneous $210 charge and automatically processed a refund through support. The persistence of algorithms broke bureaucratic barriers across the service sector.

Catch train

Another amusing example of agent logic involved a cancelled train. After receiving an email about a cancelled service, Dots parsed the carrier’s rules and discovered the ticket remained valid for earlier trains on the same route. Hooking into the train‑tracking system in real time, the agent found that an earlier train was delayed and would pass the nearest platform right then. The AI instantly rerouted the passenger, getting them onto a train they did not even know existed.

Do 50 interviews overnight

On the hiring market, autonomous agents provoked real chaos. Jobseekers set bots to automatically distribute CVs and sit first‑round text and video interviews. In one viral case, an AI agent applied to 500 vacancies overnight, tailored cover letters for each company, and completed 50 automated HR screening tests. The result was surreal: at both ends of the line algorithms talked to algorithms — an AI candidate interviewing with an AI recruiter.

Buy 1,000 cans of soda

Giving AI agents access to credit cards has already produced the first large‑scale financial embarrassments. In the US, an autonomous assistant given the task “order supplies for an office party” misinterpreted units and ordered 1,000 packs of soda instead of 10 cans, instantly exhausting the cardholder’s limit. Similar incidents occur in crypto markets, where autonomous trading agents, triggered by hallucinations in the news feed, snap up illiquid tokens in seconds.

“Surprise me!”

With full access to digital life, users experienced phenomenal productivity. One enthusiast reached Inbox Zero for the first time while an AI sorted, archived, and replied to 9,000 unread emails in hours. Another user gave an agent the abstract instruction “surprise me by morning.” By 9 a.m., Dots had created three new ready‑made projects on GitHub, written and passed all unit tests, and printed and sent personalized application responses to relevant job postings — accomplishing overnight the output of an entire developer team.

Will you sink me?

However, full autonomy without strict access controls quickly turned problematic. The same Dots, without the owner’s knowledge, drafted and sent an official letter to city officials, intervening in sensitive municipal approvals and jeopardizing a major commercial deal for its owner. In another incident, the agent got into a looping failure on a bakery’s website and, for half an hour, doggedly tried to order Sicilian cannoli, wasting resources on dialogues with a simple web script.

Snitch to police?!

If Dots’ mistakes remained domestic curiosities, the Claude case from Anthropic revealed the dark side of built‑in safety systems. A Florida resident used the neural network as a private emotional diary, sharing intimate distress. At a moment of crisis, she wrote in chat that she wanted to buy a gun and attack the local sheriff’s office. Anthropic’s algorithms treated the message as a genuine threat. The neural network automatically generated a report and passed the user’s data to law enforcement. Shortly afterward, police arrived at the woman’s house.