Ford isn’t alone. According to CNBC, Commonwealth Bank of Australia laid off more than 40 customer service staff last year and replaced them with an AI voice bot. Complaints and call volume both went up instead of down, and the bank reversed the decision and rehired the staff.
IBM replaced parts of its human resources function with AI and found the system could handle about 94 percent of routine requests just fine. It was the remaining 6 percent, the ethically tricky, judgment-heavy edge cases, that it couldn’t manage. IBM’s response was to triple its entry-level hiring across the U.S. in 2026.
For a more entertaining, but genuinely instructive, look at the same weaknesses, watch Kitboga’s video, “Don’t Hang Up On AI Scammers. Do THIS Instead.”
In it, Kitboga (a well-known scambaiter) gets a call from an AI-powered scam bot posing as a student counselor, the kind of scam that targets international students with extortion and identity theft. He confirms it’s AI by asking it to do things no real human would ever agree to, like pronouncing punctuation marks out loud mid-sentence. The bot complies instantly and cheerfully, which is the giveaway.
From there, he escalates. He gets it to stack commas and exaggerate periods, then tells it that every time it uses the letter “A” in a word, it also has to say “Albuquerque, New Mexico.” Since the letter A shows up constantly, the instruction compounds on itself. Within minutes, the bot is stuck in a repeating loop, garbling words, switching voices mid-sentence, and effectively breaking down completely, all while the scam operation keeps paying per minute for a call that is going nowhere.
It’s funny to watch, but it points to one of the most important weaknesses in current AI systems: it cannot reliably operate outside its training data, and, more importantly, it cannot keep track of long-term or even mid-term goals once a task or conversation gets complicated. In AI terms, this is sometimes called context degradation or goal drift. The system’s original purpose gets buried under everything that has piled up since, and it starts reacting to whatever is most recent instead of staying anchored to what it was actually supposed to be doing. This is closely related to what’s known as poor long-horizon task management, the inability to hold onto a goal and stay coherent across a long or complex interaction.
That same flaw isn’t just comedic on a scam call. It’s the exact reason unsupervised, AI-generated websites tend to fall apart the bigger they get. Ask an AI to write one page, and it usually does fine. Ask it to manage dozens of pages, or build out a whole site over time, and it starts losing track of what it already said, built, or promised on page three by the time it gets to page thirty. This is where a professional’s oversight stops being optional.
Here’s what that typically looks like in practice:
AI is a genuinely useful tool. It’s fast, it’s tireless, and it’s great at handling scale. But scale is exactly where its weaknesses compound, the same way “Albuquerque, New Mexico” compounded into a total breakdown on a scam call. A website is a long-horizon project by nature. It needs someone tracking the whole picture, not just the next sentence.
Yes. It’s called Vending-Bench / Vending-Bench 2.
It tests whether AI agents can run a simulated vending-machine business over a long period. The AI has to manage inventory, set prices, order supplies, negotiate with suppliers, handle refunds, deal with disruptions, and stay profitable.
Andon Labs created and runs the benchmark. Epoch AI also tracks and displays the results.
Not every AI model participates. The test is run on selected models that researchers choose to evaluate. Current examples include models from Anthropic, OpenAI, Google, xAI, Zhipu/GLM, Moonshot/Kimi, and other frontier or open model groups.
So it is not a required industry exam. It is more like a serious third-party benchmark that compares selected AI agents.
It’s important, but still niche. It measures something normal AI tests often miss: whether AI can stay useful, coherent, and profitable over a long, messy task instead of just answering one prompt correctly.
It is significant as an AI agent benchmark. The vending machine itself is not the point. The point is that running even a tiny business exposes whether an AI can plan, adapt, negotiate, remember decisions, and avoid bad loops over time.
Not official in the sense of an SAT, IQ test, or government standard. It is a research/industry benchmark created by Andon Labs and tracked by Epoch AI, which makes it more credible than a random demo, but it is still not a universal standard.
