Use AI to solve problems, not just to use it
Em algum ponto do mundo corporativo, uma conta de quinhentos milhões de dólares em um único mês tornou-se o espelho que o mercado de inteligência artificial generativa precisava encarar. A promessa de uma ferramenta ilimitada encontrou, inevitavelmente, os limites da economia real — e o entusiasmo sem disciplina revelou-se tão custoso quanto a ignorância. O que está em curso agora não é uma rejeição da tecnologia, mas a passagem difícil da adolescência para a maturidade: da adoção por impulso à adoção por propósito.
- Uma empresa gastou R$ 500 milhões em um único mês com o modelo Claude, da Anthropic, simplesmente porque ninguém havia definido um limite de uso — e os funcionários usavam a IA até para consultar a previsão do tempo.
- A história viralizou nas redes sociais como símbolo do absurdo: tokens de texto consumindo fortunas que comprariam ilhas, jatos e iates, enquanto nenhum valor real era gerado.
- Microsoft, Uber e Amazon estão cortando orçamentos, encerrando licenças irrestrictas e removendo incentivos internos ao consumo excessivo de IA, após constatar que mais uso não se traduz em melhores resultados.
- A Amazon eliminou um placar interno que premiava quem mais consumia tokens de IA, e um vice-presidente sênior reorientou a cultura: use a inteligência artificial para resolver problemas reais, não para demonstrar adoção.
- O mercado começa a transitar do modelo 'adote agora, avalie depois' para estratégias centradas em retorno mensurável — uma mudança que promete redefinir como e onde a IA será implantada nas empresas.
Em algum lugar no mundo corporativo, chegou uma fatura que paralisou todos: quinhentos milhões de dólares gastos em um único mês com o modelo de inteligência artificial Claude, da Anthropic. O problema não era sofisticação tecnológica — era a ausência total de qualquer controle. Sem limites de gasto definidos, funcionários usavam a ferramenta livremente para tarefas triviais: consultar o clima, fazer perguntas básicas, coisas que qualquer pessoa resolveria em segundos. A conta só crescia porque ninguém estava olhando o medidor.
O caso, relatado por um consultor de IA ao Axios, tornou-se uma parábola para um mercado ainda em formação. Durante o último ano, empresas adotaram a IA generativa com o entusiasmo de quem encontrou um brinquedo novo e decidiu não perguntar o preço. O fracasso espetacular de uma empresa expôs o que acontece quando a disciplina desaparece por completo — e as redes sociais não perdoaram a ironia de fortunas evaporadas em tokens de texto.
Mas o episódio revelou algo maior: uma indústria começando a acordar para o abismo entre entusiasmo e economia. A Microsoft substituiu licenças do Claude Code pelo seu próprio GitHub Copilot CLI em busca de eficiência. A Uber esgotou todo o orçamento de IA para 2026 já em abril, e um executivo da empresa disse em voz alta o que muitos sussurravam: a relação entre mais IA e melhores resultados simplesmente não estava se confirmando.
A Amazon, por sua vez, havia criado um placar interno que incentivava funcionários a consumir mais tokens — como se o uso em si fosse uma virtude. Esse placar foi extinto. Um vice-presidente sênior reorientou a mensagem para a equipe: não use IA por usar. Use para resolver problemas de clientes, problemas de negócio, para inovar. A mudança era sutil, mas total — do consumo como fim ao consumo como meio.
O que se desenha agora é uma virada de fase. A corrida do ouro da adoção irrestrita está cedendo lugar a uma pergunta mais difícil e mais cara de responder: para que, exatamente, devemos usar isso?
A company somewhere in the corporate world received a bill that stopped everyone cold: five hundred million dollars, spent in a single month on Claude, Anthropic's artificial intelligence model. The culprit was not a sophisticated deployment or a carefully architected system. It was the absence of any guardrail at all. No one had bothered to set a spending limit. Employees, given free rein to use the tool, did exactly that—running queries for tasks that required no artificial intelligence whatsoever. Checking the weather. Asking basic questions. The kind of work a person could do in seconds. But it was cheaper, or felt cheaper, because no one was watching the meter.
The story, first reported by an AI consultant speaking to Axios, became a kind of parable for a market in its adolescence. For the past year or so, companies had been adopting generative AI with the enthusiasm of people who had found a new toy and decided not to ask the price. The spending was real, the adoption was real, but the discipline was missing. This one company's catastrophic bill exposed what happens when that discipline vanishes entirely.
Social media seized on the absurdity. Five hundred million dollars. That was enough to buy private jets, yachts, an island. All of it evaporated into tokens—the tiny units of text that AI models consume and charge for. The darker jokes circulated too: what happens in the meeting where you explain this to the board? Do you fire the person who approved unlimited access? Do you fire everyone who used it?
But the real story was not about one company's spectacular failure. It was about what that failure revealed: a market beginning to wake up to the gap between enthusiasm and economics. Microsoft started replacing Claude Code licenses with its own GitHub Copilot CLI, hunting for better cost efficiency. Uber burned through its entire 2026 Claude budget by April. Andrew Macdonald, an executive there, said plainly what others were starting to whisper: the connection between more AI and better results for customers simply was not there.
Amazon had created an internal leaderboard—a scoreboard, really—where employees could see who was consuming the most AI tokens. There was pressure, implicit but real, to use more. The company was incentivizing consumption for its own sake. That leaderboard is gone now. In a recent message to staff, Dave Treadwell, a senior vice president, reframed the entire conversation. Do not use AI just to use it, he said. Use it to solve customer problems. Use it to solve business problems. Use it to innovate. The difference was subtle but total: from consumption as a virtue to consumption as a means to an end.
What was happening across these companies was a reckoning. The initial phase of generative AI adoption had been characterized by a kind of gold-rush mentality—get it in everywhere, figure out the value later. The bills were arriving now, and the value was harder to find than expected. Companies that had embraced the technology aggressively were now stepping back, not abandoning it but reorienting it. The question was no longer how much can we use, but what should we use it for. The answer to that question, it turned out, was much more expensive to figure out than anyone had anticipated.
Notable Quotes
The connection between increased AI usage and customer returns simply is not there— Andrew Macdonald, Uber executive
Use AI to solve customer problems, business problems, to innovate—not just to use it— Dave Treadwell, Amazon senior vice president
The Hearth Conversation Another angle on the story
How does a company even let a bill get to five hundred million dollars without noticing?
Because no one was watching. There was no limit set on the licenses. Once you remove the friction—the cost visibility—people use things differently. They use them for everything.
But surely someone saw the charges coming in?
That's the thing about cloud spending and API costs. They're often invisible until they're catastrophic. The bill arrives, and by then you've already spent the money.
So this is really about governance, not about the technology itself?
Exactly. The technology works fine. The problem is that companies treated it like it was free. They incentivized usage without asking whether the usage was creating value.
Why would Amazon create a leaderboard for token consumption?
Because in the early days, everyone believed more AI was better. It felt like innovation. It felt like progress. No one wanted to be the person using it the least.
And now?
Now they're asking harder questions. Is this actually making us money? Is this actually helping our customers? Those are much harder questions to answer.