No‑BS AI Briefing
No‑BS AI Briefing is for builders who don’t have time for hype. Each episode focuses on a handful of high‑signal stories in AI and AGI, unpacked in simple language with a builder’s perspective. You’ll hear what changed, why it matters, and how you can experiment with the tools, ideas, or strategies yourself—whether you’re leading a team, shipping a startup, or exploring AI side projects.
No‑BS AI Briefing
AI's Hidden Costs: Klarna & Uber's $55M Reality Check
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Today on Nobs AI Briefing, Klana tried to replace 700 customer service reps with an AI and it ended up costing them over $15 million. Uber's CTO burned through their entire 2026 AI budget before summer even started, and Google is rolling out a whole new platform for building the very agents that are causing these problems. We'll talk about what actually matters for builders in all of this. NoBS AI briefing brought to you by Proactive AI. Welcome back. I'm your host, Vikash, and this is where builders get straightforward AI news without the fluff. Alright, let's get into the headlines because this week we got a serious reality check. First up, Klarna's AI customer service failure is a story every founder needs to hear. They announced a plan to replace 700 human customer service agents with AI, aiming for a massive $40 million in annual savings on paper. It probably looked like a genius move. In reality, it was a disaster. According to Forbes, customer satisfaction completely collapsed. The AI just wasn't good enough. It couldn't handle the complexity, the nuance, the empathy that real customers needed. By mid-2025, they had to hit the brakes and start rehiring human staff. And the cost to unwind this mess, rehiring, retraining, reputation damage, has already exceeded $15 million. That turns a $40 million savings plan into a $15 million loss. And that's not even counting the customer churn. For builders, this is proof that AI without strong governance and human-in-the-loop feedback isn't a strategy, it's a liability. It shows the hidden costs of failure go way beyond the tech. And speaking of costs, our next story shows the other side of that coin. Uber's AI budget burnout is a cautionary tale about the raw expense of this technology at scale. The same Forbes report revealed that Uber's CTO managed to burn through the company's entire 2026 budget for AI token usage before the year was even half over. I mean, think about that. A company with the resources of Uber and they still misjudged the cost of inference so badly they blew their budget in less than six months. This just hammers home that token costs are still the primary hard constraint on AI adoption. It's not about whether the models are smart enough, it's about whether the economics actually work when you go from 100 test users to millions of real ones. For builders, the lesson is stuck. Your product's financial viability is directly tied to token pricing. Your budgets have to account for exponential consumption, not linear. And optimization strategies, things like caching common results, batching requests, and using smaller, cheaper models for simpler tasks, they aren't nice to have anymore. They are absolute table stakes. So while big companies are struggling with the cost and quality of AI, what's happening on the platform side? Well, Google is building the tools for the next wave. At their I.O. conference, they announced the Google anti-gravity platform, which includes a new SDK and command line interface for building and deploying AI agents. This is a big deal. They're trying to create a full developer platform, not just an API endpoint. One of the most interesting parts is that their managed agents in the Gemini API will now support provisioning remote Linux environments. This means an agent can have its own little sandboxed computer to do complex, stateful work like coding, testing, and running scripts without you having to build and manage all that infrastructure yourself. That's a huge unlock. And to top it off, they're proposing an open standard called WebMCP for letting agents interact with web tools and they're even building agent-specific features into Chrome DevTools. For us, this signals that the game is moving from simple API calls to building sophisticated autonomous systems on managed infra. The WebMCP standard, if it gets adopted, could be huge for preventing lock-in and making agent tools more portable. And finally, to power this new agent stack, Google also announced Gemini 3.5 Pro. It's the more capable successor to their speedy Gemini 3.5 flash model. And it's scheduled for a June 2026 release. This is pretty standard stuff. The next model is always just around the corner, promising to be better, faster, and smarter. But what matters for builders is the timing. A June release means it's time to start planning for evaluation. It's an opportunity to benchmark it against whatever you're using now, whether that's OpenAI's models, Anthropics, or even Google's older models. You have to ask, is the performance jump worth the switching cost? But when you look at it combined with the anti-gravity STK and the managed environments, you can see the cohesive stack Google is trying to build. They want to give you the powerful new brain and the fully equipped body for your AI agents all in one place. It's a compelling package, and one will have to watch very closely over the second half of the year. Okay, let's do a deep dive. The story I can't stop thinking about this week is the combined cautionary tale from Klarna and Uber. It's all about the real unhyped cost of deploying AI at scale. So what happened in a nutshell? Klarna tried to swap humans for AI in customer service and ended up with a $15 million bill and angry customers. Uber let its AI agents run so hot they burned the entire year's token budget by May. These aren't theoretical risks discussed in a white paper. These are real quantified failures from two major tech companies. Why does this matter so much right now? Because for the last couple of years, the entire conversation in AI has been about capability. Can the model write code? Can it analyze a document? Can it pass the bar exam? We've been so focused on can we build it that we forgot to keep asking, should we build it and what's the real all-in cost? The Klarna and Uber stories slam the brakes on that hype train. They prove that governance, financial discipline, and human oversight aren't boring enterprise topics to be dealt with later. They are business critical from day one. They are the difference between a successful AI feature and a multi-million dollar write-off. So who should be paying the most attention to this? Well, if you're a founder or a product leader, these stories are your new nightmare fuel. Klarna shows that customer satisfaction and AI capability are not the same thing. You can't just replace a human workflow and assume the outcome will be positive. You have to measure the impact on your users constantly. Uber shows that your AI features cog your cost of goods sold, can spiral out of control and kill your margins before you even find product market fit. Your financial model has to be built around token consumption. If you're an engineering leader or an infra engineer, your job just got a lot more focused on efficiency. The Uber story means that token optimization is no longer nice to have, it's a core competency. Your teams need to be experts in caching strategies, in model selection, knowing when to use a cheap, fast model versus an expensive, powerful one, and in designing systems that make as few API calls as possible. Building cost tracking and alerting directly into your AI stack is now mandatory. And if you're building in the infrastructure layer yourself, maybe you have a dev tool or a platform, these stories are a massive opportunity. The pain Uber is feeling is a market need. Tools for token optimization, cost management, and AI governance are no longer niche. They're becoming essential utilities. Here's how I'd think about it as a builder. Treat every AI call like you're spending real cash because you are. For years we got used to compute being relatively cheap and predictable. Server costs are understandable, but inference costs are different. They're variable, they're spiky, and they scale with user engagement in a way that can be terrifying. You need to build a dashboard that tracks your token consumption per user, per feature, per day. You need to know your cost per magic moment. If a user gets a great AI-powered summary, what did that summary actually cost you in tokens? If that cost is 10 cents and you only make 5 cents from that user, you don't have a business. It's that simple. So here's my no-biest take. We're at an inflection point. The era of just shipping cool AI demos is ending. The hype is giving way to the hard reality of unit economics and operational excellence. The companies that win in the next phase of AI won't just have the smartest models, they'll have the tightest financial discipline and the strongest governance. The plumbing is starting to matter more than the magic. If you're finding this useful, hit follow in your podcast app right now. It takes two seconds and it's the best way to make sure you don't miss the next briefing. Alright, if you want one practical takeaway from today's episode, here it is. Audit your AI budget and your real token consumption. Don't wait for your own Uber moment. You can do this in less than 30 minutes. Here's how. First, spend 15 minutes just getting the data. Log into your OpenAI, Anthropic, or Google Cloud Billing dashboard. Don't just look at the total bill. Go to the usage section. Most of these platforms let you break down usage by model, and if you've been smart, you're using different API keys for different features, sort by cost. Your goal is to find the one or two features or workflows that are responsible for 80% of your token consumption. I guarantee you'll find a surprise in there. Second, spend the next 15 minutes comparing that reality to your forecast. If you don't have a forecast, well, that's your first problem. Make one now. Ask yourself. If this feature gets 10 times more usage, what does that do to my monthly bill? If the answer is terrifying, you need to act now. Finally, based on what you found, pick one high burn workflow to target for optimization this week. Just one. You have three simple levers to pull. One, can you use a cheaper, faster model? Maybe that complex summarization task can be handled by Gemini 3.5 flash instead of GPT-4 Turbo. The quality difference might be negligible for your use case, but the cost difference is massive. Two, can you implement a simple cache? If users are asking for the same thing over and over, stop hitting the API and serve a cached result. 3. Can you re-engineer the prompt or the workflow to use fewer calls? Sometimes a better prompt can get the job done in one shot instead of three. Doing this little 30-minute audit isn't just a cost-saving exercise. It's about building a sustainable product. Your token spend is a core part of your cost of goods sold. If you don't have a tight grip on it, you're flying blind. And as Klerna and Uber just showed us, flying blind in this environment can get very, very expensive. That's it for today's NoBS AI briefing. If this helped, follow the show in your podcast app and share it with one builder you know. And if you've got questions or topics you want covered, connect with me on LinkedIn and send them over. See you in the next briefing.
