Hi {{first_name}}!
On one hand, businesses are finally getting clear on exactly where AI and automation save real time. On the other hand, the infrastructure underneath those workflows is becoming more powerful, more political, and less predictable.
The practical question is no longer, "Should we use AI?" It is, "Where do we deploy it first, and how do we build it so our operations keep running when the ground shifts?"
Here is a look at what we are covering in this week's edition:
The News that Matters: From ChatGPT's dipping market share to background agents landing inside Microsoft 365 and Android.
NotebookLM's Evolution: How the platform just shifted from simply summarizing your notes to executing real data analysis.
The Fable 5 Warning Shot: What a wild 72-hour launch-to-shutdown cycle teaches us about platform dependence.
OK, let's get into it!
New and Noteworthy
ChatGPT slips below 50% market share: Sensor Tower data shows ChatGPT’s global share of the AI assistant market has dipped to 46.4%, with Gemini climbing to 27.7% and Claude holding 10.3%. ChatGPT isn't shrinking, but its closest rivals are expanding rapidly. The operational lesson: The market is no longer a one-horse race. If you standardized your entire team on a single platform a year ago, it is a great moment to recheck the fit and see if another tool matches your workflows better.
China's GLM 5.2 matches the frontier at a fraction of the price: Released this month by Z.ai under an open MIT license, this coding-focused model matches or beats closed proprietary systems on long-horizon tasks at roughly one-sixth the cost. The operational lesson: Because the model weights are open, your business can download it and run it entirely on your own private infrastructure. No sudden vendor policy shift or price change can take it away from you.
Microsoft Launches Copilot Cowork Globally: Released generally on June 16, this upgrade operates entirely in the background across Outlook, Teams, Excel, and Word. Instead of just drafting a single email, you give it a high-level objective and it executes the multi-tool workflow end-to-end, returning a completed project. The operational lesson: Most small businesses will meet agentic AI inside the core software platforms they already pay for, not through brand-new independent applications.
Android 17 turns Your Phone Into an Active Agent: Google shipped its mobile update with Gemini Omni embedded at the system level. This lays the technical foundation for autonomous workflows arriving later this year that can reach directly into your apps to book appointments, edit logs, and place orders for you. The operational lesson: If your team relies on the Google ecosystem, keeping your data structured and clean is what determines whether these automated shortcuts actually function.
DeepSeek raised $7.4 billion at a $50 billion valuation: Chinese startup DeepSeek closed a massive funding round valuing the company at over $50 billion, with backing from tech heavyweights like Tencent and CATL. The operational lesson: The ultra-low-cost, open end of the AI market now has serious institutional money behind it. This structural backing will continue to exert heavy downward pricing pressure on Western cloud providers over the next twelve months.
Meta botches its internal AI reorganization: CTO Andrew Bosworth acknowledged an atrocious job communicating a restructuring that moved 6,500 engineers into an Applied AI unit, sparked by a massive employee pushback against background keystroke-logging software used to collect agent training data. The operational lesson: The software tools are the straightforward part. How you communicate operational changes and bring your team along is what actually decides whether a rollout succeeds or stalls.
Anthropic faces a lawsuit over Claude's usage limits: A federal lawsuit claims that Anthropic’s $200 per month Max plans deliver significantly less computing allowance than advertised, calling the capping system a black box. Heavy users reported burning through 15% of their weekly allowance in a single afternoon session. The operational lesson: Rate limits are a volatile reality of the AI economy. If you are paying for premium team tiers, track your real consumption metrics rather than relying blindly on vendor multipliers.
The big AI CEOs Push World Leaders for Global Rules: At the G7 summit in France, the heads of OpenAI, Anthropic, and DeepMind met with world leaders, including President Trump, to propose shared risk-review standards. The operational lesson: Software access is becoming a core piece of geopolitical strategy. Where your platform provider is headquartered is now a primary business risk question, not a technical footnote.
NotebookLM Gets File Creation and Code Execution
I wrote about the big NotebookLM update last week talking about an update that fundamentally changes how you interact with your own business data. NotebookLM has officially transitioned from an assistant that simply reads and summarizes text into an operational workspace that actively analyzes files and builds native, fully editable documents.
If you’re a NotebookLM user you already know that the real power of this platform lies in its ability to stay strictly grounded in your own data, rather than pulling randomly from the open web.
Here are the three practical shifts from this month's update that matter most for your daily workflows:
Direct File Creation from Chat: You can now ask NotebookLM to produce finished files directly from your sources within the chat section, located in the middle column. It can generate PowerPoint decks, Excel spreadsheets, Word docs, PDFs, Markdown text, and structured data like CSV and JSON. These are download-only assets built as native office files. This means you can download them and immediately edit or format them inside your standard software tools, unlike older static exports.
Secure, Inline Code Execution: Under the hood, NotebookLM now runs on the Gemini 3.5 architecture, meaning every notebook spins up its own isolated cloud environment to write and run code. Powered by Google's Antigravity platform, it handles more than 100 built-in software skills to execute real quantitative analysis. If you drop in a spreadsheet of your ad spend alongside a sales ledger, the AI will calculate the actual numbers and construct a visual chart rather than just describing the trend in a generic paragraph. You can even check its exact reasoning steps directly in the chat to audit its math.
Dynamic Source Discovery: Previously, you had to show up with a perfectly organized folder of documents to get accurate answers. Now, you can start with a loose idea or an abstract question, and the chat interface will actively help you build your source library from scratch, identify gaps in your current data, and surface relevant context. It only commits these new files to your notebook with your explicit approval.
The Bottom Line: Instead of forcing you to bounce between separate tools to research, calculate, and format data, NotebookLM can now manage the inputs, analyze them with code, and produce business-ready deliverables in one isolated place.
Note: This update is currently rolling out to Google AI Ultra subscribers and top-tier Workspace business accounts first, with wider availability expected later this year.
Fable 5 Lasted Exactly Three Days
On June 9, Anthropic launched Fable 5 (the widely available new model) and Mythos 5 (the model only available to vetted security partners under Project Glasswing). By June 12, both were completely gone. They were not throttled or slowed down due to a standard outage; they were entirely switched off for every customer worldwide.
The sudden shutdown followed a federal export control directive blocking access to the models for any foreign national, an order so broad that Anthropic chose to pull the models entirely rather than attempt selective enforcement. While the lab has pushed back, arguing the issue stemmed from a narrow jailbreak capability that already exists in competing models, the platforms remain down.
The Core Operational Risk
This situation represents the exact type of fragility that almost nobody prices in when choosing an AI platform. Most operators weigh a model purely on what it can execute and what it costs per token. They rarely evaluate how exposed that tool is to sudden regulatory mandates, vendor compliance disputes, or background political friction.
Fable 5 was the new flagship model and regular paying business customers were actively deploying it and very impressed with its capabilities. If you spent time wiring its API into a core workflow, your systems were completely broken by Friday morning through zero fault of your own, with no flaw in your actual setup.
AI has rapidly evolved from a conversational playground into critical infrastructure, and infrastructure demands real stability.
Building a Swappable Architecture
The takeaway here is not to avoid closed-source frontier models. We integrate them into operations every single day because they deliver incredible leverage. The lesson is that no single model should ever carry your entire business infrastructure.
When we design automated workflows for clients at Ampra, we treat the underlying model as a swappable utility part, not the permanent foundation. If your system architecture is model-agnostic, a provider going dark, your work simply routes to a fallback option that same day. Last week was the most definitive proof yet that this is the only resilient way to build.
Any questions about where to start, what to automate first, or how to make your workflows less dependent on one AI provider?
See you next week,
Julien
PS: Know someone who'd find this useful? Forward it their way, and they can subscribe at www.ampra.ai/join-our-newsletter.