Hi {{first_name}}!
Big week. Stanford just released its 2026 AI Index and one number stands out: Generative AI has hit 53% adoption in just three years. To put that in perspective, it took the PC and the internet twice as long to reach that same level.
If you feel like you are still "just testing" while the world is moving on, you aren't alone. But this week proved we are moving out of the experimental phase. We are seeing a shift toward professional, reliable tools—from Microsoft testing "always-on" agents to Google turning AI prompts into one-click workflows.
At the same time, we are seeing the first "growing pains" of this scale. Major outages and shifts in pricing models suggest the "wild west" era of free, unlimited AI is starting to close. Here is what you need to know to keep your edge.
In This Issue:
The News: Why "unlimited" AI plans are ending, Microsoft’s new 24/7 agents, and Google's local privacy win.
The Deep Dive: Why the better AI gets, the more dangerous its hallucinations become (and how to spot them).
The Pro Tip: The one "Global Instruction" setting in Claude Cowork that saves you hours of setup.
Strategy: Four ways to run AI on your own hardware to keep your business data 100% private.
OK, let's get into it!
New and Noteworthy
Google "Skills" Makes Chrome an Automation Tool On Tuesday, Google launched a "Skills" feature in the Chrome Gemini sidebar. It lets you save your most-used prompts and run them as one-click actions across multiple open tabs at once. If your team spends their day comparing products or summarizing dozens of research pages, this turns a repetitive manual process into a single click.
The "End of the Buffet" for AI Agents Anthropic recently blocked Claude Pro subscribers from using their $20 flat-rate plans with third-party agent tools like OpenClaw. Why? Because a single autonomous agent can burn through hundreds of dollars in compute costs in one day. Anthropic is moving these "power users" to a pay-as-you-go model. If you are building automation agents, the takeaway is clear: the era of "unlimited" agents for a flat monthly fee is ending. Efficiency and cost-tracking now matter more than ever.
Microsoft Tests "Always-On" Copilot Agents Microsoft is building persistent AI agents that run in the background 24/7. Instead of you asking for help, these agents watch your inbox and calendar to surface conflicts, track document changes, and alert you when a file becomes time-sensitive. This is the shift from "AI you talk to" to "AI that works for you." Expect a wider rollout at the Build conference in June.
Claude for Word is now in beta You can now use Claude directly inside Microsoft Word via a native sidebar. It can draft content, suggest redlines for contracts, and even share context with Excel or PowerPoint. If your team lives in the Microsoft ecosystem, this is a much more seamless way to work than copy-pasting back and forth from a browser tab.
Stanford AI Index: The Speed of Change Beyond the 53% adoption rate, the report found that 88% of organizations have now adopted AI in some form. However, it also flagged that model transparency is declining. For business owners, this confirms that your competitors are almost certainly using these tools, but the "black box" nature of how they work means you need your own internal verification steps.
Tech Layoffs and the "AI Washing" Trend Tech layoffs hit 80,000 in Q1. While many companies cited "AI and automation" as the reason, industry experts suggest some of this is "AI washing" to satisfy investors. The real takeaway: companies are hiring fewer entry-level roles and paying a premium for people who can bridge the gap between business strategy and AI implementation.
Google Releases Gemma 4 for Local Privacy Google released Gemma 4, a powerful "open" model that you can run on your own hardware. Unlike ChatGPT or the standard Gemini, this doesn't require a cloud connection or an internet connection of any kind for that matter. For businesses in legal, finance, or healthcare, this is a massive win for data privacy. You can run high-end AI without your data ever leaving your office.
Meta's Custom Chip Push Meta is investing heavily in custom AI chips through 2029 to reduce its reliance on Nvidia. For you, this competition is good news. More players in the hardware space eventually lead to lower subscription costs and faster processing speeds for the tools you use every day.
"CEOs are betting AI will augment work rather than displace all workers": At the Semafor World Economy conference, Anthropic co-founder Jack Clark publicly disagreed with his own CEO Dario Amodei's prediction that AI could push unemployment to 20% within five years. Clark called that outcome a policy "choice," not an inevitability. The broader consensus among executives at the event leaned toward augmentation over replacement. That tracks with what we see in practice with SMB clients: teams using AI to do more with the people they have, not gutting headcount. At the enterprise level and with publicly traded companies I can image that might be a different story.
Google tests an "Agent" tab, hinting at a Cowork competitor: A new Agent tab has appeared in Gemini Enterprise alongside the standard chat interface, featuring a task workspace with goal-setting, connected apps, file uploads, and a "require human review" toggle. The layout looks a lot like Claude Cowork's approach: give the model a goal, connect it to your tools, and let it execute multi-step work. Google is also known to be building a desktop app for AI Studio, raising the question of whether these efforts merge into a single product. Expect more details at Google I/O. The desktop agent race is heating up fast.
Why Even the Best AI Models Still Hallucinate
The better AI gets, the less people check its work. This is exactly when it becomes most dangerous for your business.
Hallucinations (where a model tells you something factually wrong with total confidence) haven't gone away. They have just become rarer. When a tool like Claude or ChatGPT is right 95% of the time, your brain naturally stops questioning the other 5%. You might miss a fake citation, a made-up statistic, or a slightly wrong date that looks perfectly reasonable.
In a professional setting, the most successful teams I work with follow three simple habits:
The "Tell Me Why" Rule: Always ask the AI to provide sources for its claims. Then, ask it to confirm those sources actually exist and say what it claims they say.
The "I Don't Know" Permission: Explicitly tell the AI in your prompt: "If you are unsure or don't have the data, tell me you don't know rather than guessing." This single sentence significantly reduces hallucinations.
The Human Filter: Never let AI-generated content go directly to a client or a public platform without a human review. One simple human-in-the-loop "review step" in your workflow is the difference between a productivity win and a reputation risk.
The Cowork Setting That Ties Everything Together
If you are using Claude Cowork, there is one "hidden" feature that will save you hours of re-explaining: Global Instructions.
Found in the Settings under the Cowork tab, this is a single text field that Claude reads before every single session. It is the best place to set your "Rules of Engagement."
Instead of starting every chat by explaining who you are and how you like to work, put it here once:
Your preferred writing style (e.g., "concise, no jargon, no em dashes").
Your "Never" list (e.g., "Never use buzzwords like 'synergy' or 'robust'").
Your standard tools and context (e.g., "I use Slack for internal comms and ClickUp for project management and HubSpot as my CRM").
This turns Claude from a generic assistant into a partner that already knows your standards before you even type your first prompt.
Open Source AI: Why it Matters for Your Business
I wanted to break this down because it connects directly to something a lot of us are already doing: building AI agents and automations. Most of the AI tools we use daily, whether it is ChatGPT, Claude, or Gemini, are powered by closed models. The companies that built them control how they work, where they run, and what they cost. You are essentially renting access.
But there is a parallel world of open source models built by companies like Google, Meta, and Mistral. These are not watered-down versions. They are production-grade AI engines like the new Gemma 4 and Deepseek V3.2. They are increasingly what powers the agents and automation platforms many of us are building with.
So why would a business owner care? It comes down to three things:
Cost: Paid AI services charge per use, and those token costs add up fast once you are running agents at scale. Open source models are free to use. The more your agents run, the wider the savings gap gets.
Privacy: When your agents use cloud based LLMs, every interaction sends data to someone else’s servers. For businesses handling client information, financial records, or legal documents, that’s a real concern especially if you’re running on a free tier that is training on your data. In contrast, pen source models can run entirely on your own hardware. Your data never leaves your control.
Customization: Open source models can be fine-tuned to your business: your tone, your industry, and your specific workflows. This is where you build an agent that truly fits your operation instead of working around a one-size-fits-all product.
Here is how to get started, from simplest to most advanced:
Option 1: Try them in your browser
Sites like Groq and Arena.ai let you pick an open source model and start chatting immediately. There are no downloads and no setup. It is the fastest way to compare what is out there. Just be aware that this is not private and there are usage limits.
Option 2: Run them on your own Device
This is where it gets interesting for anyone who cares about data control. There is a free and widely used app called Ollama that lets you download a model and start chatting within minutes. A standard modern laptop handles smaller models like Gemma 4 (4B) perfectly. A popular setup I have been seeing is a dedicated Mac Mini running AI around the clock so it does not slow down your main machine.
Option 3: Build with hosted open source models
This is for people who want to create tools or agents without managing infrastructure. Services like Groq, Together AI, and Fireworks AI host open source models and let you connect through a simple access key. You get open source flexibility without the hosting headaches.
Option 4: Rent a private server
A VPS is essentially a computer in a data center that is always on and entirely yours. Providers like Hostinger offer this for $5 to $10 a month. You get a private environment to run AI models while keeping full control of your data. This especially matters in healthcare, legal, and finance where data privacy is not optional.
My suggestion: If you are just curious, start with Option 1. If you have a modern computer and privacy matters or you are already building agents, jump to Option 2 and get Ollama set up. It takes about 15 minutes and it will shift how you think about what is possible without a paid subscription.
If any of this sparked something you want to dig into further, or if you're trying to figure out where AI fits in your business and want to talk it through, hit reply and let me know. I read every response and I’m happy to chat through an idea.
See you next week,
Julien
PS: If this was helpful, forward it to someone on your team who'd get something out of it. Always appreciate the support www.ampra.ai/join-our-newsletter.