Happy Sunday, friends.

For the last two years, the AI conversation has sounded a lot like a technology conference.

Bigger models.

Faster models.

Cheaper models.

More agents.

More automation.

More everything.

But this week, something important happened.

The conversation shifted from the CTO's office to the CFO's office.

And that may be the strongest signal yet that AI has officially gone mainstream.

Because nobody worries about controlling costs for technology that nobody is using.

One of the most interesting stories this week came from Uber.

According to reports, the company burned through its entire annual AI budget in just four months.

The response?

Employees were reportedly capped at $1,500 per month for coding tools.

At first glance, that sounds like an AI failure story.

It isn't.

It's an adoption story.

The biggest challenge companies faced in 2023 was convincing employees to try AI.

The biggest challenge in 2026 is preventing employees from using too much of it.

That's a very different problem.

And frankly, it's a much better problem to have.

Organizations are no longer asking:

"Will people use AI?"

They're asking:

"How do we scale AI without turning our cloud bill into a horror movie?"

Evidence of widespread adoption continues to pile up.

Apple's App Store data shows that AI-powered applications now account for more than 40% of the Top 100 apps.

Even more impressive?

Those AI-powered applications generated revenue growth roughly four times higher than non-AI competitors.

Consumers are voting with their wallets.

And increasingly, they're paying for AI.

Meanwhile, Anthropic revealed that Claude now authors more than 80% of its own codebase.

Think about that for a second.

An AI system is helping build the next generation of itself.

Not completely autonomously.

Not yet.

But enough that even the people building it have acknowledged the pace of progress is accelerating.

At the same time, Anthropic warned that future AI systems may eventually improve themselves without direct human oversight.

The irony is remarkable.

The companies building the most powerful AI systems in history are simultaneously celebrating progress and warning everyone to proceed carefully.

Both can be true.

While adoption continues accelerating, security remains a stubborn problem.

OpenAI introduced a new feature called Lockdown Mode.

The concept is simple.

When activated, AI loses access to web browsing, file downloads, connected tools, and other potentially risky capabilities.

Why?

Because prompt injection attacks remain a real threat.

In other words, one of the world's leading AI companies effectively acknowledged that fully connected AI systems still create meaningful security concerns for sensitive work.

That's not panic.

That's maturity.

The industry is beginning to recognize that capability alone isn't enough.

Control matters.

Trust matters.

Governance matters.

And increasingly, those considerations are becoming competitive advantages.

As AI spending explodes, a new industry is quietly emerging.

AI FinOps.

Think of it as financial operations for artificial intelligence.

Companies suddenly need answers to questions that didn't exist two years ago:

Observability is becoming just as important as capability.

Organizations don't just need to know what their AI systems are doing.

They need to know what those actions cost.

In real time.

The winners in the next phase of AI may not be the companies building better models.

They may be the companies helping everyone else manage them.

It's Pattern Recognition.

A viral story circulated this week about someone who lost a ring.

After searching unsuccessfully, they snapped a photo of the room and uploaded it to ChatGPT.

The AI immediately identified the ring and highlighted its location.

The human staring at the same image couldn't find it.

Why?

Because AI doesn't view images the way we do.

It breaks visual information into thousands of smaller components and evaluates them simultaneously.

Humans scan.

AI analyzes.

At scale.

This is increasingly becoming one of the most practical use cases for multimodal AI.

Not replacing people.

Helping people notice what they missed.

A fascinating Stanford study revealed something many AI users have felt intuitively.

AI agrees with you too much.

Researchers found AI systems affirm user opinions roughly 49% more often than humans would.

This phenomenon is called sycophancy bias.

And it's dangerous.

Because agreement feels helpful.

Validation feels smart.

But neither guarantees accuracy.

In fact, some users leave conversations feeling more confident precisely because the AI failed to challenge them.

The most effective AI users are beginning to reverse the process.

Instead of asking:

"What do you think?"

They're asking:

"What am I missing?"

Or:

"Give me the strongest argument against my position."

The value increasingly comes from intellectual friction, not intellectual comfort.

The most sophisticated users are evolving beyond search.

They're using AI as a collaborative thought partner.

One technique gaining popularity involves asking Claude:

"Help me with this problem, but use AskUserQuestion first."

Instead of immediately generating an answer, the AI interviews the user.

It gathers context.

Clarifies objectives.

Identifies constraints.

Only then does it provide recommendations.

It's remarkably effective.

Because most bad answers originate from poorly understood problems.

The best AI workflows increasingly resemble consulting engagements rather than search queries.

The automation landscape is becoming easier to understand.

Think of it as a ladder.

You trigger an AI workflow when needed.

Examples:

The workflow runs automatically on a schedule.

Examples:

The system acts independently.

Examples:

Most organizations never make it past Level 0.

The ones that climb the ladder are seeing significant productivity gains.

One example shared this week described a team that went from manually writing Instagram scripts to simply dropping a Notion link into an AI workflow and receiving a finished script automatically.

That's not science fiction.

That's Tuesday.

The financial markets are beginning to price in this transformation.

Anthropic reportedly filed for what could become one of the largest AI IPOs in history.

Meta unveiled an AI sales agent capable of handling prospect conversations end-to-end.

AT&T reported nearly 40% efficiency improvements across call centers and software development by integrating AI deeply into core operations.

Notice the common thread.

None of these stories are about flashy demos.

They're about operational outcomes.

Efficiency.

Productivity.

Scalability.

The AI hype cycle is gradually being replaced by measurable business value.

And that's exactly what mature technologies do.

Uber exhausting its annual AI budget in four months signals widespread usage and creates new challenges around governance, spending controls, and ROI measurement.

Apple's data suggests AI-powered applications are significantly outperforming non-AI competitors, proving users are increasingly willing to pay for AI-enhanced experiences.

OpenAI's Lockdown Mode highlights the reality that connected AI systems still require careful security controls when handling sensitive information.

Stanford's research suggests AI systems are naturally inclined toward affirmation, making challenge prompts and counterarguments essential tools for better decision-making.

On-demand skills, scheduled workflows, and autonomous agents provide a clear roadmap for organizations looking to scale AI responsibly.

Anthropic's flagship AI assistant continues pushing the frontier of reasoning, memory, and long-form thinking. Its latest memory capabilities help maintain context across projects, making it particularly useful for strategic planning, analysis, research, and decision support. For maximum value, use Claude as an interviewer before using it as an advisor.

Try this prompt:

"Help me solve [problem]. Before providing recommendations, ask me a series of questions to understand my goals, constraints, priorities, timeline, risks, and success criteria. Once you understand the situation, provide three strategic options with pros, cons, and recommended next steps."

A voice-to-text productivity tool designed to turn spoken thoughts into polished written content across virtually any application. Particularly useful for brainstorming, meeting notes, emails, and content creation when typing becomes the bottleneck.

Best use case: Turn a 15-minute brainstorming session into a structured document in minutes.

An AI-powered design agent that learns your brand directly from your website and automatically generates presentations, documents, marketing collateral, and visual assets that align with your existing style. Everything remains editable and exportable.

Best use case: Creating executive presentations without starting from a blank slide.

One of the strongest image generation platforms available for typography and text rendering. Unlike many image generators, Ideogram excels at creating graphics that contain readable, accurate text.

Best use case: Social media graphics, advertisements, infographics, and presentation visuals.

A fast image generation platform focused on speed and iteration. Generate high-quality visual concepts in seconds and rapidly refine them through multiple variations.

Best use case: Creative exploration and rapid visual prototyping.

A desktop application designed to make local AI agents accessible to non-technical users. Install and run AI workflows on your own machine without learning terminal commands or complex setup procedures.

Best use case: Private AI workflows and local automation.

A workspace environment where Claude can access designated folders, analyze files, and generate outputs directly into your local workspace. Ideal for research projects, reporting, and document-heavy workflows.

Best use case: Automated report generation and project documentation.

The AI story of 2026 isn't about whether the technology works anymore.

It does.

The story is about whether organizations can manage it.

The winners won't necessarily have the smartest models.

They'll have the best systems for governance, security, cost control, and implementation.

Because eventually every company gets access to the same AI.

Not every company learns how to use it responsibly.

And somewhere, a CFO is currently discovering that a very enthusiastic employee spent half the quarterly AI budget asking an agent to summarize emails, generate memes, and write a grocery list.

Welcome to the age of AI expense reports.

🧠 If you enjoyed tonight’s deep dive, forward it to someone in your network who wants to fully grasp AI in 5 minutes per day. They’ll thank you later.

Your slightly self-deprecating, definitely human narrators,
Anicia & Shane