How The Shadow AI Economy Is Hurting You

Every headline this week screams that 95% of corporate AI projects are failing. MIT drops a study, and suddenly everyone is ready to declare AI “overhyped.”

But the devil is in the details. And those details are fascinating.

Yes, companies are burning millions on failed pilots. But the report also reveals something bigger: 90% of employees have gone rogue. They’re paying out of pocket for tools like ChatGPT and Claude to get their jobs done.

This is the fastest technology adoption in corporate history—and it’s happening entirely outside official channels.

That should fascinate you. And terrify you.

The productivity gains happening outside formal work systems are only the surface issue. The real nightmare is data privacy: the larger the company, the bigger the risk.

The study bluntly explains why 95% of pilots fail: corporate AI systems “do not retain feedback, adapt to context, or improve over time.” The missing ingredient isn’t human—it’s that the AI itself can’t learn.

Fixing this requires a wholesale rethink of work and leadership. Building self-improving flows from atomic tasks. Empowering even juniors to become managers—of AI agents, not people. And resisting the corporate addiction to micromanagement and visibility theater.

Large companies will struggle to turn the battleship, but small, nimble teams can make the pivot today (even inside large orgs).

Next week, we’ll launch our AI Ready CMO community—where this exact challenge will be a major focus. Join here: join.aireadycmo.com

A quick word from this week’s sponsor:

Training AI? Let’s talk. If you're actively building or fine-tuning AI models, Shutterstock offers enterprise-grade training data across images, video, 3D, audio, and more—fully rights-cleared and enriched with 20+ years of human-reviewed metadata.

Our 600M+ asset library powers faster iteration, simplified procurement, and improved model accuracy.

We’re offering a $100 Amazon gift card to qualified AI decision-makers who join a 30-minute discovery call with our team. We’ll walk you through how our multimodal datasets, flexible licensing, and premium metadata can help de-risk development and accelerate time to deployment.

Book a call

_{For complete terms and conditions, see the offer page.}

Microsoft introduced a COPILOT() in-cell function for Excel that lets analysts call the model directly from the grid, return results, and chain them like any other formula. Early access is rolling out via the Microsoft 365 Insider program. At the same time, Paradigm, a new startup, has also launched a spreadsheet where each cell can host an agent to fetch, transform, and reason over data.  

This moves AI from chat panes to first-class spreadsheet logic—versionable, auditable, and compatible with existing formulas. Microsoft’s feature is gated by Insider availability; Paradigm is early and evolving. Expect staggered access, Excel admin controls, and pricing to shape adoption velocity.

Your growth model, media plan, and pipeline sheet are about to compute narratives, not just numbers. Think: natural-language lookups, automated cohort notes, and on-sheet copy drafts that inherit cell lineage.

A new image generation and editing model, reportedly from Google, is in a stealth launch phase and is already drawing comparisons to the debut of ChatGPT. Dubbed “Nano Banana,” the model demonstrates an extraordinary ability to understand and execute complex, natural language editing commands with a high degree of precision, a task where previous models have struggled.

Early tests circulating on social media show the model accurately adding reflections to surfaces, changing specific objects while maintaining shadows, and realistically inserting new elements into existing scenes. Unlike tools that require masking or complex prompting, Nano Banana appears to interpret intent with conversational instructions like “make the car look like it's driving through a puddle.”

Creative ops live or die on edit fidelity—especially for brand, legal, and talent likeness. If this closes the “face drift” gap, you can push more edits in-house and reduce manual QA rounds.

ElevenLabs, a leader in AI voice generation, has released a suite of updates that significantly expand its capabilities beyond audio.

The conversational AI tool used in customer support can now seamlessly switch between text and realistic, human-sounding voice interactions on demand. It is even intelligent enough to measure the background noise on the human’s end and propose continuing in writing if the environment is too loud.

ElevenLabs has also launched a video-to-music feature, enabling marketers to upload a video clip and have the AI generate a fitting, royalty-free soundtrack. It is uncanny: the model analyzes the video’s content, rhythm, and mood to generate music that matches it perfectly.

Further, an alpha version of the v3 audio model is now available via API, offering more granular emotional control across 70 languages.

Voice is becoming a modality toggle, not a channel. Your support bot can start in text, escalate to voice, and never lose context—slashing average handle time and preserving brand tone. And video-to-music is just brilliant.

A large-scale study involving 70,000 candidate interviews found that an AI recruiting system was more effective than human HR departments at spotting and hiring talent. The research, which spanned 48 distinct job roles, showed that candidates interviewed by the AI, named Anna, were 18% more likely to start a job. This is the first study of its size to benchmark human versus AI performance in a real-world hiring context.

The findings indicate the AI recruiter consistently selected candidates who not only performed better but also stayed longer, with 17% higher retention after one month. The system also had a significant impact on fairness, cutting reports of gender discrimination by nearly half compared to human-led interviews. The AI achieved this by capturing more relevant data, identifying a median of nine key topics per interview versus five for humans.

This isn't just an HR story; it’s a talent story. It suggests that the "soft skills" we prioritize in interviews can be quantified and predicted more accurately by a machine, forcing a rethink of how we build and scale our organizations. Also, your ability to build a world-class marketing, sales, or creative team is now a data science problem.

The competitive edge shifts from who has the best recruiters to whose employer-brand signals (tone, values, benefits) are machine-readable before the first human ever shows up.

Small Headlines. Big Shifts.

» Meta's 'Hypernova' glasses land next month

Leaks point to a display-equipped model with a wristband, likely targeting glanceable AI experiences in the wild. If users stare at your product, what do these glasses say about it?

» Grammarly’s redesign + modes

A new UI and multi-mode writing experience aim to reduce friction in long-form drafting and revision for teams already bought in. Enterprise guardrails still apply; watch how it integrates with your IT stack.

» DeepSeek v3.1 ships (modest uplift)

The Chinese startup pushed an incremental update focused on stability and quality. For GTM, treat this as a possible cost-quality play.

» Meta x Midjourney licensing

Meta will license Midjourney’s “aesthetic technology” for future models and products. Expect higher-polish image/video across Meta surfaces; terms and timelines are not public.

» Pixel 10 debuts with more on-device AI

Google’s new phones expand multimodal features and camera-adjacent smarts, which typically trickle into consumer search and shopping behaviors.

» Mustafa Suleyman’s essay on “seemingly conscious AI”

A useful frame for governance involves avoiding anthropomorphic traps, designing for guardrails, and preserving human-to-human context in product UX. Definitely worth a read.

The lesson is simple: AI isn’t failing. Companies are. Bad systems, weak governance, and leadership theater are what kill momentum.

The real winners will be the leaders who redesign workflows around AI.

We’re building a space for those leaders. Next week, our AI Ready CMO community goes live. If you want to be in the room where these problems actually get solved, join the waitlist now: join.aireadycmo.com

See you next week,

Peter & Torsten

PS: Use your unique referral code to unlock even further discounts: