April 2026 AI Insights

I spent the last month delivering programs, workshops and keynotes explaining what’s happened so far in 2026 and explaining how productive individuals do not equal productive organizations.

I had executives respond saying, “yes, that’s exactly where I’m stuck” followed by, “do you think we’re ahead or behind?”

Wrong question! Here’s why: If you’re ahead, the curve just moved. If you’re behind, you already know. The question you need to be asking instead is: what are you actually doing about it?

The reality is that organizations are stuck in the capability gap between what the technology can do today and what they’ve been able to adopt into the business. Yet some organizations are racing ahead. This is what I call the “95% Paradox.”

The 95% Paradox

An MIT study found that 95% of enterprise AI pilots fail to produce measurable returns. Same window, AI-native companies and startups are scaling faster than any cohort in technology history. Cursor in talks at a $50B valuation on $6B projected 2026 revenue. Anthropic from $19B to $30B annualized in a single month. Vercel approaching IPO with AI agents generating 30% of deployments.

Same technology. Polar opposite outcomes.

That’s the paradox, and it’s why “ahead or behind” stopped being a useful question. The gap between two companies in the same industry, with the same headcount and the same budget, is now wider than the gap between industries used to be. One of them is in the 95%. One is in the 5%. The difference has very little to do with the tools they bought.

April made this impossible to ignore.

The Month the Tech Stopped Waiting

The AI 2027 researchers updated their forecast. Expert-level AGI is now arriving roughly 18 months sooner than they predicted last year, with capability doubling every 4 to 4.5 months. If you have a three-year AI roadmap, you have an 18-month problem.

Anthropic withheld Claude Mythos from public release because of its offensive cyber capabilities, and launched Project Glasswing as a coordinated disclosure coalition. Two weeks later, the NSA was deploying Mythos Preview to scan their own systems while the Pentagon was still calling Anthropic a supply-chain risk. Same government. The model also surfaced a 27-year-old vulnerability in OpenBSD and triggered an emergency Federal meeting. Whatever you think the safety frontier looks like, it’s already past where governance is.

Jack Dorsey cut 4,000 jobs at Block, 40% of the workforce. The detail that mattered was the sequence. He and his leadership team spent the holidays inside Claude and Codex. They came back, looked at what they’d built and learned, concluded the company “would not look the same, or be the same size,” and moved within three weeks. Most executives I talk to are running AI experiments in one room and headcount conversations in another. Dorsey put them in the same room.

Apple sent 200 Siri engineers to a multi-week AI coding bootcamp, two months before WWDC. The curriculum is Claude Code and Codex. The most well-funded engineering organization in the world publicly confirmed that one of its teams was in the 95%, and treated it as an emergency rather than a training program.

Microsoft, in the other direction, formally labeled Copilot “for entertainment purposes only” inside parts of its enterprise tier. Read that twice. The model is fine. The deployment around it isn’t.

This is what the 95% looks like up close.

The Trap Inside the 95%

Here is the part that keeps me up.

A new paper this month observed something I’ve been watching for a year. They call it the LLM Fallacy. When AI helps you produce something good, your brain takes the credit. The output was yours, so the capability must be yours. Confidence rises while real skill drops. They use the GPS analogy. The directions keep working, and your internal map keeps degrading, and you don’t notice because nothing visibly breaks.

A separate MIT study found measurably less brain activity in students who used AI for essays than those who didn’t. A CNN investigation at Yale caught students feeding seminar questions into chatbots and presenting the output as their own thinking. Twenty people, twenty versions of the same answer.

Most enterprise AI programs are built on the assumption that the goal is to get more people using the tool. They measure adoption. They reward usage. Meta ran an internal token leaderboard tracking 60 trillion tokens consumed across 85,000 employees in a month. They killed it last week after the backlash, but the instinct that built it is everywhere. That number sounds like progress. It’s the perfect metric for the 95%, because it answers the wrong question. High usage with no redesign is how you build capability into the tool and out of the person at the same time.

That’s the trap. AI bolted onto existing roles makes individuals feel more productive while the organization stays exactly as stuck as it was. The CIO sees usage going up and pilots failing and can’t square it. This is how.

Redesigning the Factory

When Amazon arrived during the early internet, Barnes & Noble put books on the internet. Same technology. Wildly different outcome. The 5% are not running better pilots. They’re redesigning the factory.

Across every engagement I’ve run this year, the work breaks into four layers. They only produce results when all four are moving at once.

Personal tooling. The individual layer. Not “AI literacy” or prompt training. Hands on the tools, every day, until the hesitation overhead from 2022 is gone. The fastest movers in any organization aren’t more technical, they just never stopped giving themselves permission to try. Most companies are still trying to skip this layer with classroom training. It doesn’t work. You build it through reps.

Team productivity. Once you have power users, you deploy them across the business. The pairing is where the lift compounds. A power technical user (or forward deployed engineer) sitting with the marketing team finds workflows neither of them would have seen alone. This is the layer where most organizations stall, because it requires explicit permission and structure, and most leaders treat it as something that should happen organically. It won’t.

Redesigned workflows. This is where the role itself changes because it’s been designed from the ground up with AI at the center and optimized what is uniquely human. You have to look at three pillars: the skills that appreciate when a digital coworker handles execution, a vision that says what people are freed from and freed to, and a growth path that reflects judgment complexity rather than task volume. If you can’t articulate what your senior people are now freed to do, you haven’t redesigned anything.

Transformational bets. The big plays. New products, new business models, new operating models that wouldn’t have been possible eighteen months ago. This is the layer the board wants to see, and it’s the layer that fails most predictably when the bottom three layers haven’t been built. You cannot place transformational bets on top of an organization that hasn’t built personal fluency.

The 5% are working all four. The 95% are picking one and hoping it’s enough.

I’ve watched this play out inside our own walls and across client engagements this month. The pattern holds every time. Power users get built through hands-on, not through training. Pairing them across functions is where the productivity actually shows up. Workflow redesign is what makes the productivity stick and hits the bottom line. And the transformational bets only land if the foundation underneath is real.

Signals from the Edges

Five April stories that didn’t dominate the news cycle, but give you clues as to where this might be heading.

• Lila Sciences ran the scientific method autonomously, producing 10x and 700x improvements over human baselines on real lab problems.
• A Unitree humanoid robot hit AliExpress for $4,900, a 99.4% price collapse in two years.
• Tesla and DoorDash-linked firms paid gig workers $25/hr to film themselves doing household chores, training the robots that will eventually do them.
• A Google/MIT multi-agent system autonomously discovered a depression biomarker in wearable data, compressing 37 person-days of research into eight hours.
• A small Missouri town ousted half its city council over a $6B AI data center, a preview of every planning meeting in America for the next decade.