PI Planning went from two days to one. Do the arithmetic and a different story appears: upfront planning falls by more than half while iteration-level learning more than doubles. Not a cut — a transfer of time from prediction to correction.
There’s a pattern in which jobs AI is automating first, and it’s uncomfortably simple: the more completely your work lives on a screen, the more exposed you are. Software AI needs no hands, no van, no job site — so purely digital, remote-capable work is falling first and cheapest. Physical presence is a genuine moat. But “physical work is safe from AI” is a dangerous half-truth, because robots are already automating the physical world too. The real distinction is sharper.
United Way of Cleveland wanted publicity and a world record, so it spent six months and half a million dollars releasing 1.5 million balloons over the city. Then the balloons came back down — clogging highways, closing an airport runway, and drifting across Lake Erie during a Coast Guard search for two missing fishermen. The event set the record and became a disaster: lawsuits, a net financial loss, a ruined reputation. Here is why Balloonfest ’86 is the definitive cautionary tale for the age of AI.
Monthly iOS app releases have climbed roughly 70% above baseline since agentic coding tools went mainstream. App reviews did not follow. Apps achieving significant usage drifted down. More software is being built than at any point in App Store history, and less of it is being used. This is what happens when an organization gets radically better at execution and invests nothing in learning.
There is a chart circulating among people who build software for a living, and it should worry every executive currently funding an AI transformation.
Drawing on research by Demirer and colleagues published this year, the Financial Times plotted two lines against each other: monthly iOS app releases, and monthly app reviews. From the point agentic coding tools went mainstream, the release line climbs steeply — roughly 70 percent above the 2024 baseline. The review line does not follow. Neither does the measure of apps achieving significant usage, which drifts gently downward across the same window.
More software is being built than at any point in the history of the App Store. Less of it is being used.
That gap is the entire story of the current moment. It is also the problem that the AI-Native SAFe framework was designed around — and the first thing its authors reach for when asked what has actually changed.
One line is worth carrying into your next leadership meeting: traditionally, efficiency comes from standardization. In an AI-Native organization, efficiency comes from adaptation.
Read that slowly, because it inverts forty years of operational doctrine.
Standardization won the twentieth century. You found the best way to do a thing, you wrote it down, you trained everyone to it, and you got efficiency through repetition and reduced variance. Every process improvement discipline most executives were trained in — Six Sigma, Lean, the entire operational excellence canon — is a machine for converting variation into standard work.
That logic held because the cost of trying something new was high. Building was expensive, so you built carefully, once, after a long process of deciding what to build.
Agentic AI collapses that cost. When a working prototype takes an afternoon instead of a quarter, the constraint moves. Your bottleneck is no longer how fast you can build the thing you already decided to build. It is how fast you can find out whether it was worth building at all.
Standardization optimizes execution against a known answer. Adaptation optimizes learning when the answer is not known yet. The chart is what it looks like when an entire industry gets radically better at the first while investing nothing in the second.
The dominant organizational approach to AI right now can be stated in one sentence: build a lot of things, ship them, and hope that a problem turns up which the things we just built might address.
That is not a caricature. It is the observable pattern in the data. Release volume is the metric that responds to agentic tooling, because release volume is the metric that tooling can move without anyone changing how decisions are made. Usage is the metric that requires an operating model.
The alternative is to start from the outcome you want to achieve and treat everything downstream as a hypothesis. That sounds obvious. Almost no organization does it. Most start at outputs — a roadmap, a backlog, a list of features someone committed to in January — and reverse-engineer outcomes afterwards to justify them.
The loop is deceptively simple: outcomes, priorities, outputs, measurements, value — with flow moving forward and feedback moving back at every stage. Underneath it sit the things that make the loop actually run: intent, context, specification, performance signals, and curated data feeding the whole apparatus.
What matters is the direction of causality, and what it implies about the shape of the work.
A set of outcomes generates a wide field of experiments and prototypes, most of which die. What survives narrows into enablers and features. Only at the far right of the funnel does customer and business value appear. Exploration is not the wasteful part of the process that discipline eventually squeezes out. Exploration is the process. Features are what is left over when exploration has done its job.
This is where most transformation programmes will quietly fail. It is straightforward to buy agentic tooling and watch output volume rise. It is organizationally painful to fund a portfolio where the majority of work is expected to be discarded, and to defend that funding to a CFO reading a dashboard that cannot distinguish a discarded experiment from a wasted quarter.
A human-centric approach here is a design principle, not a reassurance for nervous staff. Three commitments carry the weight.
Responsibility sits with the people who approve the work. Not with the tool, not with the vendor, not with the team that ran the prompt. Whoever signs off owns the output. This single sentence resolves most of the accountability confusion currently paralyzing AI governance committees.
Focus shifts from manual production to discovery and high-value problem-solving. If your people spend their reclaimed hours producing more of the same artifacts faster, you have bought a faster horse.
Verification is mandatory, not optional. Active verification is what keeps AI amplifying human judgement rather than substituting for it. An unverified AI output that enters a decision chain is not a productivity gain. It is an unpriced liability.
Every one of those is a governance choice, not a technology choice. Nothing in your model stack determines any of them.
Readers who have worked through the Mutation Readiness diagnostic will recognize the terrain immediately.
Efficiency-through-adaptation is precisely what Structural Flexibility measures — the engineered ability to reshape faster than competitors can retool. The exploration funnel is Ambidextrous Capital: the willingness to fund bets the current dashboard cannot yet justify. The verification mandate is Ethical Guardrails operating as velocity rather than as brake. And the whole outcomes-first inversion depends on Signal Sensitivity, because you cannot start from outcomes you have no instrumentation to detect.
Organizations scoring Mutation-Blind on those dimensions will adopt agentic tooling successfully and reproduce that chart in miniature — a steep release line, a flat usage line, and a leadership team wondering where the productivity went.
One: what percentage of our current delivery portfolio is explicitly funded as exploration — work we expect to discard? If the answer is zero, or if nobody in the room can answer it, you are running a standardization operating model with AI tooling bolted to the side. Your output will rise and your outcomes will not. This is measurable this quarter, and the number is usually smaller than leaders expect.
Two: when an AI-assisted deliverable reaches a decision-maker, who is named as the verifier, and what did they actually check? Not who is nominally accountable. Who performed the verification, and what specifically did they examine. If verification is implied rather than assigned, responsibility has already leaked out of your organization and nobody has noticed yet.
Three: what is our feedback loop from released value back to the next set of priorities, and how long does one lap take? If the honest answer is measured in quarters, adaptation is not available to you at the speed the market is currently moving. The lap time is the single most predictive number about whether your AI investment will produce outcomes or just artifacts.
The uncomfortable implication of the research is that the productivity gains from agentic AI are real and measurable at the point of production, and almost entirely absent at the point of value. Both things are true simultaneously. The tools work exactly as advertised. The organizations using them do not.
This is not an argument for slowing down. It is an argument that speed without a learning loop is just a faster way to arrive somewhere nobody wanted to go. The organizations that will separate from the pack over the next eighteen months are not the ones with the best model access. They are the ones that rebuilt how they decide what to build, before they made building cheap.
The tooling was never the hard part. It is now the cheapest part. The hard part is building an organization that can change its mind quickly enough to deserve the speed it just bought.
The world has changed. The leaders who notice will be the ones the next decade is built around.
Innovation is romanticized as a flash of genius. BYD tells the truer story: it started bad at cars and became the best in the world by mass-producing its way through being bad — thousands of repeated, unglamorous cycles that compounded into capability no competitor could match. Innovation does not come from the task you are already good at. It comes from repeating, at volume, the task you are bad at until you are not.
Howard Moskowitz proved there is no perfect spaghetti sauce — only perfect sauces, one tuned to each cluster of people. Executive coaching has spent three decades in exactly the wrong search, hunting for the one right method. GROW, ontological, positive psychology, CBC, co-active — none is the answer, and all of them are. But the bliss point carries a second warning coaching rarely admits: the point where it feels best is not the point where it works.
Most leaders know Geoffrey Moore for one book. In fact he built an evolving body of work across four decades — Crossing the Chasm (1991, revised 1999 and 2014), Inside the Tornado, Zone to Win, Escape Velocity, The Infinite Staircase. Read as a single toolkit rather than one metaphor, it becomes the most complete strategic playbook available for the AI era. This is the second of three pieces reimagining Moore for AI.
Sheila Taormina had one of the best aerobic engines her coaches had ever seen — and she kept improving it, year after year, while missing what actually held her back. Then she learned Eliyahu Goldratt’s Theory of Constraints: every system is limited by its single slowest step, and improving anything else is wasted effort. She found her real constraint, trained it, and became the first woman to reach the Olympics in three different sports. Here is why most AI transformations are optimizing the wrong part of the system.
In 1991, Geoffrey Moore identified the chasm between early adopters and the early majority — the gap where most technology products die. Thirty-five years later, AI has produced a chasm of a different kind. The technology has already crossed at the individual level. The enterprise is stranded between pilot and production. This is the first of three pieces on why the classic Chasm needs reimagining for the AI era.
David Solomon, CEO of Goldman Sachs, made a striking admission: AI is partly a “cover story” — the narrative that unlocks the budget for a much broader digitization push, and that justifies writing over $500 billion in data-center checks this year alone. His own bank’s research found no economy-wide link between AI adoption and productivity. Here is how to read the money.
Tell us about your goals — we'll be in touch.
