This site uses cookies to enhance your browsing experience and deliver personalized content. By continuing to use this site, you consent to our use of cookies.
COOKIE POLICY
The GenAI divide is an operating model gap, not a technology gap.
Walk any college campus and you’ll find them: worn tracks cutting diagonally across the lawn, ignoring the sidewalk that runs the long way around. Landscape architects call them desire paths: the record of where people actually want to go, written by their feet.
Your organization has them. Someone in finance is pasting a spreadsheet into a personal AI account because the sanctioned tool doesn’t do what they need. They aren’t being reckless. They found a faster route, and the official path doesn’t go where the work happens.
Most enterprises respond by running another pilot. MIT’s NANDA initiative reviewed more than 300 enterprise GenAI deployments and found roughly 95% produced no measurable P&L impact. Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027. The models work. What’s failing is everything around them.
The most common mistake I see is buying a race car and putting it on a congested highway.
An enterprise licenses Copilot for 40,000 people, sends the announcement, and waits. The tool is fast. The road is unchanged. Same approval gates, same handoffs, same quarterly planning cycle. You’ve introduced enormous speed at the task level into a system that was never rate-limited by task speed.
So people use the race car to drive the route they already drove. They use AI to write the PRD faster. Nobody asks whether the PRD is still the right artifact, whether a prototype could serve as the spec, or whether the review cycle it feeds still needs to exist. Deloitte found only 34% of organizations are reimagining products, services, or business models around AI. The rest are layering it on top of what already exists.
Top-down mandates produce this reliably. A mandate without direction is just pressure. Absent a destination, everyone speeds up the route they already know.
The quieter reason pilots stall has nothing to do with the models.
AI magnifies every gap in how your organization shares context. A model is only as good as what it can see, and what it can see is a fraction of what the person prompting it knows. Decisions live in DMs. Rationale lives in someone’s head. Point a capable model at that and you get confident output built on a partial picture.
Ask a model to justify a pricing decision and it will hand you a beautifully reasoned answer that ignores the three constraints your team argued about last quarter, because that argument happened in a meeting nobody wrote down.
Weak context doesn’t just limit AI. It makes AI expensive. You pay in tokens and in validation time for every gap the model has to guess across, and you pay again when someone downstream catches what it missed. Making decisions, domain language, and prior work retrievable is the substrate everything else runs on.
Bolted onto the old path: McDonald’s ended its IBM drive-thru voice test after viral misorders. The ordering process was never redesigned for noise or corrections. Air Canada was held liable when its chatbot answered from a different source of truth than its own policy page. Zillow Offers wound down with a $500M+ write-down after a pricing model built for a stable market kept running as prices swung.
Built where the work happens: GitHub Copilot went into the editor developers already lived in, and 73% say it keeps them in flow. Klarna rebuilt its support split to match how cases actually break down. Moderna let any employee build their own assistant instead of handing down one sanctioned tool. More than 750 appeared in 2 months, and legal reached full adoption without a mandate.
Same technology. The difference is whether anyone redesigned the road.
The opposite extreme, handing everyone a license and letting a thousand flowers bloom, doesn’t work either.
What works is waves. Small cross-functional groups given a real workflow, dedicated time, and explicit permission to get it wrong. Then you harvest: every wave surfaces 2 or 3 people who become evangelical about what they found, and you inject them into the next one. Champions carrying context forward.
Start with senior people. A principal engineer gets more out of these tools than someone early in their career, because they know what good looks like and use AI to support their judgment rather than substitute for it. Adoption is a trust problem before it’s a training problem.
The mandate used to be optimize: take cost out. Increasingly it’s show me revenue. That’s a harder assignment for a CIO, who sits nowhere near the revenue line, and it’s worth naming plainly.
In PwC’s CEO Survey, only 12% of CEOs say AI has delivered both cost and revenue benefits, while 56% report no significant financial benefit at all. The ones capturing both aren’t running better pilots. They’re 2 to 3x more likely to have embedded AI structurally.
One caution on cost: today’s inference prices are partly subsidized, and token spend compounds. Don’t marry a single frontier provider.
And set the expectation correctly on what AI actually returns. It buys you quality more reliably than it buys you speed. The honest version isn’t four hours of work now takes one. It’s four hours still takes four hours, but you ship v4 instead of v1. Effort doesn’t disappear. It migrates out of execution and into the two ends: the context going in, and the validation coming out. Promise the wrong one and you’ll be explaining a miss you never needed to have.
Most firms treat adoption as a phase. Build the thing, then go convince people to use it. We run adoption groups inside delivery instead. The people who will live with the software shape it while it’s being built, so by the time it ships they aren’t being trained on it; they helped make it. Change management stops being a launch event and becomes a property of how the thing got built. Adoption baked in, so nobody has to be sold on it at launch.
It changes where an engagement starts, too. How you frame a problem determines the solutions you’ll get. Most engagements open with requirements. We’d rather open one question earlier. What is this organization actually trying to be good at?
Not with a platform. Not with another pilot.
Name the single system, whether that‘s strategy and planning, discovery, or delivery, where your biggest AI upside and your worst real bottleneck collide. Pick the one metric that would prove a fix. Redesign that operating model, and only that one. Then take the same lens to the next path.
Your people already found the shortcut. The work is paving it.
With over 20 years at the forefront of technology innovation, I've dedicated my career to delivering strategic solutions that drive business growth. As the Chief Technology Officer of UDig, I lead our technology vision, architecting solutions that transform how organizations leverage technology to generate impact.
