Many AI implementations fail for a reason that has nothing to do with the technology. Ben Tasker, who has built applied AI certifications for more than 225,000 learners, traces the failure to a discipline most companies never think to apply to AI at all: change management. Organizations buy tools, skip the work of building the skills to use them, and then find that the investment produced nothing they can point to. The figure he attaches to that group is 95 percent.
Eighteen months ago, I knew organizations where using AI was flatly prohibited. Most were nonprofits, but the policy was real, and employees were told not to touch it. Those same organizations are now scrambling to implement, and many are moving faster than their people can absorb. That is a different mistake with an identical ending.
What makes Tasker’s claim of 95 percent worth sitting with is the shape of the loss. These companies don’t fail because the model underperformed. They fail because, in Tasker’s phrasing, “they spend a boatload of money and they get nothing out of it,” whether the AI itself works or not, since there are no people behind it.
Ask First
The first question is not which tool. That one feels strategic because it arrives with a price tag, a vendor list, and a decision date, and it’s where most executive teams start.
Tasker’s argument is that tools have a short shelf life and skills have a long one. The vendor you standardize on this quarter may be acquired, repriced, or gone inside two years. The ability to prompt well, to fit AI into an existing workflow, and to see where it creates exposure carries over to whatever replaces it. “Just because you have AI tools doesn’t mean you have an AI strategy.”
The second question follows from the first, and it’s what separates the 5 percent from everyone else. What’s the change management plan, meaning the plan for how the organization learns, not the rollout calendar? An executive team that can answer the tool question in fifteen minutes and can’t answer this one at all has just diagnosed itself.
The Layers of Learning
Tasker describes an organization as a cake, which sounds cute until you try to use it. The base is the frontline, roughly 80 percent of the workforce, and the question is narrow: which parts of my daily work can this take over, and how do I hand them off without creating a mess? The middle layer is the AI-enabled specialists, the engineers and program managers who work behind the scenes and need enough depth to stand systems up. The frosting on top is leadership, whose job is not to use the tools at all but to explain, credibly, why any of this is happening.
Each of those three layers is asking a different question. The frontline wants to know which parts of the daily job to hand off and how to do it without creating a mess. The specialists need to know how to build and maintain the systems that make the handoff possible. Leadership needs an answer to why the company is doing any of this and what happens to the people who cooperate.
A company that buys one training license for everybody has answered the frontline’s question badly and the other two not at all, and it will find that out at the layer where 80 percent of the work happens.
Just because you have AI tools doesn’t mean you have an AI strategy.
There is a version of this that looks messy from the outside. When the whole organization is encouraged to experiment, the pain points and the failures get surfaced by people who do the work. Those feed a backlog that specialists and leaders can price, rank, and attach risk to. Tasker’s own description of it is “organized chaos instead of unorganized chaos,” and the distinction earns its keep, because the alternative to sanctioned experimentation is not order. It is the same experimentation happening quietly, on personal accounts, with company data.
Responsible AI Isn’t Just Compliance
Tasker declines to file AI risk under compliance as its own separate function. He puts it inside change management and ahead of implementation, which in practice means getting ten to fifteen people in a room before anything ships, drawn from IT, HR, and the business owners who understand the process about to be automated.
The questions are unglamorous and specific. What happens when the system misclassifies a person or a ticket? If the tool is agentic and takes an action it was never meant to take, how fast can it be shut off, and does that switch exist or is it a slide? Plenty of organizations want the most advanced version of this technology without having answered a single one of those questions at the simplest version, and that gap is where the expensive surprises live.
Kodiak is Tasker’s example of the patient version. In 2018, well before driverless technology was a boardroom topic, the company began working toward autonomous trucks inside a fleet of more than a hundred. It took the long way around: design the truck, work out which routes it could handle and which states allowed it, then run a stretch with drivers still in the cab so the paper assumptions could be tested against an actual road. Drivers were paid through the training period and the case made to them was safety rather than efficiency. The outcome Tasker reports is the one nobody predicts, which is that Kodiak ended up needing more drivers rather than fewer.
The Learning Curve
How long is the AI learning curve? According to Tasker, the honest number is eighteen months, and that assumes you are starting from zero. That is Tasker’s estimate for moving an employee from level one, where they could not reliably tell you what AI stands for, to level three, where they can prompt a model, use it for routine work, take the output, and iterate on it. Levels four and five run longer, and the specialist domains like robotics and drones can take five to six years and a degree.
Executives ask how quickly the technology can be implemented. The more useful question is how quickly their own people can be brought along, and the two answers aren’t even in the same unit. Implementation gets planned against a procurement calendar. Reskilling runs on the eighteen months it takes to move one person from level one to level three, and no vendor timeline shortens that.
Tasker’s argument for starting before you have a business case is that every job will eventually be touched by AI. It means reskilling has to begin before anyone can say which roles will need which skills. The alternative is buying each new niche skill on the open market as it surfaces, at the premium the market charges, while humans are still the ones who have it.
Hiring keeps its role for genuinely specialized work, robotics and drones among it, where you cannot train your way in from a standing start. For everything else, Tasker would write prompt engineering and responsible AI into job specs in departments that have never asked for them, partly to see what talent surfaces and partly because that hire becomes the person who brings the rest of the team along.
Adaptability ranks above every technical AI skill on his list. I have watched that shift move through HR strategy sessions over the past two years. The question used to be what a candidate knows, and it’s now how fast they can learn something that did not exist when they applied.
Fear of Job Losses
What do you say to people who are afraid of losing their jobs? The standard answer is that this is about enhancement, not replacement, and employees can see it coming a mile away. A fair number of them are right to be skeptical. Reassuring a nervous workforce that nothing will change buys one quarter of calm at the cost of the credibility you need for the next three years.
Tasker’s position is transparency with something behind it. Walmart handed AI licenses to its entire workforce and told people to use them at work or at home, which moved the question from permission to practice. Other organizations have paired formal learning tracks with a commitment to no AI-related layoffs for a defined window of three to five years, which happens to be the horizon on which the return shows up. The companies that skipped that step are the ones cutting tens of thousands of roles and hiring them back weeks later when the technology turned out to be less capable than expected.
His version of the honest conversation is that jobs are going to change, and that upskilling is what prepares people for the ones that come next. “It’s not a replacement technique.”
I would put it more bluntly to the employee weighing whether to engage. Refusing to learn is a choice, and in any field, in any decade, that choice ends in the same career cul-de-sac. AI only shortens the drive.
The number worth carrying out of all this is that 95 percent, along with a clear sense of what kind of number it is. It’s not a statistic about model accuracy or vendor quality, which is how most executives hear it. It’s a management statistic, one of the very few failure rates in this category that a leadership team can move on its own.
The companies still shopping for the right tool are not early to anything. They are spending first and learning second, which is an expensive sequence.