40% of the workforce now has access to AI tools.
That same workforce has added roughly 0.1% to overall productivity. Bleak.
If your organization has provided access to AI subscriptions and is still waiting for the results to show up, you’re not alone. Most business leaders are in the same position, learning that even the most powerful tools deployed without structure, governance, or clear intention tend to produce noise rather than results.
As Pascal Bornet, an AI expert who spent over two decades implementing AI initiatives at McKinsey and EY, puts it, giving your team AI without a plan is like “giving a three-year-old a Ferrari. Lots of power but no driver’s license.” In other words, access is not a strategy.
Bornet has spent years studying what separates organizations that successfully scale AI from the ones that stall after a promising pilot. His conclusion: Fewer than 1% of companies have cracked it, and the ones that have are consistently getting five specific things right:
1. They onboard AI agents like new employees
When a new hire joins your organization, you don’t just hand them a laptop and wish them luck. You give them a job description, define what success looks like, set boundaries around what they can and can’t access, and monitor their early performance closely before extending more responsibility.
AI agents require the same deliberate onboarding, but most organizations skip it entirely.
Bornet is direct about the consequences. “When instructions lack precision, agents produce unpredictable results,” he says. One retail company he worked with learned this expensively: A poorly onboarded agent began issuing unauthorized discount codes, ultimately costing the business more than $100,000. The problem wasn’t the technology, but the absence of a clear scope, defined objectives, and explicit guidance on what the agent should (and should not) do.
The principle of least privilege applies here as it does with human staff. Just like a marketing intern shouldn’t access payroll, an AI writing assistant has no reason to see credit card data. Start narrow and expand access as trust is earned.
2. They build guardrails before they give the keys
Think about how you’d handle a teenager getting their driver’s license. You probably wouldn’t hand over the keys and say “take it wherever you want.” You set limits, like a curfew, no highways, one passenger, etc., then relax them gradually as the new driver’s confidence and responsibility develop.
Successful AI implementations work the same way. A financial services company Bornet worked with launched their customer service agent with a deliberately narrow mandate to answer basic questions and schedule appointments, nothing more. Only after the agent established a track record did they extend its scope.
This kind of structured rollout requires specific decision thresholds, automated alerts for unusual behavior, and the ability to pause agent activity if something goes wrong. It also requires observability, or logging what your agents are actually doing so you can audit their decisions, catch problems early, and identify where performance can improve. Organizations that skip this step don’t find out their agent is off the rails until the damage is already done.
3. They learn to lead in two languages
“We are the last generation of leaders who will only lead people,” Bornet says. “The next generation will lead people and agents.”
If that sounds like a distant problem, it isn’t. Organizations deploying AI agents today are already managing hybrid teams, and most leaders are underprepared for it. The typical failure mode is lurching between two equally ineffective approaches: treating AI agents like human colleagues (and getting frustrated when they miss context or nuance) or treating them like blunt instruments that just need better programming. Neither works.
What Bornet calls bilingual leadership is something different — the ability to motivate and develop human team members while also providing AI agents with the precise instructions, boundaries, and maintenance they need to perform. With people, that means helping them understand how AI makes decisions and framing it as a tool that enhances their work rather than threatens it. With agents, it means clarity, consistency, and regular calibration.
The results, when leaders get this right, are significant. One trading desk manager Bornet worked with restructured her team around this model. Within six months, trading efficiency improved by 34%, job satisfaction among human team members increased, and client relationships deepened as traders redirected their time toward strategic advisory work.
4. They scale methodically, not aggressively
The most dangerous moment in an AI implementation isn’t the launch. It’s the period right after a successful pilot, when the temptation is to move fast and replicate broadly before you’ve understood what worked.
Organizations that stumble at this stage usually do one of two things: They scatter pilots across multiple departments simultaneously without coordinating them, creating redundancy, inconsistency, and confusion … or they set expectations so high that a solid first result still feels like a disappointment, and momentum stalls.
Bornet references Pets at Home, the UK’s largest pet care company, as a model of how to do this well. They started with a single veterinary scribe agent. They built confidence through that one success. Then they expanded methodically into fraud detection, insurance integration, and store support. Each step informed the next.
The broader lesson is that scaling AI requires the same discipline as any major operational change: start where impact is measurable and the cost of failure is low, bring leadership and frontline teams into the process together, and invest seriously in helping employees build genuine AI literacy — not through classroom training that’s obsolete before it’s delivered, but by giving people dedicated time to experiment and see the results in their own work. One healthcare organization that built an AI champions network and ran regular hackathons saw adoption rates three times higher than organizations that took a top-down rollout approach.
5. They invest in the skills AI can’t replace
The organizations getting the most from AI aren’t the ones trying to automate everything. They’re the ones who are clear-eyed about what automation is actually for: Freeing people to do more of the work that machines can’t.
Genuine creativity, critical thinking, ethical judgment, and authentic human connection aren’t under threat from AI but the complement to it. “The greatest impact comes from a symbiotic relationship where AI provides technical precision and humans provide emotion, creativity, and meaning,” Bornet says. The role of leadership is to build an environment where that exchange happens, so people are relieved of repetitive, low-value work and redirected toward the contributions only they can make.
That means changing how you measure performance: less emphasis on tasks completed, more on outcomes achieved. It means rewarding the people who take smart risks with new tools, not just the ones who deliver predictable results with familiar ones. And it means recognizing that 76% of employees are already using AI tools their employers haven’t sanctioned … not because they’re being reckless, but because they can see the value and they’re not willing to wait for IT to catch up.
Your workforce is ready. The question is whether your organization is set up to channel that energy productively.
The bottom line: Go slow to go fast
When leaders ask Bornet how to accelerate their AI implementations, his answer is the same every time: “Go slow to go fast. Start slow, be thoughtful, put the right governance in place, put the right ethics around it, have the right success criteria — learn from it, get to scale, then you can go fast.”
That’s not a reason to wait, but it is a reason to start deliberately.
The technology works. Your people are more ready than you probably think. What you’re missing is the plan that connects the two.
Pascal Bornet is the author of Intelligent Automation and Irreplaceable*, and co-author of a new book on agentic AI.
