Bernard Marr published a piece on April 24 that introduces an idea worth paying attention to. He's calling it strategic restraint, being deliberate about where you don't apply AI. His argument is sharp: in 2026, knowing when to say no to an AI rollout is a mark of organizational maturity, not hesitation. The bottleneck in AI adoption, he writes, is no longer technology. It's people, skills, governance, and judgment.
His lead case is Klarna. The payments company scaled AI in customer service fast, saw the efficiency gains, and later admitted they had over-automated, damaged the customer experience, and had to bring human workers back in. Expensive, but recoverable at their scale. The quieter version of that same story, at smaller scale, doesn't get a second chapter.
What makes that kind of restraint so hard is the asymmetry of the decision itself. The upside of saying yes to AI is visible and immediate. You can point to the efficiency gain, the speed, the demo that wowed the leadership team, the headline you get to put in the next investor update. The downside of saying yes in the wrong places is much harder to see at first, and shows up later in quieter ways. A strained customer experience. A relationship that weakens. A reputation that slips just enough that you can't fully recover it. A customer who would have stayed another decade decides the experience isn't what it was. None of those costs show up on the same timeline as the savings, which means the pressure to say yes is always louder than the wisdom to say no.
That asymmetry lands especially hard in smaller businesses. Customer relationships, reputation, and trust aren't one asset among many at that scale. They are the asset. When a large company loses a percentage of customer goodwill, it gets absorbed somewhere in a quarterly report. When a small operation loses it, there's no brand equity cushion to absorb the loss. You just have less business. Which is a quiet way of saying that the cost of an AI misstep doesn't scale linearly with company size. It lands harder the smaller you are.
So what does restraint actually look like in practice?
You pick one problem that's genuinely costing you time, money, or attention, and you solve that one first. Not five. You let the tool earn its place by showing a real result, not by looking impressive in a demo. You keep the human in the loop wherever judgment, nuance, or relationship is part of the work, because judgment is basically the whole job and you don't hand that to something that doesn't have to live with the outcome. You resist the urge to rip out a working tool to chase a newer one, because most of the shift happening right now doesn't require you to tear anything up. And you give every investment enough time to compound, because returns in these shifts always lag, and the owners who stay patient are the ones who end up with something that actually works.
None of that is the absence of strategy. That is the strategy.
The quieter thing Marr's piece points at, and the thing I keep sitting with, is that strategic restraint has quietly become the moat. For a decade, the competitive advantage in technology was being first. Being the earliest adopter, the fastest mover, the one with the newest stack. In the AI era, a different kind of advantage is taking shape. The things AI can damage fastest are the things you can't rebuild on demand. Trust. Reputation. The feeling a long-term customer gets when they call you and a person picks up. Those are the assets nobody can buy their way back into once they're gone.
Holding off isn't indecision. It's protection. And in a market where everyone's rushing in the same direction, it's starting to look a lot like the moat.