# How Great Leaders Balance Innovation and Responsibility in the Age of AI > Source: > Published: 2026-07-22 10:42:03+00:00 TL;DR — Key Takeaways **Successful adoption starts with the customer problem—not the technology—and requires a clear explanation of what the system delivers.****Leaders must be honest about workforce disruption, involving employees and boards before decisions are final rather than announcing consequences afterward.****Governance works best when risk, compliance and control teams participate from the ideation stage instead of reviewing a finished product at the end.****Responsible AI innovation means judging every system by the revenue it generates, the costs it saves and the value it creates—not by how advanced it appears.** The hardest part of adopting AI inside a large organization right now is the cost. That surprises people, so let me explain it the way I see it. On a per-unit basis, token costs are coming down, but the tasks organizations now hand to AI are far more complex, and complex work consumes far more tokens. [Gartner](https://www.gartner.com/en/newsroom/press-releases/2026-03-25-gartner-predicts-that-by-2030-performing-inference-on-an-llm-with-1-trillion-parameters-will-cost-genai-providers-over-90-percent-less-than-in-2025) makes the same point: It expects the cost of running advanced models to drop by more than 90% by 2030, yet warns that cheaper tokens will not translate into cheaper enterprise AI, because agentic systems consume five to thirty times more tokens per task than a standard chatbot.1 As consumption climbs faster than price falls, the total bill goes up. The companies living this say so out loud. Uber burned through its full-year 2026 budget for AI coding tools four months into the year, then capped each employee at $1,500 a month per agentic coding tool. ServiceNow meanwhile became the second public company to say the same thing, exhausting its full-year budget for those tools within the first few months.2 Do that at scale, across a hundred thousand employees or more, and it adds up fast. The promise of these systems holds. Uber’s own chief operating officer has said publicly that he cannot yet draw a straight line from the company’s rising coding-tool spend to the consumer features it was meant to accelerate. This is the uncomfortable position more and more leaders are about to find themselves in. I start here because cost is the responsibility leaders are least prepared for, and how a leader handles responsibility is the measure of this moment. Innovation is the easy instinct. Everyone feels the pull to move faster. What separates the leaders who do this well is how seriously they take what they owe the people and the budgets on the other side of the decision. I think about that responsibility in three parts, [with cost](https://techstrong.ai/features/the-ai-stack-and-the-indispensability-trap/) threading them all together. ### The Customer The first is the customer. You can build something useful and still watch it fail if you hand it over without explaining why it matters. Pushing new technology at someone who doesn’t have a problem to solve gets you resistance instead of adoption. Before anything ships, there has to be a direct conversation with the customer that sets expectations and answers the only question they’re asking: what does this do for me? Get that exchange right and your customers carry the adoption forward themselves, instead of leaving you to push it uphill for months. ### Jobs The second is internal. Any serious use of AI will displace some roles and create others. Pretending otherwise costs a leader credibility they will need later. Your employees and your board should understand the range of outcomes before those outcomes arrive, which means planning for the best case and the worst case and keeping the people affected by both inside the conversation. The organizations that mishandle this treated it as a message to deliver at the end, when it was a decision to share at the start. The financial case cuts both ways though. While everyone is looking at the headcount it saves, no one is looking at how many tokens it burns. ### Governance Innovation and governance get framed as opposites, the creative impulse pulling against the control function. My experience runs the other way. You need room inside the organization for the creative work, the experimentation that shows how a technology can help a customer or an internal process. The harder question comes after something proves its value: how do you bring it into the main operation and scale it. That is where governance earns its place. The approach I’ve seen work is to put the risk, control, and compliance partners in the room from the ideation stage, and keep them through development and rollout. Because they are part of the discussion from day one, their input shapes the product as it is built rather than arriving as a verdict at the end. By the time the work is ready, they have already seen their controls implemented, so they become supporters of it. The same stakeholder who would have slowed down a finished product will champion one they helped shape. I am deliberately careful with the word ethics here. The ethics of AI is a deep field with its own specialists, and that depth belongs to them. I defer to the experts rather than pretending I can settle it in a paragraph. But deferring isn’t the same as dodging, so let me be concrete about the part I can speak to. A system’s output has to be bias-free, in whatever form that bias takes for the problem at hand. It also has to be a real answer to the question in front of it, not a response engineered in advance toward the conclusion someone hoped to reach. A model that tells you what you want to hear looks like it’s working right up until the decision it shaped turns out wrong. ### Training Leaders also owe their people training on these tools. The model I’ve seen work splits users into two groups. The power users become champions who carry the message into their own teams. The newer users get one-on-one training on the basics of what an AI system can do. AI literacy across an organization matters. It is also a day-two problem. Literacy belongs to the phase after you have done the innovation, identified what the technology will do, and begun implementing it. Training people before you know what you are training them for spends effort in the wrong order. ### Two Qualities The same two qualities keep coming up. The first is the ability to learn as you go. The technology changes too fast for a fixed plan, so you hold a small set of principles as non-negotiable and stay willing to amend everything else, even mid-rollout, as you watch how other industries adopt and where the technology moves. The second is the discipline to pass on the shiniest object. An AI system gets measured by how much revenue it generates or how much cost it saves for the organization, and by how well it serves a specific customer journey or internal process. Its intelligence is beside the point. Hold a system to what it delivers against what it costs, and the tension I started with mostly dissolves. Cost discipline and customer value are the same instinct, and that instinct is what responsible innovation looks like in practice.