On the evening of June 1 at San Francisco’s Bartlett Hall, Microsoft CTO Kevin Scott spoke at a joint event with Lectures on Tap, attended by approximately 150 developers, founders, media, and tech industry leaders. His talk focused on what he described as the growing perceptual disconnect between AI capability and AI reality: the tendency to mistake rapid advances in model performance for equally rapid progress in deployment, organizational transformation, trust, and real-world value creation.
Scott argued that the AI industry is at an inflection point where technical breakthroughs are arriving faster than institutions, workflows, and human systems can absorb them. While acknowledging the extraordinary pace of progress in areas like software development and agentic systems, he emphasized that the difficult challenge ahead is operationalizing these capabilities responsibly and meaningfully at scale.
Here are his five observations of the ways in which AI reality is diverging from apparent capability.
1. Capability ≠ deployment #
According to Scott, one of the biggest mistakes people are making right now is confusing technical capability with real-world deployment. The fact that a model can do something impressive doesn’t mean the surrounding systems, economics, governance, and human behaviors are ready to absorb it at scale.
“Today’s AI models are actually more capable than the things we’re using them for in the real world,” he said, addressing today’s “capability overhang,” as he has dubbed it. “We just shouldn’t have uniform faith that, as AI model capabilities improve, we’re going to get this crazy fast deployment everywhere.”
2. Closed feedback loops ≠ universal progress #
Scott explained that some areas of AI (like agentic software development) are improving extraordinarily quickly because tight feedback loops allow those systems to iterate, evaluate, and refine outputs at high speed. But that dynamic doesn’t automatically extend to domains constrained by physical systems, regulation, or long experimental cycles.
“One of the things models can already do is postulate new ideas for particle physics experiments,” he said. “And the problem with particle physics experiments is that they take a lot of expert technical labor to set up and run, and they require the use of extremely expensive infrastructure. So there really isn’t a convenient way—other than publications in the scientific literature—to get the output of those experiments and feed it back into an actual model.”
3. Software velocity ≠ organizational velocity #
AI is dramatically accelerating software development production, but that doesn’t mean organizations can suddenly move faster. In many cases, speeding up code generation simply exposes the slower-moving bottlenecks that were already present: deployment, integration, governance, and organizational change.
“I build a lot of prototypes that are greenfield, where I have no constraints whatsoever,” noted Scott. “I just get an idea and there’s nothing stopping me from using an agentic coding system to produce a brand-new thing. But in many cases, the things we want to produce are fairly highly constrained.”
Scott pointed to last-mile problems, the need for a lot of plumbing work, and human psychology as throttling issues. He also acknowledged the forecasting problem we will inevitably face as things move exponentially faster: “A lot of the stuff I’m doing right now using agentic coding to build things wasn’t even possible in November of last year,” he said. “When things are moving this fast, it’s hard for people to notice the change and snap to.”
4. Activity ≠ value #
“Just because you’re using AI to create a lot of activity doesn’t necessarily mean that the activity you’re creating is valuable,” Scott said.
The ability to generate enormous amounts of output doesn’t guarantee meaningful impact. As AI lowers the cost of creation, the defining question shifts from, “How much can we produce?” to: “What is actually worth building?”
“We can have a lot of output, we can build more complex things than we built before,” added Scott. “That doesn’t necessarily mean that the things we’re building are super valuable. When they go into a user’s hand, are they solving a real problem? As developers, we have to pay especially close attention to how we measure value and to the feedback we get on the work that we’re doing.”
5. Autonomy ≠ trust #
AI systems are becoming increasingly capable of operating autonomously, but autonomy alone does not create trust. Real-world deployment still requires governance, identity, access control, transparency, and meaningful human oversight.
“You’re always going to have human oversight, so this notion of autonomy is a little bit of a pipe dream,” Scott noted. “You have to build systems doing complex things in a way where people can trust that they’re doing them correctly and in a way that’s aligned with their interests and values. And that’s a new way of thinking about software.”
Bridging the gap between capability and reality #
Ultimately, said Scott, “There’s a lot of work for all of us to do over the next months and years to fully unlock the potential of this crazy tool that we’ve built collectively. These problems that I enumerated don’t go away just as a function of scaling up an AI model. There is no silver bullet. That means there’s a bunch of technical work to be done, a bunch of societal work, a bunch of organizational work, and just dealing with legacy systems and plumbing.”
AI capability gains will continue, but turning those gains into trusted systems that create meaningful, durable value is the harder and more important work. And that work starts today.
“We need to engage more intensely than ever before,” said Scott, “because we can see the promise of this technology to benefit the world if we’re able to overcome these obstacles.”