Merge CEO Speaks Out: Executives Who Don't Understand AI Are the Organization's Ceiling

Executives Are the Ceiling of AI Capability

Merge CEO Shensi Ding made a sharp point in a recent interview: if you're an executive, you are the ceiling of AI fluency for your entire organization. She leads the non-engineering side of the business, yet still writes a substantial amount of code herself. This posture of "non-technical executives getting hands-on" is itself the most direct demonstration for the team.

Behind this judgment is an often-overlooked management logic: the adoption of AI tools has never been driven from the bottom up. When subordinates observe that leaders themselves don't use, understand, or evaluate these tools, the organization's willingness to adopt them decays quickly. Ding's personal practice offers another path: managers become power users first, then talk about organizational rollout.

From Zero to One: Building a Feature with Astra

Ding admits she had never used Codex before, only trying it for the first time after Astra launched—and was "stunned" by its capabilities. She then used Astra to complete a real piece of product development: a feature needed to be redesigned, but she couldn't think of the best approach offhand, so she asked Astra directly—could it reference how other parts of the product were implemented, understand what the existing design system looked like, and then offer a best attempt.

The key to this request is that it wasn't asking AI to write an isolated piece of code, but to first build an overall understanding of the design system before producing a spec-compliant solution. This is precisely where most AI coding tools tend to break down: strong at single-point generation, weak at cross-module context consistency.

Design System refers to a set of reusable component libraries, style specifications, and interaction patterns used to ensure visual and behavioral consistency across multiple modules of a product. For AI coding tools, understanding the design system means they can't just generate code from local prompts—they must be able to retrieve and follow existing tokens, component naming, and layout conventions. This is also a key metric for evaluating the "context capability" of AI coding tools: whether they proactively build awareness of the codebase's overall structure before generating, rather than producing isolated fragments.

What Astra Gets Right: Sustained Execution and Common-Sense Completion

Ding compared Astra with other AI systems and offered two specific observations.

First, persistence in task execution. She noted that other AI systems sometimes "forget" earlier requirements and only execute the last instruction. Astra, by contrast, keeps following up on each task and executes them all in order. Context retention across multi-task instructions is the dividing line for tool reliability—it determines whether users need to restate the background in every conversation.

Second, common-sense completion. Astra not only completes what's explicitly stated but also uses common sense to fill in the gaps that weren't explicitly specified. For users without an engineering background, this matters especially, because their prompts are often imprecise and require the tool to actively infer intent.

Context retention refers to an AI system's ability to persistently remember prior constraints and requirements across multi-turn conversations or multi-task instructions. Limited by model context windows and how instruction priorities are handled, some systems exhibit "forgetting" of early instructions and respond only to the latest one. Sustained execution of multiple tasks requires models to have mechanisms for task decomposition, state tracking, and item-by-item verification—this directly determines whether users need to repeatedly restate background, and is an important dimension for measuring whether a tool can be used for real production work.

Unlocking Productivity for Non-Engineering Teams

Ding particularly emphasized that this set of tools has the most significant impact on team members without engineering backgrounds. In the past, when they had an idea and wanted to build something useful for their own workflow, they had to wait for engineering colleagues to schedule assistance; now they can think of it themselves, build it themselves, and put it to use immediately.

The significance of this change goes beyond efficiency. It changes the power structure within the organization of "who can build tools"—the producers of tools expand from a small number of engineers to everyone who can clearly describe a requirement. Ding summed it up in one sentence: now they can build anything.

Prompt precision refers to how clearly users specify goals, constraints, and boundaries when entering instructions. Users with engineering backgrounds can typically write structured prompts with clear boundaries, while non-engineering users often rely on vague natural-language descriptions. AI systems with common-sense completion capabilities actively infer unstated intent and fill in missing constraints, thereby lowering the barrier to use. This capability shifts the threshold of "can you build tools" from programming skill to clarity of requirement description—which is exactly the key prerequisite enabling non-engineering members to build autonomously.

Implications for Managers

Ding's sharing points to a pragmatic conclusion: the starting point of AI transformation is not a procurement budget, nor a training plan, but whether managers themselves are truly using it. Her approach provides a replicable sample—pick a real, non-demo work task, do it from scratch with AI, and experience its real boundaries in multi-task retention, context understanding, and common-sense completion.

When executives themselves can say "where the tool forgets things and where it fills in," the organization's rollout gains credibility. The height of the ceiling depends on whether the person standing on top has personally touched these tools.

One-line Summary

Whether executives personally use AI determines the ceiling of an organization's AI capability.

Paragraph Summary

Merge CEO Shensi Ding argues that executives themselves are the ceiling of an organization's AI fluency. She illustrates through her own experience: after first using Astra, she had the AI first understand the existing design system before completing a feature redesign, rather than generating code in isolation. She notes that Astra outperforms other systems in sustained multi-task execution, context retention, and common-sense completion of unstated needs—helping non-engineering members most, shifting them from waiting on engineering schedules to building tools autonomously. The conclusion: AI transformation starts not with budget or training, but with managers personally using AI to complete real tasks, thereby credibly driving organizational adoption.