The Shadow AI Problem Nobody's Talking About

The Shifting Role of the CTO in the AI Era

The role of a CTO has fundamentally evolved with the widespread availability of AI. Mert Öztekin explains that the job is no longer just about responding to business needs for features and ensuring platform stability. Now, CTOs must proactively shepherd the entire organization through a massive transformation. This involves leaving the comfort zone of technology to push boundaries across all departments, helping them understand the impending impact of AI on their processes and the ecosystem at large. It’s a shift from being responsive to requests to actively pushing the business forward, building the foundational platforms that enable this change.

Strategies for Enabling AI Productivity

To navigate this transformation, a key focus is on enabling widespread productivity. Two slightly different, yet complementary, approaches were discussed:

• Bottom-Up Experimentation: Euro Beinat advocates for providing tools to everyone in the organization, allowing them to experiment and discover use cases from the bottom up. The rationale is that the central AI team cannot possibly envision all the specific, niche problems that exist across different business units. By giving people the tools within secure guardrails, the company benefits from “collective discovery and collective awareness.” This approach empowers employees to build their own solutions and, in doing so, gain a deeper intuition for how these technologies work, where they excel, and where they fail.

• Focused Adoption: Mert Öztekin describes a more guided approach. Instead of opening the floodgates to all tools at once, they started by rolling out specific tools like GitHub Copilot, followed by Amazon Q and Cursor. This strategy helps focus the organization, build trust in the technology, and consolidate learnings. By providing internal training and guidance on how to get the most out of a specific tool, they aim to prevent users from jumping between tools without giving any single one a fair chance. However, the ultimate goal remains the same: achieving near-100% adoption. As Mert puts it, the threat isn’t AI replacing you, but “someone using AI will replace you.”

Beyond Code Generation: The Real Productivity Bottleneck

While AI coding assistants can generate 30-40% of new code, this hasn’t translated into a corresponding increase in overall development velocity. The speakers agree that software development is much more than just writing code. The primary bottleneck is often not the “hands on the keyboard” part but the alignment between teams, product managers, UX designers, and QA. While AI is beginning to help in these other areas, optimizing only the coding part yields diminishing returns.

Euro notes that while measuring coding productivity is possible, the more important question is, “Are you measuring the right thing?” The focus should be on the entire value stream, not just one component.

The Long Tail of Agentic Automation

A significant opportunity for AI lies in what Euro calls “tiny total addressable markets.” These are niche, repetitive problems faced by a small number of people—sometimes just one person—within an organization. A traditional software company would never build a product for such a small market.

This is where low-barrier agent creation platforms become powerful. By making it extremely simple for anyone to build their own automations, companies can unlock a “long tail” of productivity gains. An employee in marketing, for instance, can build an agent to solve a specific, recurring problem without needing to wait for an engineering team. When multiplied across thousands of employees, these small, individual automations become highly impactful for the organization.

The 30,000 Agents Goal: A Change Management Strategy

Prosus has set a bold goal of having 30,000 AI agents in production by March. Euro clarifies that this is primarily a change management and cultural initiative, not a technological one. The goal is to remove the psychological barrier that building agents is a complex task reserved for engineers.

By sitting with marketing or HR teams and creating a useful agent in 15-30 minutes, they demonstrate that the tools are accessible to everyone. This process encourages a percolation of knowledge, where employees share their creations and inspire others. The initial focus is on adoption and experimentation—getting everyone comfortable with building. The process of identifying the highest-value agents and optimizing processes comes later. Mert likens it to teaching a child to read: “I don’t care if she’s reading the stories that I want her to read… I want her to get better, more comfortable with what reading is.”

Governance in the Age of AI Sprawl

With thousands of employees creating agents and adopting new AI tools, governance becomes critical. The key is to implement governance that protects the business without creating bureaucracy that stifles innovation.

Key principles for effective AI governance include:

• Practical Policies: The problem of “shadow IT”—departments buying and using their own tools—existed long before AI. The solution is not to block them but to educate them on security and procurement best practices, helping them integrate tools securely (e.g., with single sign-on).
• Automated Guardrails: Governance should be automated where possible so that people feel protected without feeling restricted.
• Human Supervision: Given that LLMs can hallucinate and make mistakes, it’s crucial to maintain a level of human supervision. Agents should not be allowed to operate with complete autonomy in high-stakes situations. This can be implemented through checks where a human must approve a critical action (e.g., “Do you really want me to go to that site?”).
• Setting Clear Boundaries: Establish firm rules, such as prohibiting the use of specific sensitive data types with external tools or limiting an agent’s ability to send bulk emails.

Budgeting for Exponential AI Growth

The cost of AI is a paradox. While the unit cost of intelligence (e.g., cost per million tokens) is rapidly decreasing, the complexity and token requirements of new use cases are increasing even faster. This results in a net increase in overall spending.

To manage this, organizations must focus on:

1. Optimization: Many initial prompts and workflows are highly inefficient. There is significant room to reduce costs by optimizing prompts and redesigning processes.
2. Model Selection: Using powerful, expensive models for every task is wasteful. Open-source models, which can be self-hosted, are becoming competitive for many practical use cases and offer a more cost-effective alternative.
3. Focusing on ROI: Ultimately, the conversation with finance should be about the return on investment. The concern is not just about managing an expense line but about enabling transformative business growth. As Mert states, the real question is how AI will help “grow the company two, three times in five years’ time.” The investment in AI is a strategic imperative to avoid being outpaced by competitors.

• Tags:
• AI Agents
• AI Governance
• Developer Productivity
• MLOps
• AI Strategy
• Change Management