# The Shadow AI Problem Nobody's Talking About Euro Beinat, Prosus Group | MLOps Community | 47:25 Source: https://www.youtube.com/watch?v=WYBX6fqQyFo Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks. Page: https://mlopstalks.com/talks/the-shadow-ai-problem-nobodys-talking-about Published: 2026-02-12 Tags: cost, governance, security, team-adoption ## TL;DR - AI has changed the CTO role from delivering technology to helping the whole company change how it works. - Giving employees safe tools to create their own agents can uncover small, useful automations that central teams would never find. - Governance should limit serious risks while keeping experimentation practical, because excessive controls reduce the value of AI. ## Summary Euro Beinat argues that AI adoption is an organizational change problem as much as a technology problem. CTOs now need to help every department understand and use AI, while building platforms with sensible security limits. Prosus gives employees tools to experiment because people closest to a task often find uses that a central AI team would miss. The long tail of small automations can add up across a large company. Beinat describes the goal of 30,000 agents as a way to make employees comfortable creating agents, rather than as a simple technology target. Governance has to account for hallucinations, prompt injection, data exposure, and excessive autonomy without blocking useful work. Mert Öztekin adds that companies also need guidance around tool procurement and adoption. AI costs require attention because lower token prices are being outweighed by more complex and more frequent use. ## Key ideas ### The CTO role now includes leading company-wide change [00:42](https://www.youtube.com/watch?v=WYBX6fqQyFo&t=42s) Mert Öztekin says the CTO role used to focus on meeting requests from product, marketing, and other departments, while keeping the platform scalable and stable. The spread of ChatGPT changed that. Technology leaders now need to help departments change their thinking, processes, and way of working. They also have to build the foundations that let people use AI safely. Beinat describes this as moving beyond responding to requests and actively pushing the business to understand what is coming. ### Employees should get room to discover useful AI applications [02:26](https://www.youtube.com/watch?v=WYBX6fqQyFo&t=146s) Beinat says organizations can either mandate AI education or give everyone tools to experiment from the bottom up. He favors the second approach, with guardrails that prevent catastrophic mistakes. Early experiments with GPT-3 produced uses that the central team would not have invented, including personalized exams created from student CVs. The value comes from both the productivity of individual employees and the shared understanding of where AI works, where it fails, and how to make it useful. ### Tool adoption needs freedom alongside practical guidance [05:04](https://www.youtube.com/watch?v=WYBX6fqQyFo&t=304s) The speakers disagree slightly on how much tool choice to offer. Beinat favors testing multiple tools because the technology is changing and different use cases may need different tools. Öztekin describes a more focused rollout with GitHub Copilot, Amazon Q Developer, Cursor, and Gemini so the company can share guidance and build trust. He says Just Eat Takeaway.com reached more than 95% daily use among engineers, although estimated AI-generated code did not produce an equivalent increase in engineering velocity. ### Coding is only one part of the productivity problem [09:39](https://www.youtube.com/watch?v=WYBX6fqQyFo&t=579s) The speakers say coding is not usually the main bottleneck. Alignment across teams, product managers, UX, and QA can take longer than writing code. Beinat also says software development is easier to augment because it has established training, tools, and ways to judge quality. Other departments contain thousands of small, specialized opportunities where one person or a small group can use an agent to remove repeated work. These small opportunities are hard for a central product team to address. ### A large agent target is mainly a change-management device [15:32](https://www.youtube.com/watch?v=WYBX6fqQyFo&t=932s) Beinat explains that the goal of 30,000 agents is intended to change how employees think about AI. Prosus had seen adoption plateau among technical users because other employees assumed agents were engineering tools. Sitting with marketing or HR and creating an automation in 15 or 30 minutes helped remove that barrier. Once employees create something useful, they can share it with colleagues. The target creates attention and participation before the company evaluates which agents are most valuable. ### Useful AI adoption must begin with immediate personal value [23:11](https://www.youtube.com/watch?v=WYBX6fqQyFo&t=1391s) Beinat rejects the idea that employees should spend time learning AI now because it might help them later. He says people will use these tools only when they are immediately useful. The organization can lower the barrier with working tools, access to internal knowledge, safeguards, and examples from other employees. Prosus also uses recognition and competition to spread good agents, including a program that features selected agents and builds toward a larger event. Experimentation is acceptable when it helps someone learn what the tool can do. ### Governance should control risks inside the workflow [26:04](https://www.youtube.com/watch?v=WYBX6fqQyFo&t=1564s) Öztekin says governance should protect people without making work painful. It should vary by area, since software engineering, procurement, sales, and customer service have different risks. Beinat says agents should never have more permissions than the person using them, and he is uncomfortable with fully autonomous agents. Human checks can be added before actions such as visiting a site or executing a test, especially when prompt injection may be involved. Controls should define unacceptable outcomes while leaving room for testing. ### Shadow AI is an old procurement problem with new risks [35:29](https://www.youtube.com/watch?v=WYBX6fqQyFo&t=2129s) Öztekin says departments have always bought tools independently, even before AI. The AI wave made the problem more visible because departments can buy overlapping tools under separate contracts and send company data to external services. His technology team uses policies, finance reports, network logs, and security integrations to understand what is being used. The aim is to help employees operate safely inside a workable framework instead of blocking every purchase. He admits the company has not solved the problem completely. ### AI costs can rise even while model prices fall [41:59](https://www.youtube.com/watch?v=WYBX6fqQyFo&t=2519s) Beinat says the cost per token is falling, but use cases are consuming more tokens as they become more complex. This can make total spending rise. Teams need to optimize prompts and context, and open-source models may be competitive for practical tasks. In some cases, hosting infrastructure internally may make sense. Mert says the larger financial question is whether AI helps the company grow, while both speakers agree that AI usage can become a significant operating cost if nobody watches it. ## Notable quotes - "You are not just providing a functionality or feature to the business anymore. You also need to help the business to go through this big transformation." (01:00) - "Collective discovery on one hand but also collective awareness on the other hand." (04:03) - "30,000 agents in fact means it's a change management process." (18:08) - "There's no way that you can push these tools down the throat of people." (23:31) - "At this point in time I think we are sort of uncomfortable letting agents do their work complete in autonomy." (31:34) ## Tools & references mentioned - Prosus Group - Just Eat Takeaway.com - Toqan - PlusOne - ChatGPT - GPT-3 - GitHub Copilot - Amazon Q Developer - Cursor - Gemini - Deep Research - MCP - OpenAI - Google - Okta - Zscaler ## Who should watch - You run technology for a company where departments are adopting AI tools faster than central IT can review them. - You are planning an internal agent platform and need to balance employee freedom with permissions, human checks, and data controls. - You are measuring AI-assisted engineering work and want to understand why more generated code may not produce the same increase in delivery speed. ## Related talks - [Enterprise AI Operations: The Missing Piece](https://mlopstalks.com/talks/enterprise-ai-operations-the-missing-piece) (Rani Radhakrishnan, PwC US, 41:28) - [Architecting Modern AI Systems](https://mlopstalks.com/talks/architecting-modern-ai-systems) (, 56:58) - [The Centralization of Power in AI](https://mlopstalks.com/talks/the-centralization-of-power-in-ai) (Kyle Harrison, Contrary, 1:01:35) - [Responsible Autonomy: Building Governance Frameworks for AI That Act in the Real World via MCP](https://mlopstalks.com/talks/responsible-autonomy-building-governance-frameworks-for-ai-that-act-in-the-real) (Saurabh Mishra, Optum, 27:55) - [The Cost of AI: FinOps Strategies for Intelligent Agents](https://mlopstalks.com/talks/the-cost-of-ai-finops-strategies-for-intelligent-agents) (Advait Patel, Broadcom, 31:05)