DeepMind’s Kareem Ayoub: You can’t fully govern AI, but you can fence it like banks did

# What Ayoub said and why it matters Kareem Ayoub, Google DeepMind's vice president of AI technical strategy, told Bloomberg's Tom Mackenzie at the HumanX conference in Amsterdam that companies cannot fully govern AI systems themselves. Instead, they can build a deterministic governance perimeter around those systems—a fixed "fence" of rules and checks. He drew a direct analogy to how banks in the late 1980s and early 1990s constrained trading algorithms inside fixed, auditable rules.

# The governance problem broken into three parts

# Practical framework examples mentioned Google's Frontier Safety Framework was cited as part of the solution set. It sets capability thresholds and mitigation steps companies can apply to frontier models. Ayoub contrasted industry views: some call for pacing development, others propose industry-led regulation modeled on financial oversight bodies, and Ayoub argued that guardrails will require partnerships beyond tech companies.

# From models to agents: a product shift

# Four operational habits of successful companies According to Ayoub, the companies getting value safely share four habits:

• Build safe sandboxes quickly rather than waiting for long IT projects.
• Don't bet against continued model progress—design for it.
• Measure the outcomes that matter, not vanity metrics.
• Senior leaders use the tools themselves and set the tone.

He warned that organisations fail when staff are rewarded for hiding AI use or when leaders adopt tools only superficially. As the cost of compute falls, he urged more spending on checking outputs rather than producing more outputs.

# DeepMind's internal context and wider signals When asked about morale and recent departures, Ayoub offered limited commentary but noted public moves: Demis Hassabis has taken the chairman role, Koray Kavukcuoglu took on AI leadership in California, and the Gemini programme is a focal point. DeepMind launched a DeepMind Institute to explore AGI implications. Former DeepMind researchers are founding startups, including a world-model company named Emulate. Ayoub summarized the shift plainly: the future is less a forecasting challenge and more a design problem.

# What this implies for companies The takeaway for non-research organisations is practical: accept that models will keep improving, design deterministic perimeters you can audit, deploy sandboxes early, measure outcomes, and insist leaders be fluent users of the tools. Governance is not a single monolithic control but a set of enforceable boundaries, operational practices, and external partnerships.