BIG BANG KI Festival

One Sentence Takeaway

Clock Drift, the gap between AI’s execution speed and your company’s decision speed, is an operating-model problem that tighter control can’t fix; the fix is orchestration: track capabilities, evaluate outcomes, reallocate responsibility, retune the interaction patterns.

Summary

On the Jupiter Stage at the BIG BANG KI Festival in Berlin, Neil Redding names what broke: not your strategy, your speed. Work still moves through specs, reviews, and committees while AI-enabled work on every laptop runs at a different tempo, and he gives the gap a name: Clock Drift, when AI can execute faster than you and your company can make decisions. The evidence runs from Samsung banning ChatGPT about 20 days after approving it, to two non-technical CNBC reporters vibe-coding a working replacement for their project management tool in about an hour for about $15 of compute, to Anthropic’s associate general counsel building a legal review plugin in an afternoon with plain language. The sanctioned version shows what’s possible; the unsanctioned version, with about half of employees using unapproved AI tools, is shadow execution. The underlying shift is from tool to participant, and the one-word difference is reasoning: tools do tasks, participants co-create outcomes, and AI rises to the level of the human interacting with it, so the ceiling is the quality of our intent, context, and framing. Interactions compound into dynamics, which Neil shows by building his company’s context layer live with Claude Code, including the moment it pushes back on his design and proposes a better one. Proprietary context is your only moat as models commodify everything else, and you do not have to own the model to own the decision system. For an audience that knows how to engineer reliable systems, the next system to engineer is the relationship between human judgment and machine-speed execution. Humans are shifting from mostly doers to mostly deciders, the entry practice is delegate first, and the leadership model is orchestration, demonstrated live on Neil’s orchestration dashboard: track, evaluate, reallocate, retune, turning Clock Drift into clock drive.

Key Ideas

Clock Drift is an operating-model problem

Your best people can build a working tool in an afternoon, then wait weeks for the idea to be approved. Neil defines Clock Drift as AI executing faster than the company can decide, and stresses it comes from organizations failing to move with AI, not from AI moving too fast. Samsung’s ChatGPT ban after three leaks, the CNBC reporters’ $15 tool, and the roughly 30% software-sector drop that followed show what the gap can do, and tighter control is not the answer.

Shadow execution is already the practical reality

Mark Pike, associate general counsel at Anthropic and a lawyer with no technical background, built a legal review plugin with Claude in an afternoon: marketing copy review, privacy impact assessments, contract redlining, turnaround cut from days to hours, all from text instructions and markdown files. His version is sanctioned. Recent studies put about half of employees in most companies on unapproved AI tools. If your organization can’t decide at the speed of execution, someone will execute your strategy before you approve it, and if you’re unlucky it’s a competitor.

Tools do tasks; participants co-create outcomes

Last year, learning to prompt was enough to stay competitive. Reasoning, one of 2025’s biggest capability unlocks, changed the relationship: the moment AI reasons rather than just responding to commands, it stops being a tool and starts being a participant. A participant prompt brings the goal, the constraints, and what can’t break, and returns judgment. AI rises to our level: Anthropic’s economic index found Claude responds at the sophistication of the human prompting it, so the ceiling is the quality of our intent, context, and framing.

Interactions compound into dynamics

The nature of the interactions determines the nature of the outcomes. A traditional decision loop of briefs, reports, and committee reviews misses signals over weeks; an evolved pattern of briefed intent and continuously surfaced live options converges faster with less rework. Same people, same strategy, same AI; changed interactions, different outcomes. As these patterns compound, systems go from discrete to cumulative, prompted to reciprocal, planned to emergent, with Matt Schumer’s viral article and GPT-5.3 Codex helping create itself as the markers.

Context is the moat, and you don’t have to own the model

What drove this year’s jump in sustained autonomous work, visible in METR’s data, is mostly context: models hold vastly more of it and reason over it longer with fewer errors. What differentiates any business is its proprietary context: customer relationships, internal decisions, deep knowledge of offerings and needs. That context is the only defense as models commodify everything else, and amid the debate about who owns the models, owning the decision system matters more. Neil builds his context layer live with Claude Code, then reframes context engineering as relationship engineering: context coherence, trust and delegation, guardrails and governance, optimal human review.

Doers to deciders, and orchestration as the leadership model

Humans are shifting from mostly doers to mostly deciders. A Deloitte report finds 84% of companies haven’t changed a single role in response to new AI capabilities; the room, Neil argues, is the sixteen percent. The entry practice is delegate first, the way you’d learn any new colleague’s capabilities. The leadership model is orchestration, shown live on Neil’s dashboard as a model upgrade triggers the full cycle: track the capability change, evaluate it, reallocate two tasks from a human analyst to the upgraded agent, retune the autonomy boundaries and the analyst’s oversight role.

Notable moments:

• The host closes Neil’s introduction by revealing it’s his birthday, with an “Alles Gute zum Geburtstag” from the stage; Neil, surprised, promises to do his best for the room.

• Claude, working inside PowerPoint, reviews Neil’s own slide for this Berlin audience and flags “existential problem” as the weak part unless it’s set up fully; Neil asks the room to hold him accountable.

• A show of hands on two prompts: the task prompt (“summarize our Q2 sales data”) feels a little old, and more hands go up for the participant prompt that asks for the strongest argument and what would undermine it.

• During the live build, Claude Code answers Neil’s scaling challenge with “Good challenge,” names the core tension, and proposes the aggregated landscape with drill-down layers that becomes round two of the context layer.

In Neil’s words:

• “If your organization can’t decide at the speed of execution, someone else is gonna execute your strategy before you can even approve it.”

• “I would argue you do not have to own the model to own the decision system.”

• “Leadership is orchestration, and orchestration turns clock drift into clock drive.”