Building Institutional AI Through Workflow Reconception and Infrastructure

Building institutional-grade AI requires moving beyond an automation mindset, which simply digitizes old bottlenecks, to a reconception mindset that redesigns workflows from first principles. Chris Churchman, head of Marquee at Goldman Sachs, argues that the primary danger in the current AI cycle is outsourcing human reasoning, which risks long-term cognitive atrophy. By grounding models in proprietary data and building agentic loops that verify their own work, firms can create durable competitive advantages. The most important, non-obvious implication is that the boring, hard, and thankless work of building reliable, auditable infrastructure provides a moat that flashy, off-the-shelf demos cannot replicate. This analysis helps leaders distinguish between temporary efficiency gains and the structural evolution of their firm decision-making capabilities.

The Demo-to-Product Chasm

Most organizations fall into the trap of evaluating AI based on its best day, which is the flashy, confident demo. Churchman emphasizes that an institutional-grade product is judged on its worst day. When AI models operate in a vacuum, they cannot distinguish between facts and plausible interpolations.

The model itself cannot distinguish between a fact and an interpolation like just it actually can't. Everything goes into the context window and it will go through the same sausage machine comes out the other round.

-- Chris Churchman

The hidden consequence of relying on off-the-shelf models is that the burden of verification shifts entirely to the user. If a system is 90 percent accurate, the user must still verify 100 percent of the facts to catch the 10 percent of hallucinations. This creates a negative-sum game where the AI saves time on generation but consumes more time on auditing.

Why Automation Locks in Legacy

Conventional wisdom suggests that AI should be used to speed up existing processes. Churchman warns that this is a fossil strategy. By simply automating a bottleneck, firms lock in the constraints of the previous era rather than eliminating them.

The reconception mindset requires asking what the firm could achieve if intelligence were abundant and elastic. For example, instead of teaching a user how to operate a complex derivative pricing tool, the objective shifts to building an agent that understands the tool mechanics and interprets its output. This moves the bottleneck from user proficiency to intent expression. The competitive advantage here is delayed: it requires the boring work of mapping proprietary data and entitlements, a barrier to entry that most competitors, who are chasing the flashy demo, are unwilling to clear.

The Danger of Cognitive Atrophy

The most profound systemic risk Churchman identifies is the potential for cognitive atrophy. Historically, technology has offloaded human capabilities like wayfinding, memorization, and now, reasoning.

I think there is a huge danger here that in the era of AI we outsource our reasoning to these models and we have cognitive atrophy that stops us being able to reason from first principles ourselves.

-- Chris Churchman

If firms treat AI as a replacement for human thought rather than an extension of it, they risk losing the tacit, intuitive knowledge that defines expert performance. The system must be designed to empower human reasoning, not replace it. Over time, the firms that retain this apprenticeship culture will possess a workforce capable of reasoning under uncertainty, while others will be left with a workforce that can only prompt a machine.

Key Action Items

• Audit your Automation vs. Reconception ratio: Over the next quarter, inventory your AI projects. Are they just making legacy processes faster? If so, pivot resources toward redesigning the core workflow from first principles.
• Stop betting against the model evolution: Avoid building complex, brittle workarounds for temporary model constraints like small context windows. Focus on building infrastructure for what models cannot learn: your firm specific entitlements, proprietary data connections, and unique mandates.
• Implement Agentic verification loops: Shift from simple prompt engineering to building systems where AI agents check their own work. This pays off in 12 to 18 months by significantly reducing the human audit burden.
• Prioritize Environment Engineering: Invest in the secure, entitled environments where agents operate. This is the hard work that creates a defensible moat against competitors using generic, ungrounded models.
• Protect the Apprenticeship Culture: Proactively design AI tools that require human reasoning to interpret outputs. If a junior employee can complete a task without understanding the why behind the data, you are creating a long-term risk of knowledge loss.