Committee Function vs. Process
Committees serve a dual nature: the process (how they operate) and the function (what they achieve). The process—twelve-person meetings with two-hour slots, shared documents, hierarchical reviews—was never optimal for architectural decision-making but was the mechanism organizational structure made available. The function—generating perspective-diverse challenges to architectural assumptions, creating social accountability that makes self-examination obligatory, providing redundant error detection through multiple pairs of expert eyes—is irreplaceable. When three engineers from different specializations review a design, they see different things: attack surfaces, performance bottlenecks, user confusion. No single person sees all three because expertise is trained attention, attending to some things at the expense of others. The committee compensates by assembling different trained attentions.
In the AI Story
The Challenger disaster illustrates both what committees provide and how they fail. Morton Thiokol engineers warned about O-ring performance in cold temperatures. The warning traveled through organizational channels, was received, discussed, and ultimately overridden by managers facing schedule pressure. Standard narrative: the committee failed, organizational structure corrupted the safety signal. But the narrative is incomplete—committees don't only corrupt signals, they generate signals no individual could produce alone. The committee provided multiple pairs of eyes reviewing booster design, multiple independent risk assessments, redundancy constituting a form of error detection. The failure wasn't the committee's existence but that its communication structure allowed organizational noise to overwhelm technical signal at the critical moment.
AI does not solve the perspective diversity problem despite confident claims from builders who use Claude to simulate multiple reviews. Claude can generate plausible security reviews, competent performance analyses, common usability findings. These simulations are useful but operate from the same training distribution, sharing statistical blind spots. The genuine security expert brings what no simulation replicates: specific, hard-won intuition from personally investigating real breaches in real systems, years developing a sense for how particular architectures create particular vulnerabilities. That intuition lives in a person, not a probability distribution. The gap between simulated and genuine perspective is the gap between statistical average and individual expertise—narrowing but not yet closed.
Committees provide social accountability—a forcing function AI doesn't replace. When peers review a design, the designer must explain reasoning, justify choices, defend against informed challenge. This social process is generative: explaining often reveals weaknesses invisible from inside one's own perspective. The externalization exposes gaps while the model remained internal. The designer working alone with AI can choose to examine her work or not. The designer within a practitioner community cannot avoid it—community creates expectations, expectations create accountability, accountability creates conditions under which blind spots are discovered before becoming architectural flaws.
Origin
The function-versus-process distinction has deep roots in organizational theory, appearing in various forms across management literature from the 1950s onward. Chester Barnard distinguished between formal organization (process) and informal organization (actual communication patterns). Herbert Simon separated programmed decisions (processable) from non-programmed decisions (requiring judgment). The specific application to committees in the AI era emerged from observing that AI eliminates the process inefficiencies—the slow meetings, the coordination overhead—without eliminating the functional need for diverse expert perspectives on architectural decisions with long-term consequences.
Key Ideas
Process is obsolete, function is irreplaceable. The twelve-person meeting was never optimal for generating perspective collision—it was merely the available mechanism. The collision itself (diverse experts examining the same design) remains architecturally essential.
Expertise is trained attention with blind spots. The mechanism making experts excellent at recognizing patterns in their domain (deep repeated exposure) is the mechanism making them unable to see what falls outside that exposure. Committees compensate by assembling different trained attentions.
Social accountability forces examination. Peer review creates obligation that individual building does not. The designer must explain, justify, defend—a social process revealing weaknesses invisible from inside one's own thinking.
AI simulations ≠ genuine expertise. Claude can generate plausible reviews but operates from the same training distribution. Real experts bring domain-specific intuitions underrepresented in published literature—too recent, too specialized, too tacit to appear in training data.
Function must be preserved in new forms. Targeted architectural reviews (focused conversations with one or two experts at critical decision points) preserve perspective diversity while discarding process overhead—but require designer judgment about when review is needed.
Appears in the Orange Pill Cycle
Further reading
- Diane Vaughan, The Challenger Launch Decision (University of Chicago Press, 1996)
- Karl Weick and Kathleen Sutcliffe, Managing the Unexpected (Jossey-Bass, 2015)
- Amy Edmondson, The Fearless Organization (Wiley, 2018)
- Keith Sawyer, Group Genius (Basic Books, 2007)
- James Surowiecki, The Wisdom of Crowds (Anchor, 2005)