In 1975, American paediatrician and system theorist John Gall formulated a simple observation that has outlasted most computer architectures from that era: A complex system that works is invariably found to have evolved from a simple system that worked. A complex system designed from scratch never works and cannot be patched up to make it work. You have to start over with a working simple system. Gall originally published the principle in his 1975 book General Systemantics (specifically as Systemantic #15), but the explicit phrase ‘Gall’s Law‘ was coined later by software engineers and systems theorists, most notably popularized by Josh Kaufman in The Personal MBA.
Fifty years later, it explains why most public-sector AI initiatives stall in the pilot phase. Driven by political pressure to demonstrate digital leadership, public institutions, international organisations, and government agencies frequently attempt to deploy complex, multi-functional AI architectures from day one. When these grand systems fail (hallucinating during policy synthesis, misinterpreting administrative rules, or generating contradictory legal outputs), the instinct is to blame the underlying technology. But the fault rarely lies with the language models or the code. It lies in violating a fundamental rule of systems engineering: skipping the step of building a working simple system first. For public sector leaders and AI governance practitioners, Gall’s Law offers a clear diagnostic tool to identify why AI projects collapse and a practical framework for building tools that survive contact with bureaucracy. When public sector entities bypass incremental system design, their AI deployments almost always succumb to one of three structural traps. Government agencies and multilateral institutions often face structural pressure to pursue large-scale procurement. Tenders often call for comprehensive, department-wide ‘AI Knowledge Platforms‘ designed to handle constituent communication, internal policy retrieval, legal analysis, and decision support simultaneously. By attempting to solve every operational challenge at once, these projects introduce too many variables. A failure in data ingestion breaks the search function; an error in output formatting degrades user trust; inaccurate legal citations draw public criticism. The Governance alternative: Mandate modular, single-purpose deployments. Before building an autonomous policy engine, build a reliable semantic search tool for your institution’s archive. Make sure it works, measure its error rate, and build user trust. Once that simple system is stable, layer additional capabilities on top of it. A frequent operational mistake is attempting to use AI to fix broken administrative processes. If a regulatory clearance pathway relies on unwritten rules, fragmented legacy databases, or weekly procedural shifts, introducing an AI model will not create order. It will simply magnify the chaos across the enterprise. Generative models require clarity, consistent data structures, and predictable rules to produce accurate outputs. When fed ambiguous inputs, they produce plausible-sounding fabrications. The Governance alternative: Mapping of the process must precede algorithmic deployment. If a workflow cannot be clearly documented on paper or executed consistently by a human staff member, it isn’t ready for automation. AI deployment should be the final step of institutional simplification, not the first. Public trust in government technology is fragile and hard to rebuild once lost. Organisations eager to demonstrate efficiency often leap directly from manual operations to automated outputs, bypassing human oversight in the pursuit of speed. When an autonomous system fails in the public sector, whether by issuing incorrect benefits advice or generating inaccurate policy summaries, the damage to public trust is disproportionate. The immediate reaction is often a total freeze on AI adoption across the entire organisation. The Governance alternative: Design for progressive capability. Systems should evolve through three distinct phases. They begin with assistance, where AI drafts text or retrieves information while humans evaluate every output. They advance to augmentation, where the system handles routine end-to-end tasks while staff step in to resolve edge cases and high-stakes decisions. Institutions should consider autonomy only after empirical reliability is proven, reserving it strictly for low-risk, deterministic tasks with high fault tolerance.
Public sector AI failures carry high stakes globally, while diplomatic posts, embassies, and international secretariats operate under even tighter resource and risk constraints. In these high-constraint policy environments, Gall’s Law functions as an essential filter for project viability. Small diplomatic missions do not need multi-million-dollar institutional AI models to benefit from machine learning. The most effective applications are small, focused, and low-risk. Missions can focus on targeted media summarisation by converting foreign-language press monitoring into structured daily briefs, deploy tools for textual comparison to track minute adjustments across consecutive treaty drafts, or improve archive accessibility by indexing historical diplomatic cables for internal semantic search. Each of these represents a simple system that works. By proving value and safety on routine tasks, missions can build the internal digital literacy, data hygiene, and governance frameworks needed to evaluate more ambitious tools down the line. Before pursuing a new AI deployment, public sector leaders and project directors should evaluate the proposal against four systemic criteria: If a proposed project answers ‘no‘ to any of these questions, it is attempting to build a complex system from scratch. Re-scoping the project back to its simplest working component will save both public funds and institutional credibility. In current AI strategy debates, starting small is often mischaracterised as a lack of vision. Public sector leaders face immense pressure to announce groundbreaking digital initiatives that match technology vendors’ marketing promises. Gall’s Law reminds us that real institutional transformation is cumulative. The most ambitious AI governance strategy isn’t the one that promises to automate an entire ministry overnight; it is the one that deploys a single, reliable tool, establishes robust human oversight, and builds a working foundation for the future. If your institution’s AI initiative is struggling under its own weight, the solution is rarely more data or a larger model. The solution is usually to ask a simpler question: What is the smallest version of this system that actually works? Author: Slobodan Kovrlija
The three failure points of public AI
1. Monolithic procurement vs. modular evolution
2. Automating institutional chaos
3. Premature autonomy and trust destruction
Translating Gall’s Law to diplomatic missions and secretariats
The Gall’s Law assessment: Four questions for public sector AI pilots
Reframing ambition in AI governance
