conways-law-ai-architecture-check
Diagnoses whether an organization's communication structure will get mirrored in whatever AI system it builds — fragmented teams produce fragmented, siloed AI tools with no shared capability — and app
SKILL.mdai-strategy-and-governance/skills/conways-law-ai-architecture-check/SKILL.md
GitHubConway's Law AI Architecture Check
Purpose
Melvin Conway's 1967 observation — organizations design systems that mirror their own communication structure — applies to AI systems with unusual force, because AI value depends on data and capability flowing freely across whatever boundaries already exist. A company that builds AI on top of its existing silos, without addressing this first, doesn't get an AI-native business; it gets several disconnected AI tools, one per silo, that reflect and often amplify the fragmentation that was already there. Use this check early in any significant AI initiative — before the technical architecture is designed, not after.
Anchored in research
Conway's Law itself (Melvin Conway, 1967) is a well-established, uncontested base concept. Its extension to AI specifically is an active, independently-arrived-at 2026 discourse theme across multiple unrelated sources making the same point: organizations tend to design systems that mirror their own communication structures, and AI systems make this reflection louder, faster, and harder to ignore. A commonly cited supporting data point: AI projects fail at roughly twice the rate of non-AI technology projects (RAND Corporation research), with miscommunication about a project's intent and purpose cited as the most frequent cause — directly consistent with a structural, not technical, root cause.
Method
- Map the current communication structure, not the org chart — who actually talks to whom regularly, who shares data and tooling decisions, and where the real boundaries sit (these often don't match the formal reporting lines).
- Predict the AI architecture this structure will produce if nothing changes. The common failure pattern to check for explicitly: each team buys or builds its own AI tool independently — marketing runs one assistant, sales another, operations a third — each reflecting the data, language, and assumptions of the team that built it, with no shared capability visible anywhere in the organization. If the current initiative looks like this pattern already forming, name it now, before more is built on top of it.
- Check whether the operating model is actually ready for the INTENDED AI architecture, not just whether the technology is ready. If the intended architecture requires data or capability to flow across a boundary that today's communication structure doesn't cross, the AI system will inherit that gap and reproduce it at machine speed — this is the core mechanism, not just an analogy.
- Apply the resequencing rule directly: decide on structure before building, not after. Reorganizing communication structure, decision rights, or team boundaries is comparatively easy before an AI architecture has been built around the current structure — and very hard afterward, because by then the technical architecture has calcified around exactly the boundaries that needed to change. If a structural fix is needed and is being deferred "until after the AI project," name this explicitly as a decision with a cost, not a neutral sequencing choice.
- Use
../../../business-design-frameworks/skills/layer-based-business-structuring/SKILL.mdas the complementary structuring tool once this check identifies a genuine need to redesign boundaries — that skill designs the modular layer structure and build/partner interfaces; this skill's job is narrower: diagnosing that the CURRENT communication structure will distort the AI architecture, and forcing the sequencing decision about when to fix it. - Revisit the check at major milestones, not just at the start — an organization's communication structure can shift during a multi-quarter AI initiative (team reorganizations, new hires, a department gaining or losing authority), and a structure that passed this check at kickoff can silently fail it by the time the system ships.
What this skill does NOT do
- Doesn't redesign the organization — it diagnoses whether the current
structure will distort the AI architecture and names the sequencing
decision; the actual reorganization is a leadership and change-
management decision (see
../../../change-and-communication/README.mdfor the communication side of executing a structural change). - Doesn't replace
layer-based-business-structuringfor designing the target modular structure — it identifies that a structural problem exists and needs resequencing, that skill designs the actual layers. - Doesn't guarantee restructuring first will prevent every AI project failure — RAND's finding is that miscommunication is the MOST frequent cause, not the only one; this check addresses a common, specific failure mode, not every failure mode.
Refinement notes
- What's a real case where this check caught a silo problem before an AI project calcified around it — and what happened when a client chose to defer the structural fix anyway?
- How do you get a client's leadership to take on a reorganization BEFORE the AI project, when the natural instinct is always to build first and fix structure later?
- Is there a lighter-weight version of this check for a smaller initiative, where a full communication-structure mapping (Step 1) is disproportionate?
Continue from here
- Use early: before the technical architecture of a significant AI initiative is designed.
- Related:
../../../business-design-frameworks/skills/layer-based-business-structuring/SKILL.md— designs the target modular structure once a redesign need is identified here. - Related:
../ai-reshuffle-opportunity-framing/SKILL.md— a 3rd-order reshuffle opportunity often requires exactly the kind of structural change this check is designed to surface. - This pack's shared guardrails:
../../CLAUDE.md
References
../../references/ai-native-reshuffle-heuristics-research.md— selection and grounding notes for this skill and its siblings../../references/— the pack's shared background material../../CLAUDE.md— the pack's shared guardrails