Skip to content

Toolkit — Cynefin framework

Gate: G3 Route (Q3 routing deliberation). Category: routing substrate.

What problem it solves

Different problems demand different decision substrates. A problem where cause and effect are obvious can be solved with categorisation and best practice; a problem where cause and effect are knowable only retrospectively needs experimentation. Treating every leaf the same — as a known problem or a research problem — misroutes half the engagement. Cynefin, Snowden's sense-making framework [1], places each leaf into one of five domains (clear, complicated, complex, chaotic, confused) so that the routing substrate matches the problem's actual epistemology.

How it is used

A 30-minute G3 conversation per leaf or per problem cluster. The chair asks three questions: is the cause-effect relationship knowable in advance? is it knowable only retrospectively? is it knowable at all? The answers route the leaf to one of the five domains. The routing recommendation follows from the domain: clear → rules; complicated → expert analysis (ML fits here); complex → empirical experimentation (AI with controls, shadow stages); chaotic → act-first, sense-later (not an AI domain); confused → decompose until individual sub-problems can be placed.

Inputs

  • The decomposition leaf or cluster.
  • The issue-tree / pyramid framing (so the leaf is well-defined).
  • Domain practitioners to judge whether cause-effect is knowable.

Outputs

  • A Cynefin domain tag per leaf: clear / complicated / complex / chaotic / confused.
  • A routing implication per domain, linked to the AI capability levels and the other G3 tools.
  • Flagged confused leaves for further decomposition (don't route the leaf as-is).

Visualisation

Cynefin framework — four quadrants plus a central confused region; clear, complicated, complex, chaotic, and confused in the middle Complicated cause-effect knowable with analysis sense → analyse → respond good practice; expert + ML fits Complex cause-effect only knowable afterwards probe → sense → respond emergent; empirical reads required Clear cause-effect obvious; known answers sense → categorise → respond best practice; rules / DMN fits Chaotic no knowable relationship; crisis mode act → sense → respond not an AI domain Confused domain unknown; decompose further

Four domains around a confused centre. Clockwise from top-left: Complicated, Complex, Chaotic, Clear. The confused centre is where the framing itself is unresolved.

Anatomy

Clear. Cause and effect are obvious; the problem has known answers. Routing: rules, decision tables, DMN. A leaf in this domain does not need AI.

Complicated. Cause and effect are knowable — but require analysis, sometimes by experts. Routing: classical ML fits naturally; the prediction is learnable from history because history contains a stable mapping.

Complex. Cause and effect are knowable only in retrospect. Patterns emerge; they do not hold reliably. Routing: empirical reads (shadow → canary → progressive) are required because offline evaluation cannot predict live behaviour. AI is useful only with substantial controls.

Chaotic. No knowable relationship; the system is in crisis. Routing: act-first; AI is not the substrate. The engagement's role is to get out of this domain, not to model within it.

Confused. The domain itself is unclear — practitioners disagree on which of the four above applies. Routing implication: decompose further; each sub-leaf may land in a different domain.

Example

Paper trail — Cynefin tagging for the freight-yard leaves

G3 Q3 Cynefin session, 45 minutes covering five leaves. Chair: Ada. Team: Priya, Raj, Alex.

Leaf 1 — slot assignment under normal conditions. Ada: "is cause-effect obvious, analysable, only-retrospective, or unknown?" Raj: "we have 18 months of data. Assignments produce unload outcomes. The relationship between assignment and outcome is analysable." Complicated. Routing: ML fits. Alex: "a regression model on historical cycle-time-at-slot is the right shape."

Leaf 2 — priority arbitration. Ada: "same question." Chen (dispatcher, joining): "the rules exist in policy, but practitioners encode extra rules (like don't interrupt active docks) that the policy doesn't cover." The decision-tables work done earlier surfaced this. Clear, once the hidden conditions are extracted. Routing: DMN.

Leaf 3 — operator-override rate at new yard. Priya: "we're opening a new yard. We have no historical override data for it. Predictions don't hold — we'd need to observe live." Alex: "classical ML trained on other yards would transfer poorly; override rate is yard-specific." Complex. Routing: the allocator at the new yard starts in shadow, and the override-rate leaf waits for live data. No AI to ship at launch.

Leaf 4 — safety event response (hazardous-material spill in yard). Ada: "is this a predictable-and-analysable event, or a crisis?" Raj: "crisis. Rare, immediate, every one is different." Chaotic. Routing: not an AI domain; the engagement's role is to make sure hazardous-material events escalate out of the allocator's scope.

Leaf 5 — weather-adjusted unloading. Team disagrees. Alex: "the effect of weather on unloading is knowable in principle, but we have few events." Raj: "it's a known effect; we just don't have data." Priya: "it's not predictable in the next quarter; we haven't measured." Confused. Decompose further: rare-event operations adjustment (arguably complex, needs more data) vs. known-weather categories (arguably complicated, fits rules). Split the leaf; re-tag the pieces individually.

Close. Five leaves tagged; three clean routings (ML, DMN, drop-from-scope), one complex leaf (empirical read required; no AI at launch), one confused leaf (decomposed further). Ada: "half the value here is the 'confused' tag. That leaf was going to be routed to ML; it shouldn't be until the decomposition is redone."

Pitfalls

Everything is complicated. The default answer for engineering teams. Ada's job is to push — is the historical data representative of the future? If not, the domain is complex, not complicated.

Complex treated as complicated. The most expensive mistake. Complex leaves get classical ML, which trains on historical patterns that don't hold. The production model is wrong; the team blames the data.

Chaos as a domain to model. Crisis conditions are not learnable. Building an AI for a crisis is building a system that won't run when it's needed and will be a liability when it's not.

Confused as procrastination. Tagging a leaf confused is valuable only if followed by decomposition. If it stays confused, the tag has done no work.

Cynefin as a category system. The framework is a sense-making aid, not a taxonomy. Good use allows domains to shift as learning happens (a complex leaf becomes complicated once data accumulates).

When not to use

  • Leaves whose routing is already decided by an earlier gate (Tier 0 refusal, DMN-covered rules).
  • Engagements too small to warrant the conversation — one-leaf engagements don't need the domain tagging.

Provenance

Cynefin was introduced by Snowden and Boone in their Harvard Business Review article [1] and expanded in Snowden's Cynefin compendium [2]. The five-domain structure has variations (the terms simple and clear are both used); the underlying epistemological distinctions are what matter, not the labels.

  • AI canvas. Applied to complicated and complex leaves; Cynefin tag is the canvas's first parameter.
  • Pre-mortem. Run on complex leaves to surface the modes of empirical read needed.
  • Shadow deployment (G3a). The primary mode for complex leaves.

Verification

[1] Snowden DJ, Boone ME. A leader's framework for decision making. Harvard Business Review. 2007;85(11):68–76. [verified] The canonical treatment.

[2] Snowden DJ, Greenberg R, editors. Cynefin: Weaving Sense-Making into the Fabric of Our World. Cognitive Edge; 2020. [verified] Extended treatment and later developments.