Pillar 2: Embrace Complexity¶
v0.8 — Working Draft
This page is under active development. Content is directionally accurate but subject to revision. Suggest an edit →
"You cannot plan your way through emergence, but you can design for it."
The Argument¶
Traditional change management rests on assumptions borrowed from engineering: define the future state, analyze the gap, build the plan, execute the plan. Kotter's 8 steps, ADKAR, Lewin's unfreeze-change-refreeze — all presuppose that organizational change is a complicated problem amenable to decomposition, sequencing, and control.
AI adoption is not complicated. It is complex.
The distinction, formalized in Snowden's Cynefin framework (Snowden & Boone, 2007), is consequential. Complicated systems (building a bridge, implementing an ERP) have knowable cause-and-effect relationships that experts can analyze in advance. Complex systems (ecosystems, markets, organizations adopting emergent technology) exhibit nonlinear dynamics, feedback loops, and emergent properties that cannot be predicted or controlled through advance planning. In complex systems, cause and effect are only coherent in retrospect.
AI adoption is complex for several reinforcing reasons. First, the technology itself is evolving on timescales shorter than organizational planning cycles — a roadmap written in January may be obsolete by April. Second, the interaction between AI tools and human work practices produces emergent behaviors that neither technologists nor change managers can anticipate: teams find unexpected uses, develop informal workarounds, and generate novel failure modes. Third, adoption in one part of the organization creates cascading effects on adjacent functions, producing system-level dynamics that no stakeholder analysis can fully map.
The implication is not that planning is useless, but that the type of planning must change. Complex systems require probe-sense-respond strategies (Snowden, 2007) rather than sense-analyze-respond. Organizations must run small, safe-to-fail experiments, amplify what works, dampen what does not, and iterate. Strategy becomes a portfolio of bets rather than a sequential plan.
This is deeply uncomfortable for leaders trained in predictability. It requires tolerating ambiguity, accepting that the "future state" is unknowable, and shifting investment from upfront design to rapid learning loops. It also requires new sensemaking capabilities — the organizational capacity to detect weak signals, recognize emerging patterns, and adjust course without waiting for quarterly reviews.
Complexity science does not counsel abandoning structure. It counsels designing enabling constraints — boundary conditions that channel emergent behavior without attempting to dictate it. Guardrails, not blueprints.
In Practice¶
A financial services firm abandoned its 18-month AI transformation roadmap after recognizing it was operating in the complex domain. It replaced the roadmap with a "portfolio of probes": 15 small AI experiments across different business units, each with a 6-week cycle, explicit learning objectives, and pre-defined criteria for scaling, pivoting, or stopping. Within two quarters, three probes had scaled to production — none of which had appeared on the original roadmap.
A government agency adopted Cynefin-informed triage for AI initiatives. Before committing resources, each proposed use case was categorized as simple, complicated, or complex. Simple cases (e.g., automating form classification) received standard project management. Complex cases (e.g., AI-assisted policy analysis) received experimental designs with iterative learning cycles and explicit uncertainty budgets. The triage itself became a sensemaking exercise that improved organizational judgment.
A retail organization created cross-functional "sensemaking squads" — small teams whose mandate was not to deliver AI solutions but to detect emerging patterns in how AI tools were actually being used (and misused) across the organization. These squads fed insights to leadership monthly, enabling adaptive course correction rather than plan adherence.
The 4×1 Matrix¶
| Dimension | Example |
|---|---|
| Tools | Cynefin categorization templates, safe-to-fail experiment canvases, signal detection dashboards |
| Processes | Probe-sense-respond cycles, portfolio-of-experiments governance, iterative replanning cadences |
| Behaviors | Tolerating ambiguity, stopping failed experiments without blame, sharing emergent findings rapidly |
| Change Skills | Sensemaking facilitation, complexity triage, designing enabling constraints rather than detailed plans |
Diagnostic Questions¶
- Planning assumptions: Does your AI adoption plan assume a knowable future state? If yes, you are treating a complex problem as merely complicated.
- Experiment portfolio: How many concurrent safe-to-fail AI experiments are running? If the answer is zero or one, you are over-concentrating risk.
- Adaptation cadence: How frequently does your AI strategy materially change in response to new information? If the answer is "annually" or "never," your sensemaking loop is too slow.
- Failure tolerance: When an AI experiment produces unexpected or negative results, what happens? If the response is blame or cancellation rather than learning, your culture is incompatible with complexity.