Reducing risk in projects, increasing resilience to unexpected change - Dave Snowden (Talk Outline)

Dave Snowden (Cynefin creator, from Wales) on applying complexity science — emphatically not systems thinking — to projects. His golden nugget for developers: “entangle yourself with your users.” Core premise: a complex system is one where you can’t know in advance what will happen; the only certainty is that whatever you do will have unintended consequences, so you must manage differently. The talk sets up theory (emergence), then walks through Estuarine mapping, distributed decision-making, a new portfolio risk matrix, micro-narrative sensing, and framework recombination.

1. Framing — Complexity Science ≠ Systems Thinking

• The “clue is in the second word” — complexity science and systems thinking are not the same thing.
• What you can manage in a complex system is quite limited (comes back to emergence).
• Roadmap: theory → affordance mapping (the main topic) → distributed decision-making → a new risk matrix → weak-signal detection → breaking frameworks → resilience.
• Affordance mapping (from ecology, Gibson): “no point having a strategic plan to be the world’s best cactus grower if you live in a paddy field.” What does your environment even make possible? Also grounded in constructor theory and quantum mechanics — “the science of can and can’t.”
• Cynical truth from his project-management career: no (male) project manager will report his project is going wrong until he’s certain he’s covered up his own guilt and isn’t responsible (gender used deliberately).

2. Emergence & Entanglement

2.1 Entangled systems and “joining the dots”

• A complex adaptive system = an entangled system — everything entangled with everything else (a mangrove swamp / thick forest); you can’t discern connections until you act and “nasty things happen.”
• “Why didn’t we join up the dots?” is always said in hindsight (e.g., 9/11: someone training to take off but not land was an obvious signal only in retrospect).

2.2 The combinatorial explosion

• 4 dots → 6 connections → 64 possible combinations.
• 10 dots → ~3.4 trillion; 12 dots → ~4.8 quadrillion (boards guess ~150; programmers guess “a few million”).
• The emperor’s chessboard parable (double the rice per square — not enough rice in China). “How many dots in a human system?”
• Hindsight does not lead to foresight — the flaw of root-cause analysis and lessons learned: they use hindsight, not present reality.

2.3 What emergence actually means

• Reductionism done wrong = assuming you can determine the whole from the properties of the parts. As parts interact, new unpredictable patterns form.
• Physics example: theoretical physics said superconductors couldn’t exist (not deducible from parts) — but smashing atoms together produced them.
• Management implication: trigger emergence and monitor for patterns — give more energy to good patterns, less to bad. The more you commit to an outcome, the higher your risk (option theory: hold things back, don’t commit).

3. Preconditions for Emergence (the manageable few things)

3.1 A significant number of actants

• Actant (from actor-network theory): not only humans have agency. “All actors are actants, but not all actants are actors.”
• Three types:
  • Actors — obvious agents.
  • Constraints — either contain (create predictability) or connect (less predictability). Regulations are constraints.
  • Constructors — from theoretical physics: something that transforms things but doesn’t itself change in the act → predictable. Software analogue: object orientation (consistent output for given input). Processes are constructors; human rituals are important constructors.
• Ritual example: changing clothes changes thinking (cognitive activation patterns; relevant to school uniforms). Experiment: coders who change into formal clothes before testing find more errors.

3.2 Rich local interaction

• A termite nest’s sophistication arises from termites responding to pheromone traces — no directing intelligence.
• Nature optimizes without direction, purpose, or goal — via multiple micro-level interactions, seeing what works.

3.3 No agent thinking holistically

• A great error of systems thinking: wanting everyone to think holistically — you can’t (too many dots), and pretending to perverts the system toward whatever has the most power.
• Consequence: you need active monitors to see patterns emerging — e.g., replacing project reports with continuous micro-activity reporting by engineers (a human sensor network); presenting a problem to the whole workforce and iterating responses to consensus in minutes, instead of commissioning conventional research.

3.4 Scaffolding

• Make a constraint/constructor/actor so rigid it can’t change → predictability.
• Example: drive on the left (UK) vs. right (Germany) — you don’t want emergence every morning.
• South of Naples / Amalfi coast: informal swarming behavior — follow the next car, match speed, avoid collision, ignore the rules (like birds flocking). This is heuristic-based control. Anecdote: he caused a pileup by braking on a red light — the informal Naples rule is to speed up through it.
• Complexity vs. systems thinking on direction: complexity says understand where you are and start journeys with a sense of direction; systems thinking says know the end state and close the gap — dangerous under high uncertainty. Too many treat North Stars as destinations rather than navigation aids. (Seneca quote: we can know the present.)

4. Method 1 — Estuarine (Affordance) Mapping

• Low-cost, two half-days: afternoon to build the map, next-day half-day for actions after reflection → “massive risk reduction.”
• Step 1: brainstorm everything operating in the system as actors, constraints, or constructors; place on a map. Vital rule: if you can’t agree what something is, break it down until you agree — no discussions/arguments allowed (used in peace & conflict resolution too). Goal: as objective an assessment as possible, free of power/authority.
• Step 2: map onto a grid of energy cost of change × time to change (energy = proxy for resource/money/attention; time currently linear — two other time types in R&D for next year, since activity and knowledge compress time). Again, disagreement → break it down.
• Three lines drawn by the group:
  • Counterfactual line — “counter the facts”: items whose energy/time to change is so high they realistically won’t change. (Run in parallel with systems-thinking vision statements, the vision usually lands in the counterfactual domain — over-idealistic; easier for consultants to be aspirational.)
  • Liminal domain (added with Novartis clinical-trials work): things you can’t change but a higher authority could — an alert mechanism. A CEO once said “nobody’s asked me” — the team had just assumed.
  • Volatile domain: things that can “turn on a dime” — high risk (little energy, potentially high impact); overlay impact here.
• Day two — nudge the environment: decide “I’d like that to never change / cost less / etc.” → gradually shift the dispositional state so the project is more likely to succeed. A pre-process before a project to reduce risk — don’t just assume success.
• Discovered by accident: senior executives use it to map their own coaching environment (actors/constraints/constructors among their workforce and colleagues).
• Dark actants: phenomena you observe but can’t attribute to a known actant (like dark matter — visible effect, unknown cause) — a significant risk factor. All methods documented on an open-source wiki.

5. Method 2 — Distributed Decision-Making

5.1 Humans are bad individual decision-makers

• Malcolm Gladwell’s Blink is “a truly appalling book” — cherry-picks supporting examples; the cognitive-neuroscience community wrote The Invisible Gorilla to counter it.
• The invisible-gorilla / radiology experiment: a gorilla image 48× the size of a cancer nodule on the final X-ray → 83% of radiologists don’t see it even while scanning it; the 17% who do come to believe they were wrong after talking to the 83%. “You do not see what you do not expect to see” (only fully autistic people are exempt — and they can’t operate). Critical for projects: unexpected events are very unlikely to be seen.

5.2 Natural group sizes (theoretical biology)

• Loosely maps to Dunbar’s number but better-researched:
  • Sexual pairs — primary caregiving.
  • Work group / extended family — ~5, never more than 7 active decision-makers. We evolved to compromise in groups of ≤7; larger groups fall back into silos and won’t compromise.
  • Demes — hunting parties, ~≤20.
  • Macro-demes — hunting parties uniting in winter to cooperate/survive → where monument-building (Stonehenge) comes from. Give them a big project, don’t just tell them to be nice. (“Market economics only work under conditions of plenty, not starvation.“)
• Example: a room of doctors/nurses/ambulance drivers can’t break out of silos; one person from each role in a small group finds novel solutions (UK hospital triage work).

5.3 The Panopticon effect + resource allocation

• Panopticon effect: if you know you’re observed but not by whom, you’re honest. (Snowden — ex-deputy financial director — “does not trust anybody who tells me to trust them.“)
• Mechanism: identify four roles (via a half/full-day simulation — currently running one on end-of-life decisions in a British hospital on real cases). If you can assemble those roles, document context/decision/consequences at the micro level, accept one anonymous role (identity unknown), you may spend $500/$1,000/$10,000 without permission.
• Then a senior exec running 100 such projects waits to see what works and puts real resources behind what works, not who advocates. Also used in development banks: e.g., village headperson + priest + oldest girl still in school + youngest boy considered a man + an anonymous bank agent agree to spend $500 → auditable, and real money follows those who make things work.

6. Method 3 — New Portfolio Risk Matrix (world-first reveal)

• Two axes:
  • Natural emergence (wait and respond) ↔ Forced emergence (make things happen, decide what to respond to).
  • Fine granularity (many small parallel things) ↔ Coarse granularity (big things, lots of resources).
• Four resulting modes:
  • Traditional strategy (coarse + forced): market research, interview experts, commit resources; fine in some contexts, but has prioritization issues.
  • Parallel safe-to-fail probes (fine + natural, Cynefin-style): generate hypotheses, let any remotely coherent hypothesis run a safe-to-fail experiment, see what works.
  • Opportunistic (coarse + natural): monitor for favorable events, then move resources fast to commit.
  • Many small stimulated experiments (fine + forced): small groups make decisions/run experiments; monitor emergence with micro-stimulation and many small interventions.
• The grid ironically forms a chaos symbol; map each aspect of your project into a domain to choose its mitigation — often start bottom-left and see what works before committing top-right.
• “People are objects too” — architect a people-object with standard I/O that handles ambiguity software objects can’t, and which also generates AI training data. “If you don’t know where the training data sets came from, you shouldn’t be using the AI.” Scathing on AI’s cultural bias (“built by misogynous white males on the US West Coast who take Ayn Rand seriously after puberty”); 25 years of DARPA work on epistemically balanced training data sets. Finely-grained project data → a training set for anticipatory AI alerts bringing in human factors.
• Back-of-a-table-napkin test: if you can’t draw a framework from memory, it has no sensemaking utility (contrast: the SAFe diagram). Sensemaking frameworks must be simple/memorable.

7. Weak-Signal Detection — Walking the Gemba & Micro-Narratives

• First used at the US Army, West Point / Afghanistan deployment (Snowden teaches just war theory at West Point). Concerned about stress, they asked soldiers to keep records as micro-narratives in the field (not patrol reports) → 100% compliance, plus human metadata and interpretation in real time → real-time exception reporting.
• Now applied to large public transport infrastructure and software projects: replace project reports with continuous micro-capture (“human metadata”) → sense patterns early. Avoid structured/periodic reporting — reports written with hindsight get changed for how they’ll be read.
• Multimedia capture: text is <10% of what you know — use pictures, voice, voice tone. Netherlands old-people’s-homes work: a nurse photographs a patient, writes and records their story → a micro-narrative. Open API for integration into project-management/operational systems (seeking partners); creates a real-time sensor network (ask 2,000 employees, get patterns back — “some of them will see a gorilla”; you need non-experts too).
• Whistleblowing problem: nobody reports safety/fraud/misogyny/racism until it’s so serious it’s too late.
  • Anecdote: at an Agile conference, a male speaker slapped a woman speaker on the bum (“Well done, lass, see you in the bar”); she refused to report for fear of secondary abuse and social-media pile-ons.
  • Boeing metal fatigue: engineers felt “something doesn’t feel right” but wouldn’t report yet — exactly the signal you want.
  • Solution: continuous capture tags a “micro-grow report” / microaggression; strip identity, keep keywords + metadata, and report to the board: “You’ve got an emerging problem in this division / a compliance issue on this project.” No need to report individual cases → weak-signal detection.
• (Credit for the “walk the gemba” naming: Nigel — Snowden pointedly “acknowledges my sources.“)

8. Assemblages, Culture Mapping & HeXi

8.1 Culture as a pattern map (assemblage)

• Partnerships with Comic Agile and Gapingvoid: present six cartoons to the whole workforce; each person picks the cartoon representing the culture they think they live in and tells a story about why, indexed onto non-gameable index structures → draw pattern maps (not abstract spider diagrams).
• Each cluster = an attitude to culture; change management becomes “I want more stories like this, fewer stories like those” (everyone understands it) instead of “an Agile mindset” or “transparency.” One cluster = the 17% who saw a gorilla. Theory basis: strange-attractor theory, complexity assemblage theory, Deleuzian epistemology. Stories at the water cooler determine culture more than structured questionnaires.

8.2 HeXi — breaking frameworks

• Take every Agile method/tool/framework (including foresight/strategy; he wants Wardley maps — jab that Simon Wardley “wants other people to do his work for him… calls it democracy; I call it laziness”). Scrum.org just produced theirs.
• Break each method/tool/concept to its lowest coherent unit, then recombine in different combinations (soon software-based; store a “product” as an assembly others can modify). E.g., peel a sprint out of Scrum and replace it with a 3-month time-box from DSDM. Instead of adopting a single framework, assemble the right combination (the HeXi kit).
• Mentions the European Union field guide on managing complexity in a crisis (co-authored) — free copies if your country stayed in the EU (else pay postage).

9. Closing Thesis — From Engineering to Ecology

• Historical arc of management metaphors:
  • Scientific management = military models (through the ’80s) — actually quite adaptive (critics don’t know it well).
  • Systems thinking (’80s, business process re-engineering, Six Sigma — ”Gary Klein calls it that”; “I wish I’d invented Six Sigma”), learning organization = an engineering metaphor (engineers dislike uncertainty and try to eliminate it).
  • Now: an ecological metaphor — manage what physics calls the substrate / biologists call the ecology: manage the environment and see what works, rather than deciding in advance.

10. Q&A

• Constraints vs. actants? The original constraint mapping work now extends to actants — you can have dark constraints, dark constructors (hard), dark actors; use actants as the general tool.
• Micro-sensing? = “Gemba di Flo”: everyone keeping micro-level records = a micro sensor network; an executive can send an infographic to the whole workforce and get results in 5–10 minutes, iterating three times — huge cost/risk reduction. Launching tooling to make this easy.
• Non-tech companies applying this? Yes — website has many; just won an award with AstraZeneca for AI-adoption + risk-reporting work; also two (sensitive) oil companies.
• Early-warning project-failure indicators? Go back over failed and succeeded projects, capture what was reported at the time, have naive agents (blind to the project) interpret it → build a training data set of failure-pattern stories → anticipatory alerts (AI is good at this). Same method built an Al-Qaeda training data set (sponsored a university project in the right region) to troll the web for indicators — focused on the training set, not general intelligence over everything.
• Real-time micro-sensing of software developers? “Agile is a mindset” is “a total disaster” (blames failure on teams lacking the right mindset). Most decisions are made by body and social environment, not the mind; behavior is an emergent property of interaction, causal only once habituated. Example constraint from running a software business (wrote decision-support systems for Guinness PLC): reusable code not under proper version control → warning twice, then fired — a constraint that produced the right behavior (write once, allow reuse). The point: generate behaviors at scale via constraints.

People, Works & References Cited

• Dave Snowden — speaker; creator of Cynefin.
• Gibson (ecology/affordance), actor-network theory (actants), constructor theory, Seneca.
• Malcolm Gladwell — Blink (criticized); The Invisible Gorilla (counter); invisible-gorilla/radiology experiment.
• Dunbar’s number; Gary Klein; Ayn Rand (as cultural-bias critique).
• Organizations/cases: Novartis (clinical trials, liminal domain), AstraZeneca (AI-adoption award), Boeing (metal fatigue), Guinness PLC, US Army / West Point, Netherlands old-people’s-homes, development banks.
• Partners/tools: Comic Agile, Gapingvoid, Scrum.org, DSDM time-box, Wardley maps / Simon Wardley, Nigel (gemba naming), DARPA, European Union field guide.
• Methods: Estuarine/affordance mapping, dark actants, Panopticon effect, distributed decision-making, the portfolio risk matrix, micro-narratives / walking the gemba, HeXi.

Video: https://www.youtube.com/watch?v=PQtFrmYdAz4 — Transcript via yt-transcript.sh; outline generated from the transcript.