Cynefin for Founders: Matching Your Decision Style to the Mess in Front of You

Sort the problem before you pick the method. Cynefin sorts on one axis: whether cause and effect are obvious to everyone, visible to a specialist, or legible only after the fact. Run one decision style across all three and you will be right by accident and wrong on a schedule.

The expensive version runs one way. You take a problem where nobody can know the answer in advance and apply the method built for problems where somebody can: hire the expert, build the model, spend six weeks, decide once.

The cheap version runs the other way and gets no attention. You spend a quarter testing your way toward something a specialist would have settled in an afternoon.

Why this ensemble, and what I left out

This piece runs the Wire Model. Score the decision’s features, route to an ensemble of formal models, make the ensemble hand back dated actions. Six scores did the routing:

• Emergence and many interacting agents (0.85). Customers, rivals, regulators and your own team move in response to you, so the system you analysed is not the system you act on.
• Deep uncertainty (0.8). In the class that costs most, only the price of finding out is knowable in advance.
• Regime-break risk (0.7). A licence regime, a settlement partner or a currency can relocate an entire industry inside a week.
• Cognitive distortion (0.65). Snowden and Boone call attachment to one method entrained thinking, and place it in both ordered domains.
• Adversarial structure (0.3). Low. No counterparty is choosing moves against you. That score kept every equilibrium model off the card list.
• Historical-analog density (0.4). Low by construction: the complex class is defined by the absence of repeatable precedent.

Three models carry it. Agent-based emergence explains why analysis fails on a whole class. The luck-skill continuum sets how much to believe any single result. Markov state-transition tracks a problem moving between classes while you work on it. They span the complex, random and cycle-regime outcome types, so their errors point in different directions.

Behavioral is folded in rather than shipped as a fourth card. Entrained thinking explains why founders keep reaching for the method that made them competent, which is a diagnosis. It adds no lever, since knowing you are attached to a method still does not tell you which method the problem wants. It appears in the first card and again at the end.

The framework: a problem has a class, and the class picks the method

Snowden and Boone sort issues into five contexts by the relationship between cause and effect. Simple contexts have a clear relationship everyone sees, so leaders sense, categorise and respond against established practice. Complicated contexts have a clear relationship not everyone sees, which calls for experts: sense, analyse, respond. In complex contexts you understand why something happened only in retrospect, so the order inverts to probe, sense, respond. The fifth context, disorder, is where it is unclear which of the four is dominant.¹

Two revisions matter and most founder summaries skip both. Snowden renamed Simple to Clear in 2020, after his co-author Mary Boone argued for it.² The centre stopped being plain disorder and became the aporetic or confused domain, which separates confusion you know you are in from confusion you do not.³

1. Emergence: why the expert’s method fails on a whole class

In the complex class the parts interact, so the behaviour of the whole is not recoverable from analysis of the parts. Snowden and Boone are direct: instructive patterns emerge when a leader runs experiments that are safe to fail, leaders who try to impose order in a complex context will fail, and the complex domain is far more common in business than most leaders realise.¹

Go one step past that. Snowden’s instinct is to probe first and let the pattern arrive. The strongest field evidence says the probe pays only when you wrote down what you expected. A randomised control trial of 116 Italian startups gave both arms ten sessions on getting feedback from the market. The treated arm was also taught to build an explicit framework predicting how their idea would perform, then test the hypotheses rigorously. Treated firms performed better, were likelier to pivot, and were no likelier to drop out early. The authors attribute it to precision: fewer projects with false positive returns pursued, more with false negative returns rescued.⁵

A probe with no written prior is a hope with a budget line attached.

The reference case for African operators is the M-PESA pilot. It ran from 11 October 2005 to 1 May 2006 across eight agent stores and roughly 500 participants in three parts of Nairobi and Thika. Initial scope included a microfinance integration for loan disbursal and repayment, which hit back-office and connectivity obstacles. Field research on the trial users found the killer application was national money transfer, and the launch was rebuilt around it as the send-money-home campaign.⁶ A complicated-class plan met a complex-class reality, and the probe returned the product.

Agent-based emergence, the class lens

Assumes: interacting agents produce whole-system behaviour that no analysis of the parts recovers in advance.

Fits because: emergence topped the list at 0.85 while analog density bottomed out at 0.4.

Breaks when: the problem is genuinely complicated. A specialist settles in a day what probing takes a quarter to reach, and the experiment is waste.

Counteracts: method attachment, the entrained thinking that dominates both ordered domains.

May reinforce: experiment theatre, probing questions that already have written answers.

2. Validity: how much to believe a single result

Complex problems return noisy feedback, so one result rarely separates a good decision from a good outcome. Mauboussin’s paradox of skill carries the mechanism. Where participants interact and luck is present, rising skill compresses the spread: absolute results improve, relative results cluster, and luck decides more of what remains. Measure relative position rather than absolute improvement.⁷

Kahneman and Klein turn that into a purchase decision. Two conditions must hold before anyone’s intuition is worth buying. The environment needs high validity, meaning stable links between identifiable cues and what happens next. And the person needs practice inside it, with feedback that arrived fast and unambiguously. They put medicine and firefighting on the high side. Individual stock prices and long-range political forecasts sit near zero validity, where confidence indicates nothing.⁸

They also name the trap that sells most consulting hours: fractionation of skill. Genuine expertise in one task gets stretched to judgments nothing ever trained the person for, and the edge of the real competence stays invisible from both sides.⁸ That is the growth advisor with fifteen years in one channel pricing your entry to another.

Before you buy an expert, ask which feedback loop taught them. If they cannot name it, you are buying confidence.

Luck-skill continuum, the calibration lens

Assumes: results mix skill and luck in proportions you can estimate from how much performance persists, and validity is a property of the environment.

Fits because: a 0.8 on deep uncertainty leaves a single outcome carrying almost no information.

Breaks when: the sample is one and the stakes are terminal. Regression to the mean consoles nobody with six weeks of runway.

Counteracts: crediting your method for a result the environment produced.

May reinforce: paralysis, treating every domain as too noisy to act in.

3. Transitions: a class is a state with a movement rate

Classify once and you own a label. Instrument the boundaries and you own a state with movement rates that run faster one way than the other. Kurtz and Snowden call the fall from known into chaotic an asymmetric collapse. Organisations settle into stable relationships in known space and miss that the environment changed until the system breaks, and the longer the stability the likelier the break. The usual return path is imposition, a forced march back to order that installs a fresh rigidity.⁴ The current wiki draws that boundary as a catastrophic fold, a cliff, and warns that the liminality inside Clear is invisible, so confidence in past practice walks you off it blind.³

Kenyan digital lending ran in the Clear state for years: no licence, a widely copied playbook, best practice traded between operators. The Digital Credit Providers Regulations of March 2022 installed a gatekeeper. Four years and 800-plus applications later, the Central Bank of Kenya’s April 2026 tally stood at 227 licensed providers, whose lending reached Ksh 133.5 billion across 7.5 million loans by February.⁹ Every unlicensed operator’s best practice became an offence on one date. Nothing in their unit economics forecast it. The observable ran in the gazette.

Order is a state you can be evicted from, and the notice usually arrives as a rule change rather than a bad quarter.

The chaotic state has its own literature. Weick’s study of the 1949 Mann Gulch fire, where thirteen men died, records the foreman shouting at his crew to drop their tools, lighting an escape fire, lying down in its ashes and surviving. Nobody followed him in. Weick proposes four sources of resilience against that collapse: improvisation and bricolage, virtual role systems, the attitude of wisdom, and norms of respectful interaction.¹⁰ In the chaotic state you act first and sense second, and nobody follows because a novel act looks like an error to people still running the previous method.

Markov state-transition, the movement lens

Assumes: problems occupy states with watchable transition rates, and those rates run faster in one direction than the other.

Fits because: regime-break risk came in at 0.7.

Breaks when: the transition matrix itself gets rewritten, which is both the case the model handles worst and the case that costs most.

Counteracts: classifying a problem once and filing it.

May reinforce: false comfort, since a state you can name feels like a state you control.

GEER: rank the levers by what they cost to undo

Four channels carry the exposure: classification, method fit, feedback validity, transition exposure. Reversible levers first.

1. Sort the open list. One question per decision. Does a written answer exist, does a payable specialist hold it, or does nobody hold it. Hits classification. Costs an hour.
2. Write the retirement condition. Before analysis starts, write the number that ends it. A complicated problem has one. A blank means you misclassified. Hits method fit.
3. Write the prior before the probe. One sentence, one number, plus what would surprise you. Ten minutes converts a launch into an experiment. Hits feedback validity.
4. Cap and parallelise. Three to five probes, each bounded by a share of monthly burn and a fixed day count, each survivable alone. Hits method fit.
5. Run the validity test on every advisor. Stable cues, repetition, fast feedback. Where the environment fails, price the advice as opinion. Hits feedback validity.
6. Instrument two boundary observables. A gazette notice, a settlement delay, a partner rewriting terms. Hits transition exposure. Costs an afternoon.
7. Write the chaotic playbook. Who decides, the first four actions, what cash is reachable in 24 hours. Hits transition exposure.

No-lever flag: when the same problem is Clear for one part of the business and complex for another, you sit in the aporetic state and no single method fits. Break the issue down until each piece sits inside one domain, which is what the aporetic liminal exists to do.³ Averaging across classes produces a method that fits none of them.

RADAR: put dates on it before the week fills up

DO NOW, by T+3 days. Reversible, dominant in every scenario.

1. Take the seven decisions eating your week. Put a class on each, written where the team can see it.
2. Next to each, write the method you have actually been using. Circle every mismatch. Those are losses already incurred.
3. Stop any analysis on a complex-class item with no retirement condition. Move those hours into probe design.
4. Write the prior for every experiment already running. With no stated expectation, the result teaches nothing whichever way it lands.

HEDGE, by T+14. Cheap insurance against a class you got wrong.

1. Run three parallel safe-to-fail probes on your largest complex-class item. Each capped, each survivable alone, each with a written amplify rule and dampen rule.
2. Apply the two-condition validity test to your two most expensive advisors. Keep the ones whose environment could have taught them.
3. Write the one-page chaotic playbook. Store it where someone other than you can reach it.

DEFER AND TRIGGER. None of these reverse. Wait, and write down the observation that releases the wait.

1. Defer: repricing the whole book, rebuilding the product, entering a second market, restructuring the team.
2. Trigger to commit: two of three probes cross the threshold you wrote in advance, by T+28. Commit the following week and close the question.
3. Counter-trigger: none cross, and the prior missed in the same direction all three times. A consistent directional miss is a classification failure. Go back to the sort, not to the tactics.

For anyone on the board side of this. DO NOW: ask which class an agenda item occupies before you ask for the number, since a complex-class item carrying a confident forecast is the tell. HEDGE: fund probe budgets separately from operating budgets, so a founder never chooses between an experiment and payroll. DEFER: judging a founder on one complex-class outcome until you have watched four.

CHAIN: the second and third order, and where the rules move

Draw the comparison set from operators who had to act while cause and effect were visible only afterwards. Incident command, drug discovery, new-market entry and counter-insurgency share that structure, and structure carries the base rate. An industry label carries nothing.

The largest test of this discipline is the 2024 replication of the Italian trial: four randomised control trials, 759 firms, 11,463 data points. The intervention raised idea termination across all three new trials, and its effect on radical pivots was non-linear. Treated firms became likelier to pivot radically once or twice, and less likely either to never pivot or to keep pivoting indefinitely.¹¹ Plan against that base rate: teaching operators to write a theory and test it makes them kill more ideas and wander less between them.

Three modifiers tilt it. Expert-grade analysis is close to free now, so complicated-class output looks identical whether or not the problem was complicated, which raises the price of a misclassification. Capital is tighter, so fewer parallel probes are affordable, which raises the return on writing the prior. Regulatory activity across African digital finance is heavy, lifting the exit rate from the Clear state for whole categories at once.⁹

Now take out what would have happened anyway. Both arms of those trials received ten sessions of general training on reading market feedback, so the measured gain sits on top of that floor and reports the marginal value of an explicit theory, not the value of thinking at all.⁵ Part of what founders credit to sharp classification is the ordinary return on writing anything down.

Matrix-break flag. Two rewrites are live. The price of complicated-class work has collapsed, moving the scarce skill from producing analysis to deciding whether analysis is the right instrument. And regulators are creating gatekeeping states faster than operators instrument them, which shortens the warning before a Clear-to-Chaotic fall. If both hold, the sorting step gains value while the analysis step loses it. Move budget in that direction.

What this ensemble cannot see

The widest gap sits inside the framework itself.

Cynefin will not classify for you. Its authors treat it as a sense-making device whose worth lies in its effect on the people using it, and separate that firmly from a categorisation framework that files you into a quadrant.⁴ The judgment stays yours.

Whose method wins the argument. Kurtz and Snowden watched people compete to interpret the undecided centre according to their preferred mode of action. Rule-makers pull toward order, experts pull toward research and data, and the more important the issue, the harder each pulls it toward the domain where their own capability makes them powerful.⁴ Classification inside a real company is settled partly by budget authority, and nothing here prices that.

Simultaneity. Payroll can be Clear while pricing is complex and a partner’s compliance review is chaotic, all in one week. These models each run on one problem. None allocates attention across four at once.

Whether your probe carried signal. In a low-validity environment a probe sometimes crosses its threshold on luck, and no write-up fixes a sample of one.

Act on the part that survives all four. This week, list the seven decisions eating your calendar. Beside each, write the one observation that would prove your classification wrong. Any blank line is a guess wearing a classification’s clothes. Fill the blanks by T+7, before you spend another shilling on analysis or on experiments.

Sources and notes

1. Snowden, D. J., and Boone, M. E. “A Leader’s Framework for Decision Making.” Harvard Business Review, November 2007. The five contexts, the sense-categorise-respond and sense-analyse-respond sequences, probe-sense-respond, safe-to-fail experiments, entrained thinking, and the observation that the complex domain is more prevalent in business than most leaders realise are all in the article text. hbr.org serves a script shell rather than the article body to a plain fetch, so this note links an accessible full-text mirror: full text PDF.
2. Snowden, D. J. “Cynefin St David’s Day 2020 (1 of 5).” The Cynefin Co, 2020. The section headed “The change from Obvious to Clear” records the rename and Mary Boone’s advocacy for it. Post.
3. “Cynefin Domains,” Cynefin.io wiki. Source for the aporetic and confused central domain from March 2020, the liminal line intersecting all domains but Clear, and the Clear-to-Chaotic boundary described as a catastrophic fold or cliff where excessive confidence in past practice carries you over blind. Wiki entry.
4. Kurtz, C. F., and Snowden, D. J. “The New Dynamics of Strategy: Sense-Making in a Complex and Complicated World.” IBM Systems Journal 42(3), 2003, 462-483. Source for asymmetric collapse and imposition at the known-chaos boundary, for the distinction between a sense-making framework and a categorisation framework, and for the finding that people pull an important issue toward the domain where their own capabilities empower them. Full text PDF.
5. Camuffo, A., Cordova, A., Gambardella, A., and Spina, C. “A Scientific Approach to Entrepreneurial Decision Making: Evidence from a Randomized Control Trial.” Management Science 66(2), 2020, 564-586. 116 Italian startups, 16 data points over roughly a year, both arms given ten sessions of general training. Full text PDF.
6. International Finance Corporation, “M-Money Channel Distribution Case: Kenya, Safaricom M-PESA,” IFC mobile money toolkit, World Bank Group. Source for the pilot dates of 11 October 2005 to 1 May 2006, eight agent stores and nearly 500 trial participants across three locations, the Faulu Kenya microfinance integration and its obstacles, and the finding that national money transfer was the killer application behind the send-money-home campaign. Case study PDF.
7. Mauboussin, M. J. “The Paradox of Skill: Why Greater Skill Leads to More Luck,” ChangeThis manifesto 100.03, drawn from The Success Equation. Source for absolute results improving while relative results cluster, and for measuring relative rather than absolute performance where interaction and luck are present. Full text PDF.
8. Kahneman, D., and Klein, G. “Conditions for Intuitive Expertise: A Failure to Disagree.” American Psychologist 64(6), 2009, 515-526. Source for the two necessary conditions, the high-validity and zero-validity descriptions of task environments, subjective confidence as an unreliable indicator, and fractionation of skill. Publisher page bot-blocks a plain fetch; open mirror at the University of Ljubljana: full text PDF.
9. Central Bank of Kenya, “Press Release: Licensing of Digital Credit Providers,” 14 April 2026. 227 licensed DCPs, more than 800 applications received since March 2022, and 7.5 million loans valued at Ksh 133.5 billion as of February 2026. Press release PDF. Underlying instrument: Central Bank of Kenya (Digital Credit Providers) Regulations, 2022.
10. Weick, K. E. “The Collapse of Sensemaking in Organizations: The Mann Gulch Disaster.” Administrative Science Quarterly 38(4), 1993, 628-652. Thirteen deaths, the order to drop tools, the escape fire nobody else entered, and the four proposed sources of resilience. Mirror hosted by the US Forest Service Lessons Learned Center: full text PDF.
11. Camuffo, A., Gambardella, A., Messinese, D., Novelli, E., Paolucci, E., and Spina, C. “A Scientific Approach to Entrepreneurial Decision-Making: Large-Scale Replication and Extension.” Strategic Management Journal 45(6), 2024, 1209-1237. Four randomised control trials, 759 firms, 11,463 data points, higher idea termination and a non-linear effect on radical pivots. Open-access copy: City, University of London repository PDF.