Complex Problem Framing & Systems Thinking for Global Teams
The expensive mistake is treating a complex problem as a hard one
Most international work that goes badly is not under-analysed. It is misclassified. A global team treats a living system as a puzzle, hires more experts, writes a thicker plan, and then wonders why the plan dies on contact with the first country that was not in the workshop.
Dave Snowden and Mary Boone put the distinction in the November 2007 Harvard Business Review essay that still does the most work here: A Leader’s Framework for Decision Making. They sort situations by how cause and effect relate. In the complicated domain, experts can find a good answer. In the complex domain, “cause and effect can only be perceived in retrospect.” The prescribed move is not Sense–Analyse–Respond. It is Probe–Sense–Respond.
That is the whole argument of this piece, applied to work that crosses borders. If you leave with one change in behaviour, let it be this: name the domain before you name the solution. Then frame the problem as a system with stocks, delays, and conflicting purposes, not as a ticket with an owner. Employers who still rank “analytical thinking” as the top core skill (the World Economic Forum’s Future of Jobs Report 2025, 69 percent of surveyed employers, more than 1,000 firms covering over 14 million workers in 55 economies) are not asking you to memorise a playbook. They are asking whether you can stop the playbook from running when the situation will not sit still.
I would not bother with another “systems thinking” poster of interlocking circles. I would bother with domain diagnosis, a written problem frame that survives contact with legal and payroll, a causal map that shows where the delay actually lives, and a small set of safe-to-fail probes. The rest is decoration.
Analytical thinking is not the same as problem framing
The WEF 2025 Skills Outlook is often quoted as if “analytical thinking” and “complex problem-solving” were one skill. They are not. Analytical thinking, as employers mean it, is the ability to take a defined problem apart. Framing is the prior act of deciding what counts as the problem, whose loss function you are using, and which interactions you refuse to cut out of the picture.
Saadia Zahidi, introducing that same report, notes that employers now expect 39 percent of workers’ core skills to change by 2030, down from 44 percent in the 2023 edition, still high enough that a memorised solution set has a short half-life. The EU Digital Skills and Jobs Platform’s write-up of the same survey adds the operational sting: 63 percent of those employers name skills gaps as the largest obstacle to business transformation. You cannot close a gap of that size by teaching people yesterday’s answers.
Horst Rittel and Melvin Webber, in Dilemmas in a General Theory of Planning (Policy Sciences, 1973), already described why the “define, then solve” sequence fails for social and organisational problems. Their first two properties are still the ones global teams violate every week:
There is no definitive formulation of a wicked problem.
Wicked problems have no stopping rule.
Plain reading: you cannot finish the requirements document and then “do delivery.” Every formulation of “we need a single global CRM” or “we need one headcount model” already contains a theory of the organisation. If Legal, Sales, and the country entity disagree about the purpose of the system, you do not have a data-model dispute. You have an unresolved wicked problem wearing a Jira ticket.
Rittel and Webber’s third property is the one finance people hate: solutions are not true or false, they are good or bad. Transfer-pricing policy, a data-residency choice, a 24/7 follow-the-sun rota, a decision to keep people on a home contract versus an Employer of Record: none of these has a unique correct answer. They have distributions of harm across jurisdictions, time zones, and balance sheets. Framing is the work of making those distributions visible before someone “decides.”
Cynefin first, or you will staff the wrong room
Snowden built Cynefin at IBM in the late 1990s; the 2007 HBR piece with Boone is the version executives actually read. Five contexts: Clear (once called Simple), Complicated, Complex, Chaotic, and Disorder in the centre. The Welsh word means a place of multiple belongings, which is already more honest than “matrix organisation.”
Clear: cause and effect are obvious. Sense, categorise, respond. Best practice. Expense policy, visa photocopy checklist, how you file a statutory accounts pack in a country you already know. Do not workshop these.
Complicated: cause and effect exist, but you need experts. Sense, analyse, respond. Good practice, not best. Aircraft engines, tax treaties, a multi-entity consolidation. Several right answers can coexist. Bring specialists. Do not confuse “we need a lawyer” with “we need an experiment.”
Complex: patterns only after the fact. Probe, sense, respond. Culture, market entry, hybrid work across six time zones, any change whose success depends on how other people react to the change. Enabling constraints, not governing ones. Snowden’s June 2026 definition on The Cynefin Company’s site is worth quoting because he wrote it to stop systems-thinking people collapsing complexity into “it depends how you look at it”:
A complex system is a system of actants and interactions, carrying propensities and dispositions shaped by path dependency and entanglement, whose emergent properties cannot be derived from any prior state of those actants or their interactions.
Complex systems are characterised by enabling constraints rather than the governing constraints of ordered systems. Enabling constraints shape what can happen without determining what does.
Plain reading: you can change the plausibility of outcomes (who talks to whom, what information is visible, which experiments are cheap) without being able to specify the outcome. A Gantt chart that pretends otherwise is a category error. Snowden’s De Iudicio paragraph is the part I would print for any team that wants to outsource this to a model: judgement “cannot be disentangled from those conditions without ceasing to be judgement,” and pattern-matching at scale is not the same as bearing accountability for what emerges.
Chaotic: act first to impose a constraint, then sense. A ransomware event, a sudden travel ban, a payroll file that will not run on the 25th. Analysis is delay.
Disorder is the state you are actually in at the start of most global programmes: different functions argue from different domains. Finance thinks the problem is complicated (model it). Product thinks it is complex (ship and learn). Legal thinks it is clear (the statute says X). Until you split the work by domain, the meeting is theatre.
The failure mode I care about is the one Snowden has been repeating for twenty years: treating the complex as complicated. You commission a 90-day diagnostic, produce a target operating model, and then discover that the Singapore entity, the German works council, and the US product org do not share a purpose. More analysis will not create a purpose.
Write the frame as a system, not as a sentence
Peter Checkland’s Soft Systems Methodology exists because “the problem” in human activity systems is contested. CATWOE is the memory aid; it is also the cheapest way to stop a global team talking past each other.
Customers: who is helped or harmed. In a cross-border payroll change, that is not “the business.” It is employees in each tax residency, the local entity directors who carry personal liability, and the customer who still expects a named human in their morning.
Actors: who actually does the work. HQ process owners are usually not the actors. Country HR ops, a local accountant, an EOR coordinator, a night-shift SRE: those are the actors.
Transformation: input to output, in one clause. “From people paid under five local contracts with inconsistent benefits, to people paid under a documented model that is legal in each country of work.” If you cannot say the from–to, you do not have a transformation. You have a slogan.
Weltanschauung: the worldview that makes the transformation worth doing. This is the field people skip, and it is where programmes die. “We are one company, so one process” is a worldview. “Directors in Germany will not sign anything that conflicts with the BetrVG consultation duty” is another. Both can be true. CATWOE forces you to write which one you are serving in this intervention.
Owners: who can stop it. Not the steering committee. The person who can refuse a works-council filing, a banking KYC pack, or a data-transfer mechanism.
Environmental constraints: GDPR, labour codes, time zones, banking cut-offs, language. Constraints are not “risks to be mitigated.” They are the shape of the system.
A root definition then looks like this: a system, owned by the regional CFO, operated by country payroll and an EOR partner, to convert hours and local statutory data into compliant pay and filings for employees working outside their contract country, because the firm will not carry undeclared permanent-establishment risk, within the constraint that personal data of EU staff does not leave an adequacy decision without a documented transfer tool.
That sentence is ugly. It is also a frame. Compare it with “fix global payroll.” One of those can be argued with. The other cannot.
I read Keeney-style value-focused thinking as the sibling of CATWOE for the cases where the fight is over objectives rather than actors. Write the fundamental objectives first (stay legal in every place of work; keep time-to-pay under the local statutory deadline; do not split the employee’s tax year by accident). Then generate alternatives. Teams that start with “should we use an EOR or incorporate?” have already smuggled in a means objective and will spend six months comparing vendors.
Stocks, delays, and the reason last quarter’s fix is this quarter’s incident
Jay Forrester’s system dynamics, and Donella Meadows after him, give you the grammar once the frame exists. A stock is an accumulation: headcount in a country, open tickets, cash in a local entity, trust between a country GM and HQ. A flow changes the stock. A delay is the time between action and visible effect. A reinforcing loop compounds. A balancing loop pushes back.
Peter Senge, in The Fifth Discipline (1990; 2nd ed. 2006), made this usable for organisations. SEBoK still treats the book as a primary reference for “what is systems thinking,” and Senge’s claim is blunt: systems thinking is the discipline that makes the other four (personal mastery, shared vision, mental models, team learning) capable of handling complexity. His laws are the failure catalogue of global programmes. Two of them do most of the damage.
“Today’s problems come from yesterday’s solutions.” The follow-the-sun support rota that saved overnight coverage now produces a 14-hour wait for anyone whose incident falls in the overlap gap. The “one global grade structure” that simplified compensation now collides with local collective agreements.
“Cause and effect are not closely related in time and space.” You cut a vendor in Dublin in March. The works-council complaint appears in Munich in September. The person who cut the vendor has already been promoted.
“The harder you push, the harder the system pushes back” is the one that looks like resistance and is usually a balancing loop doing its job. Mandate a single ticketing tool from HQ and the country teams will open a shadow spreadsheet, because their local regulator still wants an audit trail in the local language. Pushing harder on the mandate strengthens the shadow system.
“Dividing an elephant in half does not produce two small elephants” is the org-chart version. Splitting “global process” from “local compliance” so that each has a KPI does not give you two healthy functions. It gives you a handoff where the delay lives.
Meadows’s 1999 essay Leverage Points: Places to Intervene in a System (Sustainability Institute) is the ranking people cite and then ignore. She defined leverage points as places “where a small shift in one thing can produce big changes in everything.” The ranking, weakest to strongest, starts at constants and parameters (tax rates, headcount caps, SLA minutes) and ends at paradigms and the power to hold a paradigm lightly.
Almost every global transformation I have seen argued about spends its political capital at point 12: numbers. Add five people. Cut the budget 8 percent. Change the NPS target. Meadows’s point is that parameters are where debate is loudest because they are the least threatening. Information flows (point 6) are usually the first real lever a distributed team can actually pull: who sees the queue, in which timezone, in which language, with which lag. Rules (point 5) are the employment-model choice, the data-transfer mechanism, the approval matrix. Goals (point 3) are “are we optimising for identical process or for local legality?” Paradigms (point 2) are “HQ is the centre” versus “the entity director is the legal person.”
If you only change the number of coordinators, you have not found leverage. You have bought a buffer (point 11) and called it a strategy.
Eliyahu Goldratt’s Theory of Constraints is the operations-shaped cousin, and it is the right tool when the system is closer to complicated than complex: a throughput chain with a bottleneck. The five focusing steps from The Goal are Identify, Exploit, Subordinate, Elevate, Repeat. The error in global ops is identifying a local inefficiency (a country team that looks slow) and “elevating” it with more staff, while the constraint is a weekly HQ approval that sits in a US afternoon mailbox. Subordinating everything to a constraint you have not identified makes the rest of the system busier and the output unchanged.
Use TOC when you can point to a queue. Use Cynefin plus Meadows when the queue is a symptom of conflicting goals.
A worked frame: “We cannot hire in Country X”
Take the request that lands on every mobility and people-ops desk. Sales wants a person in a market. Legal says no entity. Finance says no PE risk. The country manager says competitors already have people on the ground. Someone asks for “a process.”
Domain first. Entity setup, payroll registration, and a work-permit checklist are complicated. Whether customers, regulators, and your own managers will treat a remote employee as a local presence is complex. The tax authority’s view of a dependent agent PE is complicated-to-complex depending on the jurisdiction and the facts; it is not something you vote on in Slack.
CATWOE, compressed. Customer: the employee (who needs a lawful contract and social insurance), the customer (who wants a local contact), the entity directors (who do not want unplanned PE). Actor: mobility, a local counsel, perhaps an EOR. Transformation: from an informal promise of “someone on the ground” to a documented right-to-work and payroll path. Worldview: pick one and write it down. “We will not create PE” and “we will look local to the buyer” are not the same programme. Owner: whoever can refuse the contract. Environment: immigration rules, labour law, corporate tax tests, data residency if the person will touch EU personal data.
Now the loops, not the org chart. Reinforcing: more local wins → more pressure to hire locally → more PE risk → more legal friction → slower hires → missed pipeline → louder sales pressure. Balancing: legal review slows hiring, which reduces PE exposure, which is the point of legal review. If you “remove blockers” without changing the goal, you have disabled the balancing loop on purpose.
Leverage. Adding a recruiter is point 12. Publishing a written decision tree (when EOR is allowed, when a trip is a business visit, when you must incorporate) is point 6, information. Changing the rule that sales compensation ignores delivery cost of a non-compliant hire is point 5. Changing the goal from “headcount in-market” to “coverage without unplanned taxable presence” is point 3. I would spend the week on the decision tree and the incentive rule. I would not spend it on another vendor bake-off until those two exist.
Probes, because part of this is complex. A 90-day EOR hire in one city, with a pre-agreed kill criterion (if the person cannot be paid on the statutory date, stop). A documented business-visitor pattern for the sales lead, with day-count tracking. A local counsel memo on the PE tests that names the facts you must not create. Amplify what does not blow up. Do not write a three-year target operating model first.
What goes wrong if you skip this. The hire happens on a home contract. Six months later you have a permanent establishment argument, a social-security gap, and an employee who cannot be dismissed under the law they actually work under. That is Rittel’s fifth property: every attempted solution is a one-shot operation. You can unwind a vendor. You cannot unwind a year of facts on the ground.
Trade-offs you should weigh in public
Identical global process versus local legality. Identical process wins when the work is Clear or Complicated and the statute is actually the same (it rarely is). Local legality wins the moment a director has personal liability. I would default to a thin global spine (what must be true everywhere) and thick local flesh. The spine is: right to work, a contract that matches the place of work, pay on time, data handled under the correct transfer tool. Everything else is local.
Speed versus reversibility. In the complex domain, reversibility is the constraint. A policy you can turn off in a week is worth more than a better policy that requires a works-council negotiation to unwind. Snowden’s “safe-to-fail” is not a slogan for timid people. It is an admission that you cannot know.
Central expertise versus distributed sensing. Complicated problems want a centre of excellence. Complex problems want many sensors. A follow-the-sun engineering org that routes every incident through a single US manager has built a delay into the balancing loop. The stock that then grows is unresolved severity-1 time, not “alignment.”
Home contract, EOR, local entity. These are not three equal options. Home contract is a short-stay instrument; it fails when presence becomes employment in the destination. EOR buys speed and a local employer of record at the cost of another intermediary in the information flow (Meadows point 6 again: you now have a vendor who sees the employee data before you do). A local entity is the right answer when the goal is durable presence and you will accept the overhead. The frame decides. The vendor does not.
Time zones as delays, not as culture. A decision that needs three signatures across eight hours of offset is a delay longer than the rate of change of the incident. Shorten the delay (point 9) or change the rule so two of the signatures are not required. Do not run a workshop on “collaboration.”
Language and legal meaning. The English translation of a labour code is often softer than the original. Where the topic is jurisdiction-specific, the authority’s page in its own language is the source. If the English FAQ and the statute disagree, the statute wins. That discrepancy is itself publishable inside the team: write both wordings in the frame.
Engineers, counsel, and country managers will not use the same artefacts. Engineers will tolerate a causal-loop diagram. Counsel want the statutory test and the facts pattern. Country managers want to know who they call at 02:00. Produce all three from the same frame. If you only produce the diagram, you have done theatre for the strategy offsite.
What “good” looks like in the artefacts, not in the adjectives
A problem statement that names the domain in the first line. “Complex: we do not know how local customers will treat a remote technical account manager; we will run two probes.” Or: “Complicated: we need a correct PE analysis for a dependent-agent fact pattern in France; instruct counsel.” If the statement starts with “align stakeholders,” throw it away.
A CATWOE of one page, with the worldview in a sentence a director can reject.
A loop sketch. Boxes and arrows on a whiteboard are enough. Label at least one delay with a unit of time (hours, payroll cycles, quarters). If you cannot name a delay, you do not yet understand the behaviour.
A leverage choice that is not a parameter. Write the point on Meadows’s list. If you cannot, you are probably arguing about a number.
A probe design: what you will try, what you will measure, what will make you stop, who is accountable for the mess if it fails. Time-box it. Snowden’s language is “safe-to-fail,” which means the failure has to be affordable in money, law, and reputation.
A decision log that records the frame you rejected. Six months later someone will propose the rejected frame again. That is not stupidity. That is a system without memory (Cilliers: distributed memory, no central controller). Write it down.
For technical readers who want a working object, the naive version is a linear RCA:
What breaks: you have treated a delay-and-rule problem as a person problem. Senge’s eleventh law is “There is no blame.” The structure produced the lateness.
The version that matches the system:
That block is not software. It is a specification. A team can argue with it.
Who this does not cover, and what does
If your work is genuinely Clear (a known filing, a known API, a known visa photocopy list), do not apply any of this. You will annoy competent people. Use the checklist.
If you are a single specialist doing complicated analysis (treaty interpretation, a performance model, a circuit), stay in Sense–Analyse–Respond. Cynefin does not make expertise obsolete. It stops you using expertise as a substitute for experiment when the system can talk back.
If you are in a true emergency, you are in Chaotic. Act. Stabilise payroll, isolate the network, get people out. Frame later.
If your organisation will not write down a worldview because two executives need the contradiction to remain implicit, you do not have a methods problem. Meadows would call that a goal conflict. I would not run a systems workshop on top of it. I would name the two goals in a document and force a choice, or leave the programme.
Distributed teams with no overlapping hours need the delay made explicit more than they need another values statement. Co-located teams in one HQ will underweight local legality; their frame should overweight the owner who can be prosecuted.
Nationals working in a third country, contractors, and EOR employees are different actants. Do not put them in one loop and call it “workforce.” Immigration status, tax residency, and who is the legal employer change the constraint set. If you cannot say which of those three a person is, you cannot frame the problem.
Do this in the next two days, not after the offsite
Take one live international issue that is already late. Write the Cynefin domain on the first line. If two functions disagree, you are in Disorder; split the issue until each piece has a domain.
Fill CATWOE in twenty minutes, out loud, with one person who will be harmed if you are wrong. If the worldview will not fit in a sentence, you are not ready to buy software or hire a vendor.
Mark one delay with a clock. Mark one rule you could change without a budget committee. That pair is your first leverage, and it will feel too small. Meadows’s eighth law, via Senge: the areas of highest leverage are often the least obvious. The obvious move is the extra headcount.
Then run one probe that can fail in public. The World Economic Forum can keep ranking analytical thinking at 69 percent. The skill that actually travels is the refusal to analyse a system that needed a probe, and the refusal to probe a statute that needed a lawyer. Get that classification right and the rest of the work is ordinary. Get it wrong and you will produce a very professional plan for a country that does not exist.