Systems Thinking

Confidence 0.82 · 5 sources · last confirmed 2026-05-22

A mode of reasoning and innovation that focuses on flows, relationships, feedback loops, and unintended consequences across an interconnected ecosystem, rather than on a bounded product, service, or user. Distinguished from breakthrough thinking (Silicon-Valley “move fast and break things”) and design thinking (IDEO-style user-centric iteration).

Working definition

Per Bansal & Birkinshaw (2025), building on Bertalanffy’s general-systems theory, Jay Forrester’s system dynamics, and Peter Senge’s The Fifth Discipline (1990):

• Systems thinking = recognizing that the modern economy is a network of people, products, finances, and data; changes in one node have side effects in others; innovation must be designed with these flows in mind.
• The aim is to make entire systems more sustainable and resilient, accepting that benefits in one part of the system may be outweighed by harm done elsewhere if the cross-effects aren’t traced.

Key claims

Three modes of innovation (Bansal & Birkinshaw 2025)

Mode	Method	Strengths	Side effects
Breakthrough thinking	Slice the Gordian knot; “10× / winner-take-all”; Zuckerberg’s “move fast and break things”	Speed; dramatic progress on bounded problems	Knock-on damage; well-suited only to clearly bounded problems (rocketry)
Design thinking	Empathy-driven iteration on user need; IDEO/Senge popularization	User-centric clarity; cuts through complexity	Obsession with the user creates problems for non-users (e.g., Airbnb solving for hosts/travelers but harming local housing)
Systems thinking	Map flows + relationships; iterate problem framings	Avoids unintended consequences; embraces complexity	Slower; harder; “least common mode of innovation”

Why systems thinking is rare in practice

• Traditional approach demands modeling all flows, interactions, feedback loops — daunting in fast-changing worlds where models can’t reflect reality.
• Systems thinkers spend time figuring out exactly how the Gordian knot is tied — almost guaranteed to be overtaken by a design thinker (slicing) or a breakthrough thinker (single-strand focus).

Streamlined four-principle approach (Innovation North initiative, Ivey Business School)

1. Define your desired future state — articulate a North Star for the firm’s role in the future system. Example: Maple Leaf Foods’ shift from “meat processor” to “the most sustainable protein company on Earth” (CEO Michael McCain).
2. Frame the problem, reframe it, and repeat — wicked problems lack a single definition; iterating reframings engages stakeholders who experience the system’s dysfunctions differently. Example: U of Guelph reframing climate change → soil health for farmer engagement.
3. Focus on flows and relationships, not products or services — Co-operators insurance introduced “drying in place” and “soft contents” cleaning to redirect the flow of damaged materials away from landfills, without launching a new product.
4. Nudge your way forward — pursue an “ecology of actions” rather than a moonshot/silver bullet. Example: CSA Group’s circular built environment program — small actions across architects, engineers, developers, owners, plus a “coalition of the willing.”

Wicked problems (Rittel & Webber, 1973 — implicit reference)

• Constantly changing, hard to define, multiple stakeholders with divergent experiences of the system’s dysfunctions.
• Solutions involve difficult trade-offs.
• Systems thinking is best suited here; breakthrough/design thinking misfit.

The MIT system-dynamics articulation (Sterman 2026)

Sterman 2026 (Jay W. Forrester Professor of Management, MIT Sloan; director of MIT Sloan Sustainability Initiative) delivers the modern in-house MIT-system-dynamics treatment in 58 minutes, supplying the operational machinery that sits underneath the three-modes-of-innovation framing above. Six load-bearing claims:

Policy resistance is the universal failure mode

A well-researched, expert-vetted solution either fails outright or works near-term and the problem returns elsewhere. “Extremely common and extremely discouraging.” Sterman names five examples: urban traffic (build-more-roads → more driving → worse congestion); US healthcare cost containment (~18% of GDP, 70 years of failure); failed M&A; failed process improvement (TQM / Six Sigma / BPR — “the tools work; vastly more attempts fail than succeed, breeding cynicism about the next one”); and project mismanagement (“typically late, expensive, and wrong”).

Historical anchor: Sir Thomas More, Utopia (1516): “it will fall out as in a complication of disease that by applying a remedy to one sore you provoke another. And what removes one ill symptom produces others.”

Feedback is the load-bearing primitive, not “complexity”

Policy resistance is not caused by complexity per se. “It’s that there’s a mismatch between the mental models that we use to come up with our solutions and the complexity of the real world.” The named bad mental model is the open-loop sequence: Issue → Data → Evaluate → Optimal Choice → Implement → “Problem solved.” Sterman: “No project I’ve ever done — from co-chairing the new Sloan School building to putting together my grocery list — has ever gone that way.” The reality is feedback: decisions change the world, which generates new information conditioning the next decision. The bicycle is the load-bearing analogy — you can’t know which way to turn the handlebars without feedback telling you where you are relative to your goal.

”There’s no such thing as a side effect”

A signature Sterman reframing: there are only effects, some intended, some that “feed back and undermine your goals.” When managers explain failure as “that was an unintended consequence, nobody could have anticipated that,” what they are “really communicating is how narrow and inadequate is the mental model they’re using.” Bad explanations are mental-model diagnostics.

Multi-stakeholder system pull-back

Even with a complete mental model, you are not the only actor: “You have your employees, your customers, the financial markets, your suppliers, the communities in which you operate, and they all have their own goals.” The dynamic: “when you pull the state of the system closer to your goals, you’re almost certainly pulling it away from theirs. And they aren’t just going to sit there. They’re going to take their own actions and try to pull the system back.” Other actors’ mental models are also limited → their reactive actions also generate unintended effects. This is the operational-scale version of why systems thinking is required.

Causal mapping + group modeling — “the system in the room”

The qualitative toolkit. Causal mapping makes both above-the-waterline balancing loops (the part managers see) and below-the-waterline unintended consequences visible. Sterman’s worked example is the US prior-approval / preferred-drug-list / step-therapy machinery in healthcare: $35B/year in admin cost as of 2024 ($11K per clinician/year; $20–30/request); a 1996 study and an earlier-2026 meta-review of 25 studies both find prior approval is associated with disease exacerbation, preventable hospitalizations, prolonged stays, and lower disease-free survival — “medical jargon for more people died as a result of prior approval.” The mapping technique reveals the below-waterline feedback (appropriateness and timeliness of care → patients return sicker → costs rise).

The operational saying: “You want the system in the room” — convene a group-modeling process with diverse participants, including traditional adversaries (consumer groups, upstream-supply-chain stakeholders), so the causal map reflects the actual system rather than the convener’s pre-existing mental model.

Management flight simulators — “lectures don’t work”

The quantitative toolkit. “Research shows that showing people research doesn’t work.” Mental models are “strongly reinforced by everyday experience” — short-run feedback from bad decisions feels rewarding, long-delayed unintended consequences appear disconnected from the original choice. The only way to update a mental model is to let people experience the consequences themselves in a costless simulated environment.

The Sully Sullenberger anchor (a certified simulator instructor before the Miracle on the Hudson): “I’ve never experienced a bird strike that knocked out both engines, but we practiced that in the simulator many many times.” Sterman’s live project-sim demo in the webinar walks the audience through a worked policy-resistance trajectory: 75→70 staff (budget cut), scope creep accepted (chase 3.6% market share), “do it right, but get it done” management pressure, 78-hour work weeks, final 25% defect rate against <1% target, ~$50M NPV loss. The simulator’s realistic constraints don’t let the player escape mid-flight (“reduce project scope? Won’t let me. Got to keep all those bells and whistles in there”) — by design, because the goal is mental-model update, not score-maximisation.

Other named successful applications: kidney dialysis anemia treatment (MIT Sloan alum’s flight-simulator-style intervention, “saved millions of dollars and many many lives”); the Enroads climate-policy simulator (Climate Interactive collaboration; users can change peer-reviewed assumptions); applications named in passing across high-tech, automotive, pharma, drug development, construction, chemicals, healthcare, energy policy, and public health.

Defensive epistemic stance, quoting George Box: “All models are wrong; some models are useful.” The simulator’s value is not predictive accuracy but mental-model updating under safe conditions.

Facilitator stance — “sage on the stage” → “guide on the side”

A durable practitioner-reframing of the senior convener’s role: “We need to get away from the sage on the stage. The role of the facilitator is to be a guide on the side to catalyze the learning for the group.” Sterman applies it to himself: “I’m not the expert. I’m here to help catalyze your learning.” The implication for in-house systems-thinking interventions: the senior person convening the group-modeling exercise must not perform expertise; otherwise the group surfaces less, the system stays outside the room, and the causal map collapses back to the senior’s prior mental model — exactly the failure mode the intervention was supposed to fix.

Lineage extension to industrial AI (Carrier 2026)

Carrier 2026 (Senior Lecturer in System Dynamics, MIT Sloan) is the wiki’s third source in the MIT-system-dynamics lineage and extends the discipline to industrial AI agent design. Carrier’s load-bearing methodological claim:

“A key insight from Jay Forrester, who created the system dynamics group here at MIT — there were some missing feedback loops here. And that AI agent was designed to build a shorter, faster corrective feedback loop. That’s why the value was learned. So it’s not simply about building AI agents. It’s about using them to replace long, slow feedback loops with very fast ones.”

Worked case: a relatively simple AI agent at Heineken Mexico compressed a 6-hour changeover (containing only 15 minutes of actual information content) to 15 minutes — “a million extra cases of beer per month.”

The methodological add to Sterman 2026: where Sterman delivers feedback discipline at the general-management scale, Carrier delivers it at the industrial-operations-with-AI-agents scale. The discipline is the same; the deployment substrate is different.

Carrier also supplies the wiki’s strongest counterweight to the get-more-data-always tendency — his refinery alarm-fatigue tragedy (two fatalities; “the system was completely overloaded with safety alarms so that the information flow actually ground to a halt”) reframes the data-architecture problem as one of leading indicators in a small set beat data-lakes-of-everything, the system-in-the-room discipline at the data layer.

The engineering-team operationalisation (Forsgren & Macvean 2026)

The Google I/O 2026 talk by Forsgren and Macvean (Google’s Developer Intelligence team) closes on a one-line operational tricolon that names systems thinking as the highest-altitude capability AI-era engineers need:

“We need to shift left. We need to shift up. And we need to think about designing systems, not just bits of code.”

The talk’s structural claim: as AI takes on more execution, the human engineering work moves up the abstraction stack into system-design altitude. Three mechanisms in the talk align with the discipline’s Sterman 2026 / Carrier 2026 lineage at the engineering-team scale:

• Feedback loops as the load-bearing primitive. The talk repeatedly invokes feedback loops as the new bottleneck — “effective verification becomes the bottleneck” + “this is about setting effective feedback loops” + the agent-journaling reflection loop. Direct extension of Carrier’s “replace long, slow feedback loops with very fast ones” claim from industrial AI to fleet-scale agent orchestration.
• AI as amplifier and mirror (DORA framing). “AI is an amplifier and it is a mirror. It magnifies the existing strengths and it holds up a mirror to those weaknesses.” Systems-thinking corollary: AI doesn’t create system dysfunction — it exposes and accelerates dysfunction already present in the surrounding system (data, tooling, processes, culture). This is the engineering-team articulation of Sterman’s “there’s no such thing as a side effect” claim — the system as a whole responds to AI introduction, not just the task AI was deployed against.
• Designing environments, not vibe-coding. “What they’re doing is designing environments, setting the guardrails, creating the systems so that agents and humans can work together toward a goal and a broader purpose.” Systems-thinking applied to agent orchestration: the unit of design becomes the environment-with-agents-in-it, not the individual agent’s behaviour. Convergent with Kokane 2026’s higher-altitude systems-design framing of harness engineering.

The convergence point: four sources now (Bansal-Birkinshaw / Sterman / Carrier / Forsgren-Macvean) span the strategic-innovation, foundational-decision-making, industrial-AI, and engineering-team scales of the same underlying discipline. Forsgren-Macvean adds the engineering-leadership vantage the prior three did not address directly, and grounds the discipline in DORA’s published research programme.

Examples cited in the wiki

• Maple Leaf Foods — repositioning meat processor → “sustainable protein company” → tens of millions in new value, partnership with Meat Institute on Protein PACT.
• University of Guelph regenerative agriculture program — climate change reframed as soil health.
• Co-operators (Canadian insurance) — “drying in place” + “soft contents” cleaning to disrupt the flow of damaged materials to landfills.
• CSA Group circular built environment — UN Environment Programme cites ~37% of global carbon emissions from built environment, ~38% reducible through circular design.

Debates and supersession

• Bansal & Birkinshaw vs Sterman on what counts as the discipline. Bansal & Birkinshaw 2025 treat systems thinking as one of three innovation modes (alongside breakthrough and design thinking), positioned for wicked problems. Sterman 2026, from inside the MIT system-dynamics lineage, treats systems thinking as the foundational mental-model discipline for management decision-making at any scale where feedback loops are present — i.e. essentially everywhere. The wiki holds both framings as compatible: Bansal & Birkinshaw is the strategic-innovation framing; Sterman is the operational-decision-making framing.
• “All models are wrong; some are useful” (George Box, per Sterman 2026). The defensive epistemic stance built into the discipline is itself a concession that systems-thinking outputs are useful approximations, not predictive truth. The wiki should treat any systems-thinking conclusion as load-bearing only insofar as it is supplemented by group-modeling diversity (“system in the room”) — single-modeler causal maps are at risk of collapsing back to the modeller’s prior mental model (the failure mode Sterman’s “guide on the side” facilitator stance is designed to prevent).
• Open contradiction with the AI-tooling trajectory. AI tooling excels at the kinds of expert tells you the answer outputs that Sterman 2026’s pedagogy explicitly rejects (“lectures don’t work; research shows that showing people research doesn’t work”). The wiki has not yet ingested a source that resolves whether AI tools can be reshaped into the let-people-experience-consequences shape of a management flight simulator, or whether they fundamentally produce the wrong genre of output for mental-model update. This is a substantive open contradiction, not a vocabulary mismatch.
• No supersession events. Sterman extends Bansal & Birkinshaw; neither retires.

Related concepts

• enterprise-ai-adoption — AI deployment decisions are often systems-thinking problems (knock-on effects across users, communities, supply chains).
• automation-vs-augmentation — strategic deployment choices have systemic consequences beyond the deploying firm.
• dynamic-capabilities — systems thinking informs sensing/seizing/transforming under interconnected change.
• strategic-foresight — both approaches treat the firm as embedded in a larger interacting system.
• ai-deskilling — task-composition shifts within retained jobs are a systems-level effect of AI adoption.

Open questions

• Single primary source in the wiki so far; deeper Senge and Forrester texts would strengthen the concept. Partially closed (2026-05-18) by Sterman 2026 — the Jay W. Forrester Professor of Management at MIT Sloan delivering the in-house MIT-system-dynamics treatment in 58 minutes. Still-open subitem: a primary-source Peter Senge text (The Fifth Discipline itself) is not yet in the wiki.
• How does systems thinking interact with AI tooling? Specifically: AI systems excel at slicing the Gordian knot (breakthrough mode) and at user-centric iteration (design mode); whether they can support genuinely systems-level analysis is open. Sterman’s “lectures don’t work” + “mental models are reinforced by everyday experience” claim suggests a sharper sub-question: can AI tools update mental models (in the safe-experimentation way a management flight simulator can), or do they just deliver more experts-telling-you-the-answer output — which Sterman’s research says doesn’t change behaviour?
• The qualitative-vs-quantitative split inside the toolkit (causal mapping vs full simulation) is treated by Sterman as a continuum chosen by problem size. The wiki’s empirical anchor for when each is sufficient is still thin.