organizational-frameworks-for-ai-adoption

Synthesis: Organizational frameworks for AI adoption

Confidence 0.90 · 14 sources · last confirmed 2026-06-19

Closes organizational-frameworks-for-ai-adoption. Originally filed 2026-05-05 with 6 frameworks (MIT CISR, Anand-Wu, Cisco, Werner-Le-Brun, OpenAI, McKinsey Rewired 2nd ed).

Refreshed 2026-05-08: four additional framework-vantages have been ingested in the intervening three days, expanding the cluster from 6 frameworks to 10 frameworks across 12 sources:

• Firm-boundary lens (Nishar & Nohria 2026) — Build / Compose / Collaborate / Buy Outcomes.
• Human-reaction lens (Carucci 2026) — resistance as data; orthogonal cross-cutting lens applied to any of the other framework decisions.
• Organizational-learning lens (Ransbotham et al. 2024 + Kiron & Schrage 2026) — Augmented Learners 2×2 + verification → evaluation → learning capture flywheel. Now the largest representative-sample dataset in the cluster (N=3,467, 21+ industries, 136 countries; 8-year longitudinal panel).
• Runtime-engineering lens (Kokane 2026 + Chatterjee 2026) — agent harness anatomy (Context / Constraints / Contracts / Compounding); the engineering-stack counterpart to the organizational-learning flywheel.

The 6→10 expansion does not change the synthesis’s core thesis (frameworks operate on different layers; pick by the executive’s question, not the framework’s brand). It strengthens the empirical foundation, sharpens several disagreements, and surfaces a new structural finding: the “compounding cycle” is now visible as the same operational mechanism at two different stack layers (organizational learning per Kiron-Schrage, runtime engineering per Chatterjee), each independently reached by different practitioner communities.

Refreshed 2026-06-19: no new framework lands, but the cluster gains its first lived multi-framework operator case — DBS Bank 2026. DBS is not an eleventh-plus lens; it is a single 39,000-employee incumbent running several of the cluster’s layers at once over a decade — MIT CISR Stage-4-style embedded AI, McKinsey-Rewired-style business-led + KPI-cascaded operating-model change, the Bain/OpenAI trap-escape (workflow-not-task redesign), and Beutler-style business-owns-the-outcomes governance. It is therefore the closest thing the wiki has to a partial answer to open question #4 (“do the frameworks converge when applied to a specific firm?“): DBS’s mechanisms map cleanly onto multiple frameworks without contradiction, weak evidence for the complementary-not-competing thesis. See the open-questions section.

Refreshed 2026-06-06: an 11th framework lands — the deployment-maturity ladder from OpenAI 2026 (the wiki’s first first-party OpenAI source): Ask → Assist → Automate (read-only → human-in-the-loop → full autonomy, exceptions routed to humans). It occupies a layer none of the prior ten named cleanly: given that you are deploying AI on a workflow, how much agency do you grant it, and in what sequence? — distinct from Anand-Wu’s task-deployment layer (should you deploy on this task at all) and from Allen/AWS’s economic-decision USE/COMPOSE/BUILD (build vs rent the model). Beutler pairs it with two org-design rules that reinforce existing cluster findings rather than adding a layer: separate governance from transformation, business owns the outcomes (sharpening McKinsey Rewired’s operating-model capability) and embed engineering inside the business function (the cross-customer-deployment twin of Fung’s inside-Anthropic team-shape rewrite). The 10→11 expansion again leaves the core thesis intact and adds a deployment-maturity rung beneath the task-deployment layer.

Question

When advising an executive on enterprise AI adoption, which framework should inform which decision — and where do they genuinely disagree rather than say the same thing in different vocabulary?

Findings

The frameworks operate on different layers

The ten frameworks are mostly complementary, not competing. Each names a different layer of AI-adoption decision-making, with overlap rather than conflict. Stack them top-down (organizational → engineering), with one orthogonal cross-cutting lens that applies to any layer:

Layer	Framework	Question it answers	Source
Org-design	Tin Man / Octopus	Is your organization structurally capable of adapting?	Werner-Le-Brun
Readiness	Cisco 5 Foundations (Strategy / Infrastructure / Data / Governance / Culture)	Do you have the prerequisites?	Cisco
Capability progression	MIT CISR Four Stages + Four S	What stage are you at, and what blocks the next transition?	MIT CISR
Organizational learning (new)	Ransbotham/Kiron Augmented Learners 2×2 + verification → evaluation → learning capture flywheel	Does your org metabolize AI interactions into compounding knowledge, or consume them and forget?	Ransbotham et al. 2024, Kiron & Schrage 2026
Transformation playbook	McKinsey Rewired 6 capabilities	How do you actually execute the journey?	Lamarre, Smaje, Levin et al. 2026
Firm-boundary (new)	Nishar-Nohria 4-model (Build / Compose / Collaborate / Buy Outcomes)	Which workflows should the firm own, and which should it rent / outsource / outcome-source?	Nishar & Nohria 2026
Trap escape	Bain/OpenAI 4-step transformation	How do you avoid the micro-productivity-trap?	Dutt, Chatterji et al. 2026
Task deployment	Anand-Wu 2×2 (cost of errors × type of knowledge)	For this specific task, deploy AI today and how?	Anand-Wu
Deployment maturity (new)	Beutler Ask → Assist → Automate ladder	Given you’re deploying on a workflow, how much agency do you grant the agent, and in what sequence?	OpenAI 2026
Runtime engineering (new)	Agent harness (Context / Constraints / Contracts / Compounding); 4-layer stack with swappable model layer	What runtime infrastructure must exist between your model and your user?	Chatterjee 2026, Kokane 2026
Diagnostic	Anand-Wu 6 leakage points	If gains aren’t showing up in P&L, where are they leaking?	Anand-Wu
⊥ Cross-cutting: human reaction (new)	Carucci resistance-as-data: 3 leader traps (personalize / moralize / rush) + 4 signal categories (Loss / Anxiety / Lack of control / Flaws in change)	When pushback arrives at any layer above, what is it telling you about the change itself?	Carucci 2026

The orthogonal lens. Carucci’s human-reaction lens is not a stack layer. It is a diagnostic posture applied when any of the other 9 layers triggers organizational pushback. A reorganization (Werner-Le-Brun’s Octopus transition) generates resistance; a maturity-stage transition (MIT CISR Stage 2 → 3) generates resistance; a firm-boundary shift (Nishar-Nohria Compose → Collaborate) generates resistance; a workflow redesign (Bain/OpenAI) generates resistance. Carucci’s framework is what you reach for while running any of the other nine.

A decision tree for which framework to use

Sequenced by the typical executive question (top-of-house → engineering-floor):

1. “Are we organizationally capable of changing?” → Werner-Le-Brun (Tin Man vs Octopus). If you’re a Tin Man, lower frameworks won’t stick.
2. “Do we have the foundations?” → Cisco 5 Foundations. Infrastructure, data, governance must clear a bar before pilots scale.
3. “Where are we in the maturity curve?” → MIT CISR Four Stages. Locate your stage; the Four S diagnostic identifies what blocks the next transition. The financial inflection sits specifically at Stage 2 → Stage 3.
4. “How does our org metabolize AI interactions?” → Ransbotham/Kiron Augmented Learners 2×2. Score yourself on org-learning capability and AI-specific-learning capability; the 15% Augmented-Learner cohort outperforms Limited Learners 1.6× on uncertainty management and 1.4× on revenue benefits. Then layer Kiron-Schrage’s verification → evaluation → learning capture flywheel as the operational mechanism that turns the 15% advantage into compounding returns.
5. “What’s our transformation playbook?” → McKinsey Rewired 6 capabilities (business-led roadmap, talent, operating model, technology, data, adoption-and-scaling). Anchor: ~20% EBITDA uplift, $3:$1 ROI, 1–2yr breakeven across deep-dive AI-leader companies.
6. “Where does the boundary of our firm sit?” → Nishar-Nohria 4-model (Build / Compose / Collaborate / Buy Outcomes). Decide per workflow, not at the firm level. The decision is no longer cost-driven — it is strategic about where differentiation lives. Empirical anchor: enterprise GenAI app spending $1.7B (2023) → $37B (2025); SaaS valuations 30–60% below 2021 peaks.
7. “How do we avoid the common failure mode?” → Bain/OpenAI 4-step (narrow possibilities; reimagine workflows; engage those closest to the process; measure outcomes plus evals). Names the trap (offering lock-in / process lock-in) explicitly. Anchor: 10–25% EBITDA gains.
8. “Which task do we point AI at first?” → Anand-Wu 2×2. No regrets zone (low cost of errors / explicit data) is where to start; agents thrive there today.
9. “What runtime infrastructure must we build?” → Agent harness (Chatterjee’s Context / Constraints / Contracts / Compounding; Kokane’s 4-layer stack with swappable model layer). Build constraints before cleverness. The model is rented and replaceable; the harness is owned and compounds. Caveat (Kokane): ~90% of the harness work is mature systems engineering applied to a new substrate — hire systems engineers, not AI specialists; the 10% genuinely novel is non-determinism at the execution layer + context as a degrading resource.
10. “Why isn’t our investment paying off in P&L?” → Anand-Wu 6 leakage points. Diagnostic, not prescriptive. Trace which of the 6 points is leaking.

Cross-cutting (apply at any layer above):

• “How do we read the pushback we’re getting?” → Carucci resistance-as-data. Avoid the three traps (personalize / moralize / rush to resolution). Diagnose the four signal categories (Loss / Anxiety / Lack of control / Flaws in change). Resistance is rarely noise; it is feedback on the change itself.

Where they genuinely disagree

Most apparent disagreements are vocabulary differences, but a few real splits:

• Stage 2 → Stage 3 financial inflection (MIT CISR-specific). MIT CISR claims firms in Stages 1–2 underperform their industry; Stages 3–4 outperform. The other frameworks don’t make this stage-gradient claim — they describe the destination of transformation but not where the inflection lives. If true, the inflection is the most actionable framing.
• “Reinvent the business” vs. incremental capability building. McKinsey Rewired and Bain/OpenAI both push aggressive posture: process redesign first, technology second; business-led, not IT-led; “we don’t have a single success story where senior business leaders weren’t in the driver’s seat.” MIT CISR is more incremental — the Four Stages permit gradual progression. The Bain/OpenAI thesis explicitly contradicts the incremental approach by naming the micro-productivity-trap (offering lock-in, process lock-in).
• Org-design as upstream prerequisite (Werner-Le-Brun) vs. downstream outcome. Tin Man / Octopus treats org structure as the precondition for AI adoption to work. The other frameworks largely assume the org can act, then prescribe what to do. If Werner-Le-Brun is right, the lower frameworks may not stick on a Tin Man without org-design work first.
• Pilots as a stage (MIT CISR) vs. pilots as a trap (Bain/OpenAI). MIT CISR Stage 2 (“Build pilots and capabilities”) frames piloting as a needed phase. Bain/OpenAI argue most companies get stuck in pilots — exactly because they don’t reimagine workflows. Same observation, opposite valence.
• Firm-boundary as variable (Nishar-Nohria) vs. organizational property (Werner-Le-Brun). Nishar-Nohria explicitly argue “the boundary of the firm becomes a variable rather than a given” — workflows can move fluidly between Build / Compose / Collaborate / Buy Outcomes as the cost curve of custom software collapses. Werner-Le-Brun frame organizational structure as a deeper property requiring multi-year transformation. Both can be true at different time-horizons: Nishar-Nohria’s per-workflow boundary fluidity is a near-term commercial decision; Werner-Le-Brun’s org adaptive capacity is the multi-year underlying variable that determines whether the firm can act on boundary fluidity.
• “Process redesign is the value” (Bain/OpenAI, McKinsey) vs. “the harness is the moat” (Chatterjee). Bain/OpenAI claim “process redesign is the most challenging part of AI deployment and creates most of the value” — i.e., organizational/workflow work dominates. Chatterjee claims “the harness is the only part of our stack that gets more valuable with every customer we ship to” — i.e., runtime engineering work dominates. Resolution: they sit at different stack layers. Both are required; Bain/OpenAI is right at the org-process layer; Chatterjee is right at the runtime layer. The synthesis-level claim is that both compounding cycles must be running — the organizational verify-evaluate-capture loop and the runtime telemetry → harness adjustment → workspace overrides loop.
• Kokane’s sceptical vantage on the entire cluster. Kokane argues ~90% of “agent harness” engineering is mature systems engineering applied to a new substrate. Generalised: this cuts against the breathless framings implicit in some consulting frameworks (everything-must-be-rebuilt-for-AI). The wiki holds Kokane’s vantage as a useful epistemic discipline — when a framework claims AI is uniquely transformational, ask whether the underlying work is genuinely new or rebranded.
• Augmented Learners (Ransbotham/Kiron) as upstream of MIT CISR Stage transition. Ransbotham/Kiron’s 5-question org-learning battery + 4-question AI-specific-learning battery measure capabilities that map onto MIT CISR Stage 3 attributes (test-and-learn, architected for reuse, human-feedback-loop AI). A Stage-2 firm with strong organizational learning culture is on the Augmented-Learner trajectory; a Stage-2 firm without one will stall. This sharpens MIT CISR’s Stage 2 → 3 inflection by naming the upstream variable that determines whether the transition happens.

Cross-framework agreement

All ten frameworks converge on a few claims, with different vocabularies. The post-refresh evidence is substantially stronger:

• Adoption breadth ≠ transformation depth. AI Index pegs maturity at 1%; MIT CISR pegs Stage 4 at 7%; Cisco pegs ready at 13%; Ransbotham/Kiron peg Augmented Learners at 15%; Werner-Le-Brun’s 12% transformation-success baseline. Same diagnostic, five numbers — all clustering 7–15%. When five independent measurement instruments arrive at a similar upper-tail size, the cohort is structurally narrow, not measurement-artefactual. See ai-maturity-measurement-comparison.

• Process redesign is the main cost, not technology. Bain/OpenAI: “process redesign is the most challenging part of AI deployment and creates most of the value.” McKinsey Rewired: “business-led roadmap” and “operating model” are 2 of the 6 capabilities. MIT CISR Four S “Synchronization” is reskilling and role redesign. Anand-Wu 2×2’s strategic value comes from picking the right tasks, not the technology. Now corroborated by 7 sources (the original 6 frameworks plus Nishar-Nohria explicitly: “data architecture, governance, and ownership as integral to the transformation rather than afterthoughts”) — among the strongest cross-source agreements in the wiki.

• The “compounding cycle” is the same operational mechanism at two stack layers. Two practitioner essays from independent vantages (Kiron-Schrage at the organizational-learning level; Chatterjee at the runtime-engineering level) describe operationally identical machinery:

  Kiron-Schrage step (organizational)	Chatterjee equivalent (runtime)
  Verification — does this output meet the standard?	Constraints layer — pre/post-tool hooks score outputs
  Evaluation — what does this output reveal?	Contracts layer — formal evaluable specifications
  Learning capture — how do we ensure this insight persists?	Compounding layer — telemetry → harness adjustments → workspace overrides

  This is the strongest concept-level convergence in the cluster — same mechanism, two layers, two practitioner communities, independently reached. The synthesis-level implication: AI products that compound require both cycles running (organizational + runtime); a firm with one but not the other is partially stuck.

• Senior-leader ownership is non-negotiable. McKinsey: “We don’t have a single success story where senior business leaders were not in the driver’s seat.” Bain/OpenAI: “boardroom imperative.” MIT CISR: united front of CEO / CIO / CSO / Head of HR. Cisco: “Strategy” pillar means top-down. Werner-Le-Brun: org-design ownership at C-level. Refined by Carucci 2026: senior leadership is necessary but not sufficient — leaders’ three traps (personalize / moralize / rush to resolution) systematically suppress the operator-level signal that would otherwise correct course. Senior-leader ownership + resistance-as-data discipline together are what works.

• Talent density shifts over hiring breadth. McKinsey Rewired’s 70% benchmarks (70%+ in-house, 70%+ “doer” engineers, 70%+ at competent-or-expert skill levels) — distinctive but consistent in spirit with MIT CISR’s “Synchronization” and Cisco’s “Culture and Talent.” Sharpened by Chatterjee 2026 at the engineering layer: “the most under-resourced role on most AI teams is the engineer who specializes in [harness work]… someone who thinks about agents the way SREs think about distributed systems.” The under-hire problem is specific, not general.

• Most enterprise tasks today are augmentative, not automative. Anand-Wu’s “no regrets zone” (where agents thrive) is the minority of business tasks. MITTRI/Cisco’s chatbot → agent → multi-agent progression places multi-agent as future-state. The empirical record (AI Index 2026) confirms: agent deployment in single digits across nearly all business functions. Worker-attitude data corroborates (Ransbotham et al. 2024): 84% hopeful that AI will assist with their tasks (vs. 70% in 2017); only 20% fearful (vs. 31% in 2017). The displacement panic that surged with ChatGPT has not been borne out at the worker-attitude level — at least within the 3,467-respondent MIT SMR × BCG sample.

• Models commoditize; what’s owned compounds. Chatterjee 2026 makes this explicit at the engineering layer (“the model is rented from a vendor whose competitor will outperform them within the year”). Nishar-Nohria 2026 makes it explicit at the firm-boundary layer (the Build model produces hard-to-replicate institutional knowledge over time as it encodes the firm’s data and decision logic). Kiron-Schrage 2026 makes it explicit at the organizational level (compounding economics = asset appreciation; consumption economics = asset depreciation). Three independent vantages, same conclusion: durable advantage in AI comes from what the firm owns and accumulates, not from which model it rents.

Where empirical validation sits — and where it is still missing

Of the ten frameworks:

Framework	Empirical anchor	Sample / instrument
MIT CISR Four Stages	28%/34%/31%/7% distribution; financial inflection at Stage 2 → 3	2022 Future Ready Survey, N=721 — pre-GenAI baseline, now 4 years old
Ransbotham/Kiron Augmented Learners	15% / 14% / 12% / 59% 2×2 distribution; 1.6× / 2.2× / 1.8× / 1.6× uncertainty-management multipliers; 99% revenue-benefit saturation in Augmented Learner cohort	Spring 2024 MIT SMR × BCG global survey, N=3,467, 21+ industries, 136 countries; 8th annual wave of comparable methodology — the largest representative-sample dataset in the cluster
Anand-Wu 2×2	Conceptual framework; cited prior tech cycles (e-ticketing, CAD/ERP, Big Law)	No per-quadrant empirical baseline
Cisco 5 Foundations	13% AI-ready / 98% urgency / 85% give themselves <18 months	Cisco’s own 2025 readiness survey; vendor-instrument data, methodology not fully disclosed
Werner-Le-Brun Octopus	12% transformation-success baseline	Case examples (Netflix, Google, Coca-Cola, U.S. Army); no distributional data; AWS-affiliated authorship
McKinsey Rewired	~20% EBITDA uplift; $3:$1 ROI; 1–2yr breakeven; 70% talent-density shifts	~20 deep-dive AI-leader companies (selected from ~200 study set); vendor-of-deployment data — selection-effect-aware
Nishar-Nohria 4-model	Enterprise GenAI app spending $1.7B (2023) → $37B (2025) ≈ 22×; SaaS valuations 30–60% below 2021 peaks; >1/3 of companies replaced ≥1 SaaS tool with custom GenAI	Industry analyst data + Adobe outcome-based pricing as named industry signal — third-party aggregate, not framework-specific survey
Bain/OpenAI 4-step	10–25% EBITDA gains; Lowe’s Mylow 2× online conversion; FabricationCo $30M profit	Bain client work; vendor-of-deployment data — selection-effect-aware
Agent harness (Chatterjee + Kokane)	Friday-in-March anecdote (n=1); architectural-pattern alignment with Anthropic Managed Agents	Practitioner essays; no empirical anchor for harness-tuning compounding rates
Carucci resistance-as-data	None named in the source	Practitioner essay; no empirical anchor at all — drawn from 30 years of consulting practice

The empirical landscape has shifted substantially since the original synthesis closed. Three observations worth flagging:

1. MIT CISR is no longer alone with N>500 representative-sample data. Ransbotham/Kiron’s 3,467-respondent cross-industry global survey is now the largest dataset in the cluster, and it is post-GenAI (spring 2024) where MIT CISR is pre-GenAI (2022). The MIT CISR vs. Ransbotham/Kiron methodology cross-walk is now a higher-priority open question than it was in May.
2. The 10–25% EBITDA range from consulting frameworks (McKinsey Rewired, Bain/OpenAI) plus the 22× enterprise GenAI app spending growth (Nishar-Nohria) cluster on roughly the same magnitude of value claim — different metrics, similar order of magnitude. Independent corroboration from non-vendor sources would still be valuable but the cluster is no longer a single sponsored data point.
3. The new sources tilt the cluster toward the prescriptive side, not the distributional side. Carucci, Chatterjee, Kokane, Kiron-Schrage all add prescriptive frameworks (resistance discipline, harness anatomy, “compounding cycle”); only Ransbotham/Kiron and Nishar-Nohria add empirical anchors. The wiki now has more frameworks than it did at original close, but the empirical-vs-prescriptive ratio has not improved proportionally.

Sources consulted

Original 6 (filed 2026-05-05):

• 2026-04-28-mit-sloan-ai-maturity — Burnham (2025) reporting on Woerner, Weill, Sebastian, Káganer at MIT CISR. Four Stages + Four S; representative-sample distribution.
• 2026-04-28-anand-wu-genai-playbook — Anand & Wu (HBR Nov–Dec 2025). 2×2 task framework + 6 leakage points; paradox-of-access argument.
• 2026-04-28-mittri-cisco-ai-enabled-enterprise — MIT Tech Review Insights × Cisco (2025). 5 Foundations + chatbot → agent → multi-agent progression.
• 2026-04-28-werner-lebrun-octopus-organization — Werner & Le-Brun (HBR Nov–Dec 2025). Tin Man / Octopus archetypes; AWS Executives in Residence.
• 2026-05-02-dutt-chatterji-ai-experimentation-to-transformation — Bain & Company + OpenAI Economic Research (HBR.org, 30 Apr 2026). 4-step transformation framework; names the micro-productivity-trap.
• 2026-05-03-rewired-second-edition-sample — Lamarre, Smaje, Levin et al. (Wiley/McKinsey & Company 2026, sample-only). 6-capability framework; 70% talent-density shifts; 20% EBITDA uplift.

Added in 2026-05-08 refresh:

• 2026-05-05-nishar-nohria-end-of-one-size-fits-all — Deep Nishar & Nitin Nohria (HBR.org Digital, Apr 2026). Firm-boundary lens: Build / Compose / Collaborate / Buy Outcomes. Enterprise GenAI app spending 22× in 2 years; SaaS valuations 30–60% below 2021 peaks; Adobe outcome-based pricing as named industry signal.
• 2026-05-07-carucci-resistance-as-data — Ron Carucci (HBR.org Digital, Apr 2026). Human-reaction (cross-cutting) lens: 3 leader traps (personalize / moralize / rush to resolution) + 4 signal categories (Loss / Anxiety / Lack of control / Flaws in change). Practitioner essay; no empirical anchor.
• 2026-05-07-ransbotham-augmented-learners — Ransbotham, David Kiron, Khodabandeh, Chu, Zhukov (MIT SMR × BCG Big Ideas Research Report, 8th annual, Nov 2024). Organizational-learning lens (empirical foundation): Augmented Learners 2×2 distribution; 1.6× uncertainty management; 1.4× revenue benefits. Largest representative-sample dataset in the cluster (N=3,467).
• 2026-05-07-kiron-schrage-compound-benefits — David Kiron & Schrage (MIT SMR Column, Apr 2026). Organizational-learning lens (operational mechanism): verification → evaluation → learning capture flywheel; reframes ROI as “return on iteration”; consumption (asset depreciation) vs. compounding (asset appreciation) economics.
• 2026-05-07-kokane-agent-harness-vs-systems-design — Kokane (Level Up Coding / Medium, Apr 2026). Runtime-engineering lens (sceptical vantage): ~90% of agent-harness work is mature systems engineering rebranded; 10% genuinely novel (non-determinism + context-as-degrading-resource).
• 2026-05-07-chatterjee-anatomy-of-agent-harness — Chatterjee (Medium, May 2026). Runtime-engineering lens (taxonomical vantage): Context / Constraints / Contracts / Compounding. Friday-in-March worked failure example; “the model is what you rent. The harness is what you own.”

Added in 2026-06-06 refresh:

• 2026-06-02-architecting-ai-native-organizations-redesign-work-at-scale-joe-beutler — Joe Beutler (Head of Solutions Engineering, Strategics, OpenAI), at IT Revolution’s Enterprise AI Summit (talk published 2 June 2026). Deployment-maturity lens (the wiki’s first first-party OpenAI source): Ask → Assist → Automate ladder; separate governance from transformation, business owns the outcomes; embed engineering inside the business function. Vendor-of-deployment cross-customer vantage; anecdotal anchors (T-Mobile $3B / 60%-automated call center; PwC 20%-finance-team benchmark).

Added in 2026-06-19 refresh:

• 2026-06-18-dumra-mit-smr-dbs-everyone-an-innovator — Bidyut Dumra (Group Head of Innovation and Future of Work, DBS Bank) on MIT SMR’s Leaders at All Levels (Ep. 9). Lived multi-framework operator case — not a new lens, but a single banking incumbent running several cluster layers at once: GANDALF sensing, the Innovation Pyramid + QPR (seizing), Managing Through Journeys (transforming), and a 20%-of-scorecard innovation KPI that operationalises the trap-escape. The cluster’s first decade-long operator-altitude instantiation; a partial worked-example answer to open question #4.

Lessons

• Pick a framework by the executive’s question, not the framework’s brand. Each framework names a different layer; mismatched layers explain most “why didn’t this work” stories. The decision tree above (now 10 questions plus 1 cross-cutting) is the operational artifact.
• The transformation-mindset frameworks (McKinsey Rewired, Bain/OpenAI) and the maturity-progression framework (MIT CISR) sit in tension on pilot stages. Rewired and Bain treat pilots-as-trap; MIT CISR treats pilots-as-stage. Reconcile by treating MIT CISR’s Stage 2 as a phase to pass through quickly, not dwell in.
• Process redesign is the load-bearing decision across all 10 frameworks. If your AI program treats redesign as something the technology team handles, it will land on the wrong side of the micro-productivity-trap regardless of which framework you cite. Now corroborated by 7 sources.
• Organizational learning capability is the upstream variable for MIT CISR Stage 2 → 3 transition. Ransbotham/Kiron’s Augmented Learner traits (test-and-learn, failure tolerance, postmortem culture, learning codification, AI-specific learning) map directly onto Stage 3 attributes. A Stage-2 firm with weak organizational learning culture will stall at Stage 2. This sharpens the most actionable claim in the original synthesis (the Stage 2 → 3 inflection) by naming what determines whether the transition happens.
• The “compounding cycle” is the same operational mechanism at two stack layers. Kiron-Schrage at the organizational level (verification → evaluation → learning capture) and Chatterjee at the runtime level (Context / Constraints / Contracts / Compounding) describe operationally identical machinery. AI products that compound require both cycles running. A firm with strong organizational learning but weak harness-runtime engineering will struggle to ship; a firm with strong harness engineering but weak organizational learning will ship products that don’t get better with use.
• The human-reaction lens is cross-cutting, not stack-layered. When applying any of the other 9 frameworks, watch for Carucci’s four resistance signals (Loss / Anxiety / Lack of control / Flaws in change). The fourth — Flaws in change — is operator-level signal that the framework’s prescription is wrong for this particular firm. Leaders’ three traps (personalize / moralize / rush) systematically suppress it. Resistance discipline is what keeps the other 9 frameworks honest.
• Models commoditize; what’s owned compounds. Three independent vantages reach this conclusion: Chatterjee at the engineering layer (“the model is what you rent. The harness is what you own.”); Nishar-Nohria at the firm-boundary layer (Build encodes hard-to-replicate institutional knowledge over time); Kiron-Schrage at the organizational level (compounding economics = asset appreciation). The competitive-advantage question shifts from which model do we choose? to what do we own and accumulate?
• Empirical maturity distributions older than 2024 are increasingly stale, but the cluster is no longer single-anchor. The MIT CISR 2022 baseline (N=721) is now joined by the Ransbotham/Kiron 2024 baseline (N=3,467, post-GenAI) — the larger and more recent dataset. Cross-walking the two is now feasible; whether they agree is an open question worth investigating.

Open questions

Resolved (or partially resolved) by the 2026-05-08 refresh:

• “No empirical validation against a representative sample beyond MIT CISR.” — partially resolved: Ransbotham/Kiron’s 3,467-respondent cross-industry global survey now sits alongside MIT CISR as a second large-N anchor. Still BCG-sponsored research, so the vendor-of-deployment caveat applies; but the methodology is longitudinal (8th annual wave) and representative.
• “MIT CISR Stage distribution post-GenAI.” — still open as a methodology question, but Ransbotham/Kiron’s post-GenAI 2×2 distribution (15% / 14% / 12% / 59%) gives an indirect signal: the upper-tail cohort size (15% Augmented Learners in 2024) is consistent with MIT CISR’s 7% Stage 4 (2022) plus some upward drift, broadly compatible with the original framework if interpreted generously. A direct MIT CISR 2025/2026 wave would still be highly valuable.

Still open:

• Cross-walk between MIT CISR Stages and Ransbotham/Kiron Augmented Learners 2×2. The hypothesis (above, in the new disagreements section) is that organizational learning capability is the upstream variable for the Stage 2 → 3 transition. A worked-example test on a panel of firms tracked through both instruments would settle this.
• How does the firm-boundary decision (Nishar-Nohria) interact with maturity stage (MIT CISR)? Is “Buy Outcomes” only viable from Stage 3+, or can a Stage 1–2 firm leapfrog by outsourcing the function entirely? This is the wiki’s strongest practitioner-relevant unresolved question.
• Quantitative measurement of the compounding cycle. Kiron-Schrage describe it qualitatively; Chatterjee’s Compounding layer describes it qualitatively. Nobody has measured the slope of compounding — how much contract-score uplift per unit of telemetry-driven harness adjustment, or how much Augmented-Learner advantage growth per quarter. Without measurement, the “compounding cycle” is a plausible-sounding reframing rather than a quantified claim.
• Whether the 10 frameworks converge on the same recommendation when applied to a specific firm. The decision-tree above is still a hypothesis. A worked example — pick a real firm, run all 10 frameworks against it, compare the prescriptions — would test the “complementary not competing” claim. With 10 frameworks now, this is a substantial undertaking, but the test would be definitive. Partial evidence now exists: DBS Bank is a single incumbent whose mechanisms map onto MIT CISR (Stage-4 embedded AI), McKinsey Rewired (business-led + KPI cascade), Bain/OpenAI (workflow-not-task trap escape), and Beutler (business-owns-outcomes governance) without contradiction — weak confirmation of the complementary-not-competing thesis. It is one firm narrated by one executive, not an independent multi-framework audit, so it falls short of the definitive test; but it is the first lived case where the layers visibly coexist.
• Cross-walk with academic strategy theory. Carucci’s resistance-as-data lens has obvious resonance with change-management academic literature (Kotter, Lewin); none of the wiki’s existing concept pages directly bridge to it. A follow-up synthesis could integrate the cluster with dynamic-capabilities (Teece), systems-thinking, strategic-foresight, and change-management theory.
• The Kokane-Chatterjee tension (rebranded systems engineering vs. moat) in actual hiring data. Kokane prescribes hiring systems engineers; Chatterjee prescribes hiring SRE-like AI specialists. Both are plausible; nobody has compared the build-quality outcomes of teams hired on each model. A 2026–27 source measuring this would settle the most concrete operational disagreement in the cluster.

Related pages

• enterprise-ai-adoption — main concept page; the 10 frameworks all live here in summary form
• agent-harness — concept page for the runtime-engineering lens added in this refresh
• micro-productivity-trap — named failure mode that several frameworks target; now confidence 0.90, the most strongly corroborated concept in the wiki (7-source agreement across 5 vantages)
• ai-maturity-measurement-comparison — sister thread (still open) on the measurement side; the 1% / 7% / 12% / 13% / 15% maturity spread (Ransbotham/Kiron’s 15% Augmented Learners is now the 5th measurement)
• automation-vs-augmentation — load-bearing distinction inside Anand-Wu’s 2×2 quadrants
• dynamic-capabilities — academic-strategy bridge that none of the 10 frameworks engage with directly