Enterprise AI Adoption

Confidence 0.95 · 79 sources · last confirmed 2026-06-27

The pace, depth, and pattern by which organizations integrate AI into their business functions, processes, and products. The dominant 2024 signal is a step-change in adoption breadth, especially for generative-ai, paired with modest realized financial impact and very low maturity.

Working definition

Multiple complementary lenses are used by sources in this wiki:

Breadth lens (McKinsey & Company / AI Index): an organization is “using AI” if it deploys AI in at least one business function. Low bar — captures adoption breadth, not depth.

Stage lens ( MIT Sloan): organizations are placed on a four-stage AI maturity ladder — from “Experiment and prepare” to “Become AI future-ready” — based on a 0–100% Total AI Effectiveness score combining operations, customer experience, and ecosystem-support effectiveness.

Readiness lens (Cisco): only 13% of companies globally are ready to leverage AI to its full potential. Two-thirds (68%) say their infrastructure is at best moderately ready. 98% feel increased urgency over the past year, 85% give themselves <18 months to deploy a strategy.

Capabilities lens (Warner & Wäger 2019): adoption is the result of a system of digital sensing / digital seizing / digital transforming dynamic capabilities — nine microfoundations including digital scouting, scenario planning, mindset crafting, rapid prototyping, portfolio balancing, strategic agility, ecosystem navigation, structural redesign, and digital-maturity improvement. See dynamic-capabilities.

Foresight lens (Webb 2024 / FTSG 2026): adoption decisions sit downstream of strategic foresight — the disciplined process of signal detection → trend identification → scenarios → strategy. See strategic-foresight.

2026 numbers (AI Index 2026 Top Takeaway #1, #9):

• Organizational adoption: 88% (up from 78% in 2024).
• Generative AI population adoption: 53% within three years — faster than the PC or the internet.
• Productivity gains: 14–26% in customer support and software development; weaker or negative effects in tasks requiring more judgment.
• AI agent deployment remains in single digits across nearly all business functions — adoption breadth is high but agent-mediated workflow depth is still nascent.
• >80% of U.S. high school and college students use AI for school-related tasks; only half of middle/high schools have AI policies and just 6% of teachers say those policies are clear — adoption is outrunning policy.

Transformation lens (Dutt, Chatterji et al. 2026):

• The micro-productivity-trap — task-level gains failing to translate to firm-level results — is a named diagnostic for the gap between 88% organizational adoption and single-digit agent-mediated workflow depth.
• Two failure modes: offering lock-in (AI optimizes existing offerings without reframing value) and process lock-in (AI automates current processes without redesigning them).
• Empirical anchor: 10–25% EBITDA gains across Bain & Company clients adopting a transformation-mindset four-step framework (narrow possibilities strategically; reimagine workflows; engage those closest to the process; measure what matters with concrete business outcomes plus continuous evals).
• Worked cases: Lowe’s + OpenAI partnership (Mylow / Mylow Companion across 1,700+ stores; >2× online conversion; +200bps CSAT in-aisle); FabricationCo (~$30M additional profit on track; 15× faster quote generation; +10pp win rate in 3 months).

Six-capability lens — McKinsey “Rewired” (Lamarre, Smaje, Levin et al. 2026, 2nd ed):

• Definition: “the process of developing organizational and technology-based capabilities that allow a company to continuously improve its customer experience and lower its unit costs; and over time sustain a competitive advantage.”
• Six capabilities that companies must be strong across to win (Exhibit I.1):
  1. Business-led roadmap — top-down aspiration, alignment on economic leverage points, reimagination of business domains
  2. Talent — upskilled business leaders + density of engineering talent
  3. Operating model — business, tech, and operations closer together
  4. Technology — modern software engineering and platforms for reuse and time-to-value
  5. Data — unified, easy-to-consume data
  6. Adoption and scaling — change management; end-to-end process reconfiguration; impact measured in business KPIs
• Empirical anchor (~20 AI-leader deep-dive companies): 20% EBITDA uplift on average, 1–2 year breakeven, $3 of incremental EBITDA per $1 invested, with focus on 1–3 business domains.
• 70% talent-density shifts: 70%+ in-house, 70%+ “doer” engineers, 70%+ at competent-or-expert skill levels.
• “We don’t have a single success story where senior business leaders were not in the driver’s seat.”
• Note: McKinsey is a recurring data partner of the AI Index; the McKinsey Global AI Survey is the instrument behind several AI-adoption headline numbers in the wiki. The 20% EBITDA range is from McKinsey client work — vendor-of-deployment data — but is consistent in magnitude with independent ranges from the AI Index 2026 (14–26% productivity gains in customer support / software dev) and OpenAI 2026 (10–25% EBITDA).

Org-design lens (Werner-Le-Brun): organizations are either Tin Man Orgs (predictability-optimized for a complicated world) or Octopus Orgs (adaptive, distributed, customer-centric for a complex world). Just 12% of transformation efforts show sustainable performance gains.

Task lens (Anand-Wu): a 2×2 matrix on cost of errors × type of knowledge decides where to deploy GenAI on a per-task basis (no regrets / creative catalyst / quality control / human-first zones).

Firm-boundary lens — four models (Nishar & Nohria 2026): a sharper reframing of the build-vs-buy decision under generative-ai. The economic logic that made standardized SaaS the only practical default is dissolving — custom software is accessible again because foundation-model + vibe-coding tools (Cursor, Replit Agent, Claude Code, OpenAI Codex) compress months-of-engineering into days. The strategic question shifts from “which tools to buy?” to “which workflows do we own?” Four emerging answers, not mutually exclusive:

• Build — directly on foundation models, for distinctive jobs (e.g., a logistics company building a system that continuously optimizes delivery time and cost based on its own data — over time becomes hard-to-replicate institutional knowledge).
• Compose — vendor scaffolding and templates configured by business users (e.g., Salesforce Headless 360 for tailoring lead scoring and outreach to a specific sales motion without building a CRM from scratch). Direction reverses: software adapts to the company.
• Collaborate — providers’ forward-deployed engineers build bespoke systems in weeks rather than the months/years of a traditional ERP deployment. Speed and alignment in exchange for external dependency.
• Buy outcomes — procure the outcome itself, not the tool. Adobe’s 2026 outcome-based pricing for its CX Enterprise / agentic AI tools is the named industry signal: Adobe deploys agents directly with the customer and charges for outcomes (e.g., successful ad campaigns) rather than per seat or per token.
• Empirical anchors: enterprise GenAI app spending $1.7B (2023) → $37B (2025) (~22× in 2 years; SaaS took ~10 years for comparable penetration); 40% of code AI-generated; >1/3 of companies have replaced ≥1 SaaS tool with a custom GenAI alternative; public SaaS valuations 30–60% below 2021 peaks.
• Strategic implication: “the boundary of the firm becomes a variable rather than a given.” Reinforces the micro-productivity-trap thesis from Dutt, Chatterji et al. — data architecture and process redesign are load-bearing, not afterthoughts. Caveat from the article: “moving quickly is not the same as moving effectively.”

Organizational-learning lens — Augmented Learners (Ransbotham et al. 2024 + Kiron & Schrage 2026): an MIT SMR × BCG 2×2 distinguishing organizations by how they combine organizational learning (test-and-learn, postmortem culture, learning codification, knowledge sharing) with AI-specific learning (using AI for new learning + performance + human feedback loops + employee learning from AI).

• Distribution (3,467-respondent global survey, spring 2024): Limited Learners 59%, Organizational Learners 14%, AI-specific Learners 12%, Augmented Learners 15%.
• Headline outcomes: Augmented Learners are 1.6× more likely to feel prepared for industry uncertainty than Limited Learners; 2.2× for talent disruptions, 1.8× for technology disruptions, 1.6× for legal/regulatory disruptions. 99% of Augmented Learners report annualized revenue benefits from AI vs 71% of Limited Learners (1.4× multiplier).
• Project-selection signature: Augmented Learners are 1.9× more likely to invest in long-term (>5-year) AI projects and 2.4× more likely to invest in high-risk projects — they treat AI as a learning substrate, not just an efficiency tool.
• The operational mechanism (Kiron-Schrage 2026) — extending the empirical finding 17 months later, the column specifies a three-step flywheel that turns the 15% advantage into compounding returns: Verification (does this output meet the standard? — binary check against existing criteria) → Evaluation (what does this output reveal? — generates new standards; requires domain expertise) → Learning capture (how do we ensure this insight persists? — version control for organizational judgment, e.g. CLAUDE.md-style files). When any step is missing, organizations consume AI outputs rather than compound benefits from them.
• Why this lens is distinct from Anand-Wu’s task lens or MIT CISR’s stage lens: the Augmented Learner construct is about how the organization metabolizes AI interactions, not where it deploys AI or what stage of maturity it has reached. A Stage 2 firm with a learning culture can be on the Augmented Learner trajectory; a Stage 4 firm without one cannot. The combinatorial 2×2 cuts across maturity stages.
• Empirical claim about strategy: orgs reporting an AI strategy are 2× more likely to generate additional business value; orgs with AI core to strategy are 2.6× more likely. But: only 39% report having an AI strategy in 2024 — back to 2017 levels (down from 59% in 2020), as GenAI’s emergence forced strategy reformulation.
• Operational claim (Kiron-Schrage): orgs that build systematic feedback loops between humans and AI are 6× more likely to derive substantial financial benefits; orgs investing in learning with AI are 73% more likely to achieve significant financial impact.

Human-reaction lens — resistance as data (Carucci 2026): a practitioner framework for interpreting how people respond to deployments. All resistance is meaningful data; the leader’s job isn’t to determine whether pushback is valid but to diagnose what it’s signalling. Three traps when leaders misread resistance (personalize / moralize / rush to resolution) and four signal categories (Loss / Anxiety / Lack of control / Flaws in change). Operationalizes the human-side mechanics of MIT CISR’s “Synchronization” pillar and Werner-Le-Brun’s Octopus principles (“make changes WITH people, not TO them”). Distinct decision layer from the deployment frameworks above — the question shifts from “what should we deploy?” to “what is the team’s reaction telling us about whether what we deployed is working?” See micro-productivity-trap for how Carucci’s category #4 (Flaws in change) aligns with the operators-see-problems-leaders-dismiss pattern.

Runtime-engineering lens — agent harnesses as the new moat (Chatterjee 2026 + Kokane 2026): a 10th lens at the engineering-stack level — see the dedicated subsection below for the full treatment. Headline: the model is rented and converging to commodity; the harness is owned and compounds. Plan for model swap, not marriage. Build Constraints before cleverness. Hire systems engineers, not AI specialists.

All ten lenses agree qualitatively: most organizations are using AI in some form, but very few are actually mature/ready/adapted. The AI Index pegs “mature” at 1%; MIT CISR pegs Stage 4 at 7%; Cisco pegs “ready” at 13%; Ransbotham et al. peg Augmented Learners at 15%; Werner-Le-Brun’s 12% transformation-success baseline gives the broader org-change context. The convergence of seven independent measurements all clustering around 7-15% upper-tail is the wiki’s strongest evidence that the high-value-from-AI cohort is structurally narrow, not measurement-artefactual. See ai-maturity-measurement-comparison for methodological cross-walk and organizational-frameworks-for-ai-adoption for the framework comparison (the Nishar-Nohria firm-boundary lens, the Carucci human-reaction lens, the Ransbotham/Kiron organizational-learning lens, and now the agent-harness runtime-engineering lens are the 7th through 10th named frameworks; see the synthesis for the running cluster).

The shape of adoption can be measured along several dimensions:

• Breadth: how many functions, regions, industries are using AI at all.
• Depth (maturity): how integrated and value-generating those uses are.
• Use-case mix: which functions and tasks are AI applied to.
• Financial impact: measured cost savings and revenue gains by function.
• Workforce impact: productivity, headcount expectations, reskilling needs.

The MIT CISR Four Stages of Enterprise AI Maturity

A stage-based progression model from Stephanie Woerner, Peter Weill, Ina Sebastian, and Evgeny Káganer at MIT CISR. Distribution is from the MIT CISR 2022 Future Ready Survey (N=721) — note the 2022 baseline predates the GenAI explosion, so current distributions may differ.

Stage	Name	% (2022)	Defining attributes	Focus
1	Experiment and prepare	28%	Workforce education; acceptable-use policies; data accessibility; humans-in-the-loop	Exploration and education
2	Build pilots and capabilities	34%	Process simplification & automation begun; use cases; APIs; LLMs (out-of-box + GenAI) augmenting work	Business cases and pilots
3	Develop AI ways of working	31%	Expanded automation; test-and-learn; architected for reuse; pretrained + proprietary models; autonomous agents	Scaling AI platforms and dashboards
4	Become AI future-ready	7%	AI embedded in decision-making and processes; selling AI-augmented services; combining traditional + generative + agentic + robotic AI	Continuous innovation, new revenue streams

The financial inflection is the Stage 2 → Stage 3 transition. Stages 1–2 firms had financial performance below industry average; Stages 3–4 firms above. Source: MIT Sloan article.

The “Four S” challenges to scale Stage 2 → Stage 3

To make the leap from pilots (Stage 2) to embedded AI ways of working (Stage 3), MIT CISR identifies four organizational challenges:

1. Strategy — Align AI investments with strategic goals; offer measurable, scalable value.
2. Systems — Architect modular, interoperable platforms and data ecosystems for enterprise-wide intelligence.
3. Synchronization — Create AI-ready people, roles, and teams; redesign work around AI capabilities.
4. Stewardship — Embed and monitor compliant, human-centered, and transparent AI practices by design — see responsible-ai.

Driving the change requires a united front among the CEO, CIO, chief strategy officer, and head of HR — not a single function’s effort.

Worked examples

• Guardian Life Insurance (regulated US insurance, MIT Sloan): Automated RFP and quoting process — turnaround 1 week → 24 hours. Embedded compliance/legal in architecture review boards. Reskilling into AI-focused roles.
• Italgas (Europe’s largest natural gas distributor, MIT Sloan): “Digital Factory” innovation hub since 2017; 300TB data platform, 23 AI models; WorkOnSite (+40% construction speed, -80% inspections); DANA (GenAI network control); 30,000 hours of AI training in 2024; commercialized WorkOnSite for €3M revenue in 2024.
• Ford (manufacturing, Cisco): AI-augmented vision systems for assembly inspection — defects on “squish tube” rubber seals dropped from 63 per month to zero. Computational fluid dynamic test for vehicle airflow: 15 hours → 10 seconds with AI prediction.
• DBS Bank (banking, MIT SMR 2026): a Stage-4-style embedded-innovation incumbent that makes AI/innovation adoption a measured org-wide obligation rather than a delegated technology project. Mechanism: 20% of every scorecard — down to each team member — is “transformation,” so capability-building cascades top-down (the Playbook Model: central transformation team holds the ready playbook + 4D training) while delivery bubbles up bottom-up (“you end up with 250 journeys or 10,000 agents”). The same engine that scaled customer-journey redesign was re-pointed at agentic AI — every managing director built an agent at the March 2026 leadership conference. Dumra’s adoption-gap framing sharpens the whole concept page: “It used to be you needed an information gap before you had the action gap. The information gap is gone. Now you’re just left with the action gap” — i.e. the binding constraint on 2026 enterprise AI adoption is doing, not knowing. Per the MIT CISR ladder DBS reads as Stage 4 (AI embedded in decision-making + selling AI-augmented services + combining traditional/generative/agentic AI).

The bottom-up/top-down complementarity + the coming data rearchitecture ( LangChain Interrupt 2026)

Andrew Ng, reporting from AI Aspire’s advisory work with Fortune-50/G2000 firms, supplies a sharp practitioner diagnosis that converges with the micro-productivity-trap and the DBS playbook:

• Bottom-up “thousand flowers” innovation isn’t paying off — but the fix is complementary, not substitutive. “All of us have invested in bottom-up innovation … and for the most part it is not paying off. So CEOs and boards are asking, where is the ROI for AI?” Keep it (it generates ideas and real incremental efficiency), but it yields point solutions; the transformation needs a top-down motion to redesign the whole workflow. This is the same bottom-up-meets-top-down structure DBS operationalises (20%-scorecard KPI top-down + journeys/agents bottom-up).
• The loan-underwriting worked example (a sibling to the Guardian Life RFP case above): automating the one-hour human loan-approval step is a small efficiency gain; rethinking the workflow into a “get approved in 10 minutes” product is the transformation — and that requires “someone with the broader scope to rethink and redesign the entire workflow” (marketing, routing, diligence, execution). The unit of value is the redesigned end-to-end workflow, not the automated task — the trap-escape in operator vocabulary.
• Cost savings vs growth; swing-for-the-fences. Cost savings are capped; “growth has almost no practical ceiling.” A counterintuitive lesson: incremental gains can be harder than transformative ones (you can’t make people “work 50% harder”). Narrowing the 300-idea spreadsheets firms send AI Aspire to a small portfolio of high-conviction bets needs hard technical and business analysis and a top-down resource-allocation motion.
• The coming unstructured-data rearchitecture — a major new claim. The most common enterprise pain point Ng names is rethinking the data architecture: the last 10–20 years organised structured data; now that AI can process unstructured data (text/images/PDF/audio/video), getting it to agents “at the right time, in the right place” is suddenly far more valuable. He finds no good off-the-shelf solution and predicts “tens to hundreds of millions of dollars” of data-rearchitecture projects across many firms over the next few years, blocked by fragmentation, governance, no consensus schema, data-on-laptops, and permissions designed for humans not agents (“does an agent inherit my permissions?”). This is the data-layer prerequisite that the Nishar-Nohria and Chatterji lenses flag as “load-bearing, not an afterthought,” and an adjacency to the Open Knowledge Format agent-data-sharing problem.
• Optionality as a buyer discipline. “I personally almost never sign longer than a one-year contract regardless of the discounts offered” — because the leading model/coding-agent a year out is unknown; plus vendor-neutral observability (LangSmith) and open-weight hedging. The buyer-side complement to the “plan for swap, not marriage” prescription.

Relatedly, Mollick supplies the same redesign-the-work-product point from the individual-productivity side: raw productivity gains just yield “100× more PowerPoint” unless the organisation rethinks what the work product is — human systems “weren’t built for an AI world.”

Business-transformation-not-technology + viral adoption + the two ROI “surefire hits” ( Snowflake, McKinsey Podcast 2026)

Sridhar Ramaswamy (CEO, Snowflake) supplies a CEO-vantage read that converges tightly with the page’s transformation lens (Dutt/Chatterji), the micro-productivity-trap, and the Ng / DBS enterprise reads:

• “This is more of a business transformation than a technology transformation.” The lever is change management and org-structure rework — “I don’t have to have the organizational silos … a certain reporting structure that doesn’t have to happen in a world of agents.” Same diagnosis the page’s micro-productivity-trap / process-lock-in sources name: value comes from redesigning the org around AI, not bolting AI onto existing structure.
• Change management via viral adoption (a “happy accident”). Snowflake shipped an unrestricted internal coding agent (Cortex Code / “Coco”, CLI + desktop); because all company data already lived in Snowflake (the Snowhouse instance), “the entirety of the company became AI literate pretty much in a matter of six weeks … I did not have to mandate. I didn’t have training programs. Coco spread virally because people got so much utility from it.” The prescription — “make your team embrace AI without forcing it down everyone’s throats” — is a counterpoint to top-down mandate, and a worked instance of the resistance-as-data / make changes WITH people stance. The data-already-in-one-place precondition is the data-architecture prerequisite made concrete.
• AI with ROI — two “surefire hits.” (1) Software engineering (Cortex Code); (2) support — Snowflake “rolled our own support system … written in six weeks on Cortex Code,” now “support queues are pretty much empty”; the SRE team rewrote its observability/alerting stack, automating 4-day Kubernetes-log investigations. A clean, named answer to the page’s recurring where-is-the-ROI question — and notably both hits are internal engineering/operations, not customer-facing product.
• Redeploy, don’t cut; growth over headcount. When AI made a dedicated demo team unnecessary (every account exec makes their own demo), the team was moved into other roles; the work is identifying the new jobs AI creates and transitioning people. Kutcher’s reframe (endorsed): “more code → grow my value proposition at an outpaced rate” + filling roles previously left open — the augmentation-with-reinvestment pattern at CEO altitude.
• Consumption pricing as the AI-era model. “Customers pay only if they get value”; tokenization pricing “is not likely to last … not every token has the same value”; per-person + per-account limits to bound runaway spend. A business-model-renewal signal for how AI value is captured (a strategic-renewal move on how value is priced).

The Anand-Wu 2×2 task framework

A complementary task-level lens from Anand & Wu (2025). Maps each enterprise task onto two axes:

• Cost of errors: low (small inefficiencies in a draft) ↔ high (reputational damage, legal liability, physical harm)
• Type of knowledge: explicit data (structured/unstructured but capturable) ↔ tacit knowledge (experiential, intuitive, context-specific)

	Tacit knowledge	Explicit data
High cost of errors	Human-first zone — Human leads, AI assists with minor tasks. Setting strategy, integrating enterprise systems, disciplinary decisions, hiring critical employees.	Quality control zone — AI produces, human verifies. Drafting high-value contracts (Harvey); writing production software code (GitHub Copilot); due diligence.
Low cost of errors	Creative catalyst zone — AI creates options, human selects. Creating advertisements, outlining sales scripts, developing products.	No regrets zone — AI does it all (no human in the loop). Bulk customer inquiries, document summarization, résumé screening. Where ai-agents thrive.

The framework’s punchline: stop debating GenAI’s intelligence; ask which tasks GenAI can assist with today to make human judgment more effective. See organizational-frameworks-for-ai-adoption for how this maps to other frameworks.

Why don’t gen AI gains show up in P&L? (Anand-Wu’s leakage diagnostic)

Anand-Wu’s most actionable artifact: six leakage points along the value chain where potential gains evaporate before reaching the P&L.

#	Leakage point	What goes wrong	Owner
1	Task efficiency	Fail to identify tasks where gen AI improves efficiency	Everyone, enabled by CTO/CIO
2	Employee adoption	Miss opportunities because employees aren’t trained	Everyone, enabled by CTO/CIO
3	Resource redeployment	Labor capacity saved isn’t redeployed to higher-value tasks	Every manager, enabled by CEO/COO
4	Organizational throughput	Fail to redesign processes to capitalize on gains	Every manager, enabled by CEO/COO
5	Market demand	Customers don’t have a need to purchase the greater output	CEO + C-suite
6	Competitive retention	Competitors use gen AI similarly; gains dissipated through lower margins	CEO + C-suite

This is the wiki’s sharpest answer to the AI Index 2025 puzzle: why is adoption 78% but maturity 1%? Because gains are leaking at multiple points, and only diagnosing the leakage at every step gets ROI to the bottom line.

The paradox of access (Anand-Wu)

A counterweight to the “AI is a competitive moat” rhetoric: “Because everyone can use it, it becomes dramatically harder to capture value with it.” Pattern from prior tech cycles:

• E-ticketing (2000s): all airlines adopted; benefits flowed to customers as lower airfare.
• CAD/ERP (1990s+): once an advantage, became table stakes.
• Big Law (1990s+): clients pulled work in-house using digital tools; nearly 90% of large law firms now offer flat-fee or favorable pricing; in-house counsel tripled 1997–2020.

Implication: competitive differentiation in 2025+ comes from complementary assets — proprietary data, unique people/processes/culture — not from “having AI.” See organizational-frameworks-for-ai-adoption.

Key claims

Empirical claims drawn across the corpus, organised by the layer they speak to: breadth (how many organisations use AI at all), depth (how mature that use is), use-case mix (where AI is being deployed), financial impact (what the gains look like in P&L), productivity (task-level effects, with the consistent equalising signature surfaced by jagged-frontier task-by-task variance), and workforce (expectations vs. realised outcomes from the Brynjolfsson canaries panel).

The 2024 jump (breadth)

• 78% of orgs use AI in at least one function in 2024 (vs. 55% in 2023; ~50% during 2017–2022). Source: AI Index 2025 §4.4.1, citing McKinsey survey n=2,854.
• 71% use generative AI in at least one function (vs. 33% in 2023). The use-gap between any-AI and GenAI shrunk from 22pp to 7pp in a single year.
• The rise was global. Regional 2024 adoption: NA 82% (+21pp YoY), Europe 80% (+23pp), Greater China 75% (+27pp — largest jump), Developing markets 77% (+28pp), Asia-Pacific 72% (+14pp). AI Index 2025 §4.4.4.

Maturity is rare (depth)

• Only 1% of C-suite executives describe their GenAI rollouts as “mature” — McKinsey complementary survey of developed-markets execs, via AI Index 2025 §4.4.5.
• Most companies report cost savings of <10% per function and revenue gains of <5%. The financial impact is real but small at typical adoption depth.

Use-case mix

Function	AI use rate (Tech industry, 2024)
IT	48%
Marketing & sales	47%
Product/service development	47%
Software engineering	45%
Service operations	42%
HR	24%

Industries by overall AI use (decreasing): Technology > Media/telecom > Financial services > Energy/materials > Health care > Consumer goods > Advanced industries > Business services. Source: AI Index 2025 §4.4.2.

Top GenAI use cases: marketing strategy content (27%), knowledge management (19%), personalization (19%), design (14%), code creation (13%), automation of sales follow-up (13%), customer-service workflow integration (12%), sales lead identification (11%), accelerated R&D simulation (11%), scientific literature review (11%). AI Index 2025 Fig 4.4.5.

Financial impact by function

Most companies that report any financial impact estimate it as modest. Cost savings <10% is the most common bucket; revenue gains <5% is the most common bucket.

Function	Report cost savings (analytical / GenAI)	Report revenue gains (analytical / GenAI)
Marketing & sales	34% / 47%	71% / 66%
Service operations	49% / 58%	57% / 63%
Supply chain & inventory	43% / 61%	63% / 67%
Software engineering	41% / 52%	44% / 57%
Strategy & corporate finance	— / 56%	— / 70%
HR	37% / 56%	—
Product/service dev	23% / 43%	56% / 51%

Source: AI Index 2025 §4.4.

Productivity (consistent equalizing effect at task level)

Five rigorous empirical studies (n>200,000 across customer support, software, materials science, knowledge work) converge on AI productivity gains in the 10–45% range, with a robust equalizing effect — low-skill workers benefit more.

Study	Task	Low-skill gain	High-skill gain
Brynjolfsson, Li & Raymond 2025 (QJE)	Customer support	30% RPH; quality up	~0% RPH; quality DOWN slightly
Dell’Acqua et al. 2023	Consulting	43.0%	16.5%
Cui et al. 2024	Software engineering	21–40%	7–16%
Hoffman et al. 2024	Software engineering	12–27%	5–10%

Plus: Microsoft workplace study, Toner-Rodgers 2025 (materials scientists, +44.1% discovery / +39.4% patents / +17.2% prototypes). Sources: AI Index 2025 §4.4 and (for the customer-support study) the primary source Brynjolfsson, Li & Raymond 2025 QJE.

Important refinement (primary-source upgrade): The Brynjolfsson, Li & Raymond customer-support study, in its Quarterly Journal of Economics version, reports the headline as +15% RPH overall, not the +14.2% from the NBER working paper version that was cited via the AI Index. More importantly: the equalizing effect is not “high-skill workers gain 0%” — it’s “high-skill workers see small speed gains AND a small DECLINE in quality of their conversations.” This nuance is load-bearing for automation-vs-augmentation: augmentation is positive overall and reliably equalizing for low-skill workers, but not strictly Pareto-improving at the top — and the long-run training-data quality of the AI system depends on top performers continuing to make original contributions, which the paper finds is being diluted.

Important caveat (added 2026-04-28 batch): the equalizing effect is measured at the task level within a role, not at the occupation level across firms. As Brynjolfsson, Chandar & Chen (2025) show using ADP payroll data, employment for early-career workers in highly AI-exposed occupations has declined ~13% relative since late 2022 — even as productivity per worker has risen. Both findings are true: AI raises individual productivity (especially for low-skill workers) and reduces the number of workers needed in automate-able roles. See ai-employment-effects and automation-vs-augmentation.

Workforce expectations (mixed and softening)

McKinsey survey, via AI Index 2025 Fig 4.4.13:

• 31% of orgs expect little change in workforce size over 3 years.
• 43% expect workforce decreases (8% by >20%, 14% by 11–20%, 21% by 3–10%).
• 23% expect workforce increases.
• The share predicting workforce reductions has declined YoY. Business leaders are becoming less convinced AI will shrink workforces in the near term.
• 46% expect >20% of the workforce to need reskilling.

Realized employment outcomes (Brynjolfsson et al. 2025)

Brynjolfsson, Chandar & Chen (2025) use ADP payroll data covering ~25M U.S. workers (Jan 2021 – July 2025) to test whether AI is yet displacing human labor. The empirical correlate of the McKinsey expectations:

• Early-career workers (ages 22–25) in the most AI-exposed occupations: ~13% relative decline in employment since late 2022 (firm-time-effects-controlled).
• Software developers ages 22–25: nearly 20% decline from peak in late 2022.
• Older workers in same occupations: stable or growing.
• Concentrated in automation uses (not augmentation) — see automation-vs-augmentation.
• Adjustments visible in employment, not wages (wage stickiness).
• The McKinsey expectation that AI may not shrink overall headcount is roughly consistent with the Brynjolfsson finding that overall employment continues to grow — but with a critical compositional twist: the decreases are concentrated at the entry level, not spread evenly across all roles.

This is the wiki’s first measurable empirical evidence for AI labor displacement, distinct from survey data on expectations. See ai-employment-effects.

Integration depth correlates with productivity payoff

Necula et al. 2024 (Romanian survey, n=233): organizations with high AI integration showed a 72% probability of significant productivity improvements vs. 3.4% for those with minimal integration.

The runtime-engineering lens: agent harnesses as the new moat (Chatterjee 2026 + Kokane 2026)

A 10th lens on enterprise AI adoption emerges from the practitioner literature on agent harnesses (April–May 2026). The vantage is engineering-level — sitting beneath the strategy/maturity/transformation frameworks above — and offers a sharp prescription that complements Nishar-Nohria’s firm-boundary lens:

• “Plan for swap, not for marriage.” Foundation models are converging to commodity status. The decision still matters, but the moat it produces lasts months, not years. Keep the agent harness model-agnostic where reasonable.
• The harness investment is a permanent allocation, not a phase. Build Context / Constraints / Contracts in year 1; build the Compounding (self-tuning telemetry → harness adjustment) layer in year 2; calibrate every quarter after. “This is the work.”
• Sequence matters: build constraints before you build cleverness. The Friday-in-March story (an agent emptied a customer’s workspace because no intent-validation existed) repeats itself in every AI product without a hook layer. “It is cheaper to build the layer than to write the apology emails.”
• The under-resourced role: an engineer who thinks about agents the way SREs think about distributed systems — failure modes, observability, graceful degradation, long-tail edge cases. Not a model researcher, not a prompt engineer.

Where the Kiron Augmented Learners lens measures organizational learning capability (the 9th lens above), the agent-harness lens measures runtime engineering maturity. They are complementary, not redundant: an org can have strong organizational learning culture and a weak harness layer (or vice versa), and AI products require both.

Critical caveat from Kokane 2026: ~90% of the harness work is mature systems engineering applied to a new substrate. If your engineering team has shipped real backend systems, “you’re already 80% of the way there.” The remaining 10% (non-determinism at the execution layer + context as a degrading resource) is where the genuine new design discipline lives. Implication for hiring: don’t hire AI specialists for harness work — hire systems engineers and let them ramp on the 10%.

The Augmented Learner advantage and its operational machinery (Ransbotham et al. 2024 + Kiron & Schrage 2026)

The eighth annual MIT SMR × BCG global survey (3,467 respondents) finds that Augmented Learners — orgs combining high organizational learning capability with high AI-specific learning capability — represent only 15% of the sample but realize 1.4–2.2× advantages on uncertainty management and revenue capture. The follow-up Kiron-Schrage 2026 column then specifies the operational mechanism (verification → evaluation → learning capture) that turns the 15% finding into a deployable playbook.

Three areas where AI augments organizational learning (Ransbotham et al. operational core):

Area	Mechanism	Worked examples
Knowledge capture	AI extracts tacit knowledge resistant to legacy codification	NASA Mars 2020 (rover learns “interesting” terrain without explicit definition); LG Nova AR glasses (real-time tacit-knowledge extraction from factory workers); Capital One (Prem Natarajan, test-and-learn approach)
Knowledge synthesis	AI systematizes data overwhelming legacy analytics	Stitch Fix (years of customer Fixes summarized for stylists); Expedia Group (1.26 quadrillion combinations across 3M properties + 500 airlines)
Knowledge dissemination	AI delivers personalized, inclusive knowledge access	Slack (native daily channel recaps; >700M messages/day); cloud services provider (TikTok-style micro-adaptive learning during COVID pivot)

The flywheel (Kiron-Schrage):

1. Verification — Does this output meet the standard? Binary check against existing criteria. Unverified AI output is “noise with a confident tone.” Verification used alone catches errors without generating learning.
2. Evaluation — What does this output reveal? May generate standards that did not exist before. Requires domain expertise. The expert as evaluator is discovering what quality means in this new context.
3. Learning capture — How do we ensure this insight persists? Version control for organizational judgment. Without it, evaluation is a one-time event; with it, every subsequent interaction starts smarter.

The trap (load-bearing claim): most organizations practice verification masquerading as evaluation — treating AI outputs as verdicts to confirm rather than starting points to interrogate. Consumption dressed up as adoption. The remedy: deploy AI first in domains where your people already have deep expertise, because evaluation requires someone capable of recognizing what “not perfect” actually means.

Five concrete moves (Kiron-Schrage):

1. Preserve evaluation expertise — domain expertise repositioned, the expert as evaluator rather than the expert as producer.
2. Build verification mechanisms — minimally viable verification (multijudge systems, consistency checks across formulations); start the cycle even when full verification is expensive.
3. Institute evaluation practices — after every significant AI interaction, ask: What worked? What failed? What was interestingly wrong? (The third question is where hidden value lives.)
4. Create capture systems — both inferential (patterns in accumulated traces) and explicit (decision journals, prompt repositories, evaluation logs). Discipline, not cost or creativity, is the true constraint.
5. Measure the cycle, not just the output — count verifications, evaluations, learning captures, and how quickly captured learning changes subsequent practice. Tools-adopted / hours-saved / tasks-completed are consumption metrics.

Convergence with prior wiki claims:

• The Augmented Learner traits map onto MIT CISR Stage 3 attributes (test-and-learn, architected for reuse, human-feedback-loop AI).
• The verification → evaluation → learning capture cycle is the operational answer to the micro-productivity trap — Bain/OpenAI’s “process lock-in” failure mode is what happens when verification masquerades as evaluation.
• “Return on iteration” (Kiron-Schrage’s coining) is a candidate replacement metric for “ROI” in AI deployments, distinct from but adjacent to the 6-capability “Adoption and scaling” capability in Rewired 2nd ed.

The deployment-surface migration: coding moved to API (Anthropic Economic Index, 5th report)

A finding worth flagging because it changes how enterprise AI deployment shape should be read: between Aug 2025 and Feb 2026, the share of coding work on Claude.ai dropped sharply while the share on the 1P API rose from 36% → 47%. The mechanism, per AEI 5: as agentic coding harnesses (Claude Code, Cursor, etc.) matured, coding work migrated off the conversational consumer surface and onto the API where many short directive task-labelled calls happen per actual task. This has two consequences for adoption measurement:

1. Single-surface adoption metrics undercount the shift. A survey that asks “do you use ChatGPT/Claude.ai” misses the place where enterprise coding adoption is now happening.
2. Where automation is rising is surface-specific. AEI 5’s “automation rising on API” finding (sales, trading, coding harnesses) is consistent with the Claude.ai split staying flat (53/45 augmentation/automation). The aggregate “AI is automating” / “AI is augmenting” framing dissolves once you separate the consumer chat surface from the agent-mediated API surface.

Reinforces the micro-productivity-trap reading: the firms capturing real value from coding-AI today are the ones operating it through agentic harnesses on the API, not through individual ChatGPT/Claude.ai seats. The deployment surface itself is part of the maturity gradient.

The AI foundation cluster: $11T market-cap accretion since 2022 (MGI 2026)

The wiki’s first economy-side scale anchor for the AI deployment race. MGI’s [[2026-03-25-russell-bradley-mgi-race-takes-off-next-big-arenas|The race takes off in the next big arenas of competition]] (March 2026, 127 pp.) names a three-arena cluster — semiconductors + cloud services + AI software & services — as the “AI foundation”, and quantifies its 2022–25 trajectory in firm-level terms:

AI-foundation metric	Value	Comparator
Market-cap accretion since 2022	+$10.77T	60% of the entire 18-future-arena market-cap accretion of $17.79T
Revenue accretion since 2022	+$490B	one-third of the 18-future-arena revenue accretion
AI software & services market-cap CAGR 2022–25	142%	vs 8% non-arena market-cap CAGR
AI software & services revenue CAGR 2022–25	55%	vs 1% non-arena revenue CAGR
AI software & services 2040 revenue projection	$1.5T–$4.6T	from $85B baseline (17–25% CAGR)
Semiconductors market-cap accretion alone	+$7.39T	~68% of the AI-foundation total

This firm-level scale is the empirical complement to the wiki’s prior adoption-depth gradient (78% adoption / 1% mature). Adoption breadth + AI-foundation valuation accretion together imply the deployment race the wiki has tracked qualitatively is now visible in the listed-equity market caps of the firms providing the substrate. Open question that MGI itself raises: “Whether investor expectations for AI’s future will be supported by companies’ sustained returns on invested capital above the cost of capital is one of the biggest open questions in business today.”

The omniscaler thesis as cross-arena platform play (MGI 2026)

MGI coins “omniscalers” for nine firms competing in three or more future arenas and sitting in the top-30 global R&D + capex spenders: Amazon (cluster — incl. Blue Origin, Prometheus), Tesla/X (incl. SpaceX), Alphabet (cluster), Microsoft (cluster), Meta, Apple, Samsung, Alibaba, Huawei. 6 of 9 are US-headquartered. Collectively in 2025:

• ~$700B operating cash flow (combined)
• ~$800B R&D + capex (combined)
• ~$2.7T total revenue (larger than Italy’s GDP)
• ~$200B average per-omniscaler arena revenue vs ~$10B average for other arena players — a ~20× advantage per arena that compounds across the 3-to-9-arena spread.

The structural advantage: “reusable infrastructure, data network effects, high risk appetites, and top talent attraction among their interlocking strengths.” Capability built in one arena (e.g., Alphabet’s search/YouTube engagement data → AI training data) becomes accretive in the next. Alphabet plays in 9 arenas; Amazon in 8; Samsung in 7.

This sharpens the wiki’s prior tracking of enterprise AI adoption: at the top end of the firm-size distribution, “AI adoption” is no longer a category that captures the operative dynamic — these firms are not adopting AI, they are simultaneously building, selling, and reselling it across arenas. The omniscaler concept is single-source so far; the wiki tracks it as a candidate concept page pending second-source corroboration. Worked example of how to read it alongside YC: Hu’s “thousand-fold engineer + closed-loop company” prescription is the startup-side analogue of the omniscaler advantage — the omniscalers operate the same closed-loop pattern at multi-arena scale.

Sharper formulations from May 2026 (Chamath, Spiegel, Ognibeni, Price, Jha, Glasgow)

Operator-narrated framings push the wiki’s existing position further:

• Firm-rebuild as the ROI gating condition (Chamath 2026). The wiki has held that shallow adoption produces task-level gains without firm-level value; Chamath sharpens it: “Unless you completely rebuild how those companies operate, it will not pay [the trillion-dollar AI capex] off and there will be blood in the streets.” The claim is stronger than the wiki’s prior formulation: not merely adoption-depth-matters but only-full-rebuild-justifies-the-capex. The mechanism named: legacy enterprise systems carry institutional tribal knowledge inaccessible to AI agents until reconstructed (see the COBOL retiree anecdote). 8090’s Software Factory is the first wiki source describing an enterprise-rebuild as a productised offering (target sectors: healthcare, financial services, manufacturing, government).
• JTBD-as-sequencing for AI transformation (Spiegel 2026). A consumer-scale operational answer to “how do you sequence AI deployment across the company without it dissolving into chaos.” Snap’s method: enumerate jobs-to-be-done per user type, map agents and cross-functional teams to those jobs, track progress against the per-job business outcome. Paired explicitly with the “thousand flowers bloom” idea-generation default — bloom at the experimentation layer, sequence at the resourcing layer. The wiki’s first operator-grade consumer-scale AI-transformation sequencing framework — complementing the consulting-firm frameworks (Bain / McKinsey / Thoughtworks) with a single-company in-product-team narration.
• The China-as-time-machine four-lessons diagnostic (Ognibeni 2026). A Berlin-Expo keynote naming the why-doesn’t-Western-AI-show-up-in-the-P&L gap from the outside-in. Four operator-grade failure modes for the BCG / PwC / McKinsey adoption gap: (1) cost-vs-growth framing — Western firms default to bottom-line savings (Klarna fired-then-rehired service team) while Chinese firms default to top-line business-model renewal (Shein, BYD, Xiaomi); (2) demo-vs-scale failure — pilots work, edge cases break; (3) silo-vs-ecosystem mismatch — Chinese firms operate as integrated ecosystems (Alibaba, ByteDance), Western firms as disconnected departments; (4) trust deficit — Alibaba’s “if the translation is wrong, it’s Alibaba’s fault” liability-shift is the trust-scaling pattern Western platforms have not yet adopted. The wiki’s first source naming concrete RMB revenue lines for deployed Chinese agentic-commerce systems (JD.com Joy Streamer ~$250M sales in one Double 11 season via virtual digital-twin live-stream hosts; Alibaba’s Qen one-sentence-purchase agent at 100M+ users with tokenized mandates as guard rails).
• Agentic AI as strategic threat to the customer relationship (DFI 2026). A named regional retail-CEO public-record articulation of the disintermediation thesis — DFI Retail Group’s CEO Scott Price on what keeps him up at night: “Agentic AI and creating a personal assistant — the idea that an individual will have a relationship with an agentic personal assistant that says, ‘I need eggs, dentist appointment, car to pick up the kids at 9 a.m.’ — basically disrupts the relationship that each one of those service providers has with customers today. With that goes loyalty, with that goes access to data. So there is going to be this arms race for retailers to understand in that agentic world, how do you ensure you maintain that relationship with the customer.” Paired with DFI’s three-axis AI investment framework (personalization / cost management / running a better business) operating at scale across thousands of supermarket, convenience, H&B, and F&B outlets. The seller-side mirror of Ognibeni’s “nobody will show up in your store when you only do search-driven e-commerce” — the two are the wiki’s most directly paired buyer-side / seller-side framing of the agentic-commerce disintermediation thesis.
• The AI-native safety inversion at the enterprise procurement layer ( Campfire CEO 2026). A YC Root Access fireside with John Glasgow, CEO and founder of Campfire (S23, AI-native ERP), naming the wiki’s clearest vendor-side observation of the enterprise-procurement narrative flip on what counts as a safe buy. “Buying the legacy version was considered very safe. But then once AI started to take off — call it end of ‘24 — it then become being the incumbent meant you were not AI-native. And so there was a flipping of the narrative that the board and the executives were saying, we want AI-native. And even if the accountant wasn’t fully ready to embrace AI, they had this blessing to go buy something new that nobody had heard of and their C-suite or their auditor wasn’t familiar with yet.” Three load-bearing properties of the flip: (1) narrative, not feature (Campfire had AI features long before the inflection); (2) C-suite/board/auditor channel, not the operator channel (day-to-day operator hesitancy is overruled, not converted); (3) the narrative had to shift even though the core product did not. Hard empirical anchor: a four-employee seed-stage vendor pulling >100 enterprise customers off NetSuite from Q4 2024 onward, “more than doubled ARR each quarter since then,” with customers signing two-year contracts “longer than your runway” — pricing the procurement-side flip directly. Companion claim: the tech-stack-turnover thesis (“everything in the finance tech stack — payroll, spend management with the Brexes of the world — had all turned over in the last 5–10 years except the core general ledger… the contrast was very acute”) is the wiki’s first source naming why the GL was the last enterprise-SaaS category to flip. Glasgow is the enterprise-B2B vendor-vantage twin of Jha/Emergent’s long-tail observation below — same vendor-side phenomenon, opposite ends of the customer-size spectrum, observed independently in May 2026.
• Adoption-velocity asymmetry favouring the long tail ( Emergent CEO 2026). A named-CEO Anthropic-Applied-AI-interview anchor for the counter-thesis that the fastest-adopting AI customers in 2026 are not Fortune 500 enterprises but the long tail of SMBs / domain experts / business operators who never had software at all. Emergent reached $100M ARR in 8 months, 7M users in 190 countries, 70–80% of whom never wrote a line of code — after a deliberate enterprise → consumer/SMB pivot: “We had an enterprise customer started working with them but realized that the adoption in enterprise is going to be slow.” The wiki’s clearest 2026 named instance of this asymmetry — sits in productive tension with the slower-enterprise framings of MIT Tech Review Insights, Nohria, and the AI-Index / MIT CISR maturity-distribution data above. Recurring procurement-side claim that resists benchmark quantification: across the four Feb–May 2026 Claude-channel customer stories (Lyft “Claude’s personality is really what stuck out”; HubSpot “Claude has really good taste”; Figma Make “every single person who has taste can just enact it all that easier”; Emergent “Opus has been our workhorse […] really great at instruction following”), the named model-selection criterion is personality / taste / voice / instruction-following, not benchmark scores. Worth tracking as a 2026 procurement-side pattern distinct from the leaderboard-driven model-selection vocabulary in the harness/eval clusters.

The MGI workflow-as-unit-of-value-capture + $2.9T US economic value scaffold ( MGI 2025)

[[2025-11-25-yee-mgi-agents-robots-and-us-skill-partnerships|MGI’s Agents, Robots, and Us]] (November 2025) supplies the wiki’s most quantitatively-precise structural anchor on the enterprise-AI-adoption value question. Four substantive contributions:

(a) The $2.9T US / $28.7T global economic-value scaffold — under the midpoint adoption scenario for 2030, AI-powered agents and robots could generate ~$2.9T US economic value per year; globally ~$28.7T (up from the earlier untimed ~$26T MGI projection in The economic potential of generative AI, June 2023 — the same model now timed to 2030). The model assumes ~27% of US current work hours automated by 2030 (range 20% healthcare → 31% manufacturing across sectors).

(b) Workflow-as-unit-of-value-capture: the central operational prescription. “Realising these gains requires more than automating individual tasks. It will mean redesigning entire workflows so that people, agents, and robots can work together effectively.” The corollary diagnostic — “Nearly 90 percent of companies say they have invested in [AI], but fewer than 40 percent report measurable gains. The gap may reflect the fact that many projects are still in pilot or trial phases or that organizations are applying AI to discrete tasks rather than redesigning entire workflows” — anchors the failure mode at structural altitude. The 190+ workflows mapped across 16 business functions form the wiki’s most granular publicly-available workflow taxonomy at MGI altitude.

(c) 60% sector-specific + 40% cross-cutting value concentration: MGI’s most actionable prioritisation guidance for enterprise leaders is invest in sector-specific workflow redesign first, cross-cutting workflows second. Breakdown of the $2.9T:

• Sector-specific domains: $1.7T (60%) — Knowledge services $773B + Frontline services $424B + Production services $556B. Each sector has its own ~90 sector-specific workflows in 3 service categories.
• Cross-cutting domains: $1.2T (40%) — IT $196B, Marketing $167B, Planning + management $161B, Administrative services $133B, Logistics $122B, Sales $90B, Product/R&D $87B, Customer support $60B, Talent + org $56B, Finance $42B, Procurement $31B, Legal $21B, Risk + compliance $11B.

(d) 77% from agents + 23% from robots — agents dominate the economic-value claim ($2.26T agents vs $0.67T robots), reflecting the 65/35 nonphysical/physical work-hour split in US labour. By sector, agent share ranges from 57% in agriculture + accommodation/food (highest robot intensity) to 92% in finance + insurance (highest agent intensity).

Sector adoption rates (avg automation as share of current work hours, midpoint scenario, 2030): 31% manufacturing, 30% finance/insurance + accommodation/food, 29% utilities + mining + agriculture, 28% information + management of companies, 27% wholesale + other services, 26% real estate + arts + retail + administrative, 25% educational + transportation, 20% healthcare. The 20%-31% range across sectors is informative — no sector approaches the 57% technical-automation potential in the midpoint scenario; adoption is bounded by integration friction, labour-vs-capital costs, regulatory speed, and customer acceptance.

Four worked operational cases (Ch.3) operationalising the workflow redesign:

Case	Industry	Agents	Outcomes
B2B sales	Global tech	5 (prioritisation / outreach / response / scheduling / handoff)	+7-12% revenue; -30-50% time across sales roles
Utility customer ops	Utility (7M+ calls/yr)	4 (inbound / intent ID / scheduling / self-service)	40% of calls handled, >80% resolved without humans; cost/call -50%; CSAT +6pp
Biopharma medical writing	Pharma	6 (planning / data mapping / drafting / validation / reviewing / submission)	Touch time for first human drafts -60%; errors -50%
Regional bank IT modernization	Banking	3+ (modernisation planning / assessment / functionality / coding / QA / testing)	Code accuracy up to 70%; human hours -50%

MGI’s framing converges with Bain’s micro-productivity-trap (task gains don’t aggregate to firm value without transformation), McKinsey (“organisational change is half-or-more of the secret sauce”), and AWS (“AI bolted on is going to fail… focus at a workflow level”). The cross-consultancy convergence on workflow as the unit of AI value capture is now the wiki’s strongest evidence-based prescription for enterprise leaders.

The 5–10-year sector-by-sector replacement speed model ( Lenny’s Podcast May 2026)

Evans (May 2026) supplies the wiki’s clearest independent-analyst-altitude speed-of-adoption frame:

“Typical big-company enterprise software sales cycle is 18 months if you’re lucky. Enterprise sales cycle is shorter than the venture-backed startup software funding cycle — longer — like it takes you longer to get an enterprise deal than it takes you to go between rounds. So no, people aren’t just going to tear out SAP and replace it with XYZ. Maybe in three, five, ten years yes that whole estate will look radically different and all those jobs will have changed, but it will take three, four, five, ten years and it will take time sector by sector and it will take time for people to work out — oh, you could do that thing with this.”

Three propositions packed inside:

(a) The 18-month-enterprise-sales-cycle floor. Adoption can’t outpace procurement. The enterprise-sales-cycle floor on adoption velocity is structural; no model-capability improvement removes it. Convergent with the Anthropic engineering-org practitioner observation that customer onboarding pace, not model release pace, sets the cadence at which AI-native workflows actually land in clients.

(b) Sector-by-sector pacing. The whole-estate-rewrites will happen — but at sector-staggered timelines spanning 3–10 years. Evans’s frame is the sales cycle is shorter than the funding cycle is shorter than the integration cycle — money flows faster than software does, software flows faster than workflow integration does, and workflow integration is the rate-limiting step.

(c) The delay-until-someone-realises-you-could-do-that-thing-with-this lag. Evans’s worked example: “There’s a company called Frame.io which is video editing — video collaboration. And there’s nothing there that you couldn’t have done at least five years earlier and maybe ten years earlier … the delay was somebody realizing — oh we could — that problem exists inside that industry and oh this is the way that we would solve it. It didn’t all happen the day after Google Docs.” The pacing-of-adoption is not bounded by capability availability; it’s bounded by the recognition that capability X applies to problem Y in industry Z — a knowledge-discovery process that runs sector-by-sector at a much slower clock than the capability-frontier itself.

Sits alongside the MIT CISR Stage 2 → Stage 3 financial inflection finding and the OpenAI micro-productivity-trap result: high pilot density and shallow task-level gains can co-exist with multi-year time-horizons for firm-level value to land — Evans’s 3–10-year frame is the macro version of MIT CISR’s between-stage gradient and Bain’s reinvention-not-task-optimization prescription.

The AI-native services company vendor-side playbook ( YC Startup School June 2026)

Warren (June 2026) supplies the vendor-side architectural mirror of [[2026-05-05-nishar-nohria-end-of-one-size-fits-all|Nishar & Nohria’s Buy Outcomes model]]. Where Nishar-Nohria frame Buy Outcomes from the enterprise-buyer firm-boundary angle (when a firm contracts for an outcome rather than licensing software, the entire job is automated by the vendor), Warren writes the YC-altitude playbook for building that vendor.

Core thesis: “Some of the biggest companies of the next decade won’t be software businesses at all. They’ll be services companies like insurance carriers and law firms rebuilt from scratch with AI doing most of the work.” Markets named as YC-internal known-good-fit: tax, audit, insurance, mortgages, parts of healthcare, parts of logistics.

Outcome-as-product vs co-pilot-as-product — the central distinction: “companies provide the outcome to the customer versus build a co-pilot that the customer uses internally.” The wiki’s clearest single-sentence YC-altitude framing of the two-vertical AI-native-company split: outcome-vendors (the Warren category) vs co-pilot-vendors (the prior YC default which the YC AI-from-the-ground-up piece operationalises).

Four market-selection traits unique to AI-native services:

Trait	What it means
Low trust	Work is already outsourced; customer cares about outcome, not how
Low judgment at the task level	Most steps automatable; judgment in a few places
High intelligence threshold	Overall work hard enough that models + humans are needed
Regulation could actually be good	Higher expectations + legal accountability → moat

Warren’s Sam Altman Test for model-disruption-resistance: “as the models get better, does your service get stronger or does the model itself commoditize you? You want to be in the first camp.” The vendor-side test, structurally adjacent to the agent-harness model-rented-harness-owned discipline applied to services-company strategy.

P&L opportunity-size frame: “Traditional services firms top out around 30% margins. Pure software and agent companies have more margin, but often smaller TAMs. The bet on these services companies is that AI operating leverage gets you closer to software margin, say 50%+, on a market that’s two to three times bigger than software. You don’t need to be there right away. But the trajectory has to be believable.” The wiki’s clearest YC-altitude quantification of the services-AI-vendor opportunity-size envelope.

Don’t-buy-your-way-in rule: only valid reason to acquire an existing services business is fast regulatory moat (e.g. insurance licensing); otherwise “legacy service businesses are legacy. They have different expectations on metrics, hiring, and performance. Adding AI on top doesn’t change any of those realities. Building is almost always better than buying.”

The Warren-Nishar-Nohria pairing now forms the wiki’s first complete buyer-side architecture + vendor-side playbook pair on outcome-as-a-service. Convergent with McKinsey (incumbent-side rewrite of consulting), Luminai (worked example in healthcare-ops), and HelloPrint (day-zero rebuild at SME scale rather than YC startup scale). The services-don’t-disappear-they-reinvent thesis is now visible at four altitudes: incumbent-firm (Sternfels), buyer-firm-boundary (Nishar-Nohria), startup-vendor (Warren), in-place-SME (Scheffer).

The Gartner $86B → $206B agent-spending + one-third of business decisions by 2028 anchors + Giles five executive recommendations ( WP Intelligence May 2026)

Giles (May 2026) supplies two executive-readership Gartner anchors + the wiki’s first executive-readership five-recommendations close:

• Gartner: spending on AI agents will more than double from $86.4B (2025) to $206.5B (2026) — the wiki’s clearest single forecast-anchor on agent-deployment capex.
• Gartner (separately): “agents could handle a third of business decision-making by 2028” — the wiki’s most concrete forward-looking-decision-share forecast.

The five executive recommendations Giles closes with (recapped here as the executive-altitude operational counter-prescription to the simple adopt-agents-quickly default reading of the spending data):

#	Recommendation	Load-bearing claim
1	Review workflows first, deploy agents second	Garaba/UiPath worked example: the period between offer-extended and onboarding looked like a simple process; her team found ~60 different subprocesses, some not suited to automation. The wiki’s clearest single-line executive-altitude restatement of the workflow-not-task discipline.
2	Remember agentic tech isn’t perfect	Hallucinations + consistent-accuracy problems → keep enough experienced employees to apply sense checks to agent recommendations. Convergent with Narayanan’s reliability-as-gating-dimension framing at executive-readership altitude.
3	Redesign entry-level roles	IBM model (the Nickle LaMoreaux capability-gaps-from-cutting-entry-level anchor): new recruits do customer-facing projects + review agent output; not traditional starter tasks. “A far better use of AI-native talent.” See automation-vs-augmentation §16 for the worked example.
4	Beware AI-linked skill erosion	The Gartner one-third-of-business-decisions-by-2028 forecast turned into a deskilling-prevention prescription: bonuses to workers to keep practicing skills like coding; regular manual checks of key agentic systems. See ai-deskilling §“AI-linked skill erosion as an executive prescription”.
5	Communicate around AI often — and carefully	”Dire predictions of job losses made by some prominent AI leaders such as Amodei have turned the deployment of agents and other forms of the technology into a communications minefield.” The HubSpot Russell-quoted prescription: “You’ve got to cut through the other noise to focus [them] on the things that matter most.”

Giles also surfaces McKinsey’s identification of IT, knowledge management, and sales & marketing as the top-3 functions ramping up agent use — a useful function-by-function deployment-mix anchor at executive-readership altitude. The wiki has multiple McKinsey-altitude sources on the same finding (Moon et al. at McKinsey-Digital tech-partner altitude, Sternfels at McKinsey managing-partner altitude); Giles’s function-level summary serves as the executive-readership-news-survey altitude statement of the same finding.

The first-party-OpenAI deployment vantage: team agents + embedded engineering + Ask-Assist-Automate ( OpenAI, IT Revolution April 2026)

Beutler (2026) supplies the wiki’s first first-party OpenAI vantage on enterprise AI adoption — Joe Beutler, OpenAI’s Head of Solutions Engineering for Strategics, reporting both how OpenAI restructures itself and the patterns his solutions team sees succeed across OpenAI’s largest enterprise customers. Four contributions:

(a) The adoption gap is a middle-layer gap. Two approaches dominate — bottoms-up (workforce tools like ChatGPT for all employees) and top-down (one-to-three large strategic initiatives, e.g. call-center automation) — leaving a hollow middle: team agents, automation at the team/department level. “What’s been missing is what’s in the middle … how can you have more automation at the team and department level.” This is the vendor-of-deployment restatement of the micro-productivity-trap and the org-altitude name for the gap MGI’s workflow taxonomy quantifies — and Beutler claims closing it is how you close “the big enterprise value gap that everyone keeps talking about.”

(b) Separate governance from transformation; business owns the outcomes. “You can’t have the same person own transformation and governance, because one or the other tends to win out … if the head of a business unit is not accountable for the result, you probably don’t have an agent. You have a demo.” Central IT owns foundations / firmwide governance / tooling / the biggest cross-company initiatives; business units own transformation in their own domain. The sharpest single articulation in the corpus of the governance/transformation split — complements McKinsey Rewired’s “senior business leaders in the driver’s seat” and operationalizes Kropp et al.’s accountability-must-be-personal prescription.

(c) Embedded engineering inside the business function. The binding constraint “is rarely the model capability … it is org design, ownership, and workflow clarity.” The fix: engineering sits inside finance / go-to-market, via an organic path (a domain expert builds tooling → full-time role → paired with an engineer interviewed by and on the same comp ladder as engineering → a head of innovation), framed as “moving centers of excellence to embedded innovation.” The three load-bearing roles: domain expert (requirements / quality / edge cases), AI expert (project selection / evals / system behavior), software engineer (context / identity / telemetry). This is the cross-customer-deployment twin of Fung’s inside-Anthropic team-shape rewrite — both land on engineering moves to where the domain work is.

(d) Ask → Assist → Automate — a deployment-maturity ladder for agentic systems: Ask (read-only — pull data, answer questions, validate sources before any write action), Assist (agent recommends, human-in-the-loop validates), Automate (the things the agent reliably gets right stop going to humans; exceptions route to people). You cannot go zero-to-full-automation, especially in regulated environments; the biggest value is at Automate, but so are the highest integration cost, quality bar, eval strength, and operational ownership. The ladder maps onto automation-vs-augmentation (Ask/Assist = augmentation; Automate = automation) and is the deployment-maturity complement to AWS’s economic-decision USE/COMPOSE/BUILD frame. T-Mobile anchors the responsible-scaling point: a $3B call-center cost center, already 60% automated, with a 75% ambition — OpenAI walked them back to start at Ask. Operational on-ramps Beutler names: a workflow inventory (map every workflow, track AI usage / frequency / time / opportunity, then pick your bets), and the emerging AI-agent-manager role (agents need ongoing management the way people need enablement — a candidate full-time job; see ai-employment-effects).

A vendor-strategy note worth flagging for the firm-boundary lens above: Beutler explicitly rejects the professional-services / outcome-as-a-service model (“we don’t have a massive professional services business … lucrative contracts that will last 10 years”) in favour of selling no-code agent builders (continuous evals + governance + connectors + shareable skills) so customers build their own agents — the deliberate counter-position to YC’s vendor-side outcome playbook and Nishar-Nohria’s buyer-side Buy Outcomes model. The same firm (OpenAI) appears on both sides of the wiki’s outcome-vs-builder debate: the Bain/OpenAI HBR article sits with the transformation-consultancy cluster, while Beutler stakes OpenAI’s product bet on the builder, not the outcome.

The workforce-readiness constraint: adoption outruns capability (LF Global + LF Europe 2026)

The two 2026 Linux Foundation State of Tech Talent reports add the workforce-readiness lens to the adoption-depth gradient: adoption intent is near-universal (97% plan to implement AI globally; 93% Europe / 99% rest-of-world are implementing), but the binding constraint is the ability to operationalise it — a survey-side restatement of the micro-productivity-trap and the BCG 5%-get-value-at-scale anchor (both cited by the reports). The gap is most acute in security and operational readiness: security/privacy is now the top barrier to adopting new technology (up from 3rd in 2025); understaffing is widespread in AI, cybersecurity, cost-optimisation, and platform-engineering; and most organisations lag on foundational AI infrastructure (the “PARK stack”). The reports’ prescription — upskill existing staff over external hiring — locates the maturity bottleneck in people and capability, not tooling (see durable-skills and ai-employment-effects). For the strategy-discipline complement (how AI changes the decision work itself), see Csaszar et al. on AI-augmented strategic decision-making.

The 30/60/90-day rollout playbook + the trust gap ( Microsoft Dec 2025)

The GitHub/Microsoft Agentic DevOps keynote contributes a vendor-practitioner adoption playbook for agentic coding: a 30/60/90-day ramp (first 30 = pilot, 31–60 = level-up, 61–90 = optimisation), with the explicit discipline of progressive ROI, don’t boil the ocean, and “break the work down for the agent like you would for a human” (giving an agent too much at once is named as the common failure). Adoption is framed as role-by-role upskilling — developers (prompt-writing, code review, security awareness), DevOps/security teams, and the C-suite — echoing the durable-skills / workforce-readiness constraint above. The keynote’s load-bearing adoption caution is the trust gap: “AI is powerful but not magic… you’re really going to have to invest in how you bring it onto your team just like you would invest in bringing another human onto your team,” plus the anti-FOMO rule don’t chase every new frontier model (each switch costs learning time). This is the developer-tooling-vendor mirror of the micro-productivity-trap adoption warning and the Beutler/OpenAI Ask→Assist→Automate maturity ladder above — adoption is an organisational-transformation problem (“DevOps is an organizational transformation first”), not a tool-procurement one.

Data-readiness as the gate + obsess over evals: the Goldman CIO vantage ( HBR June 2026)

Marco Argenti, CIO of Goldman Sachs, contributes a first-party CIO adoption playbook with three load-bearing claims that sharpen this page’s deployment-discipline spine:

• “AI transformation follows data transformation, not the other way around.” Agents without context “revert to chatbots”; data is the ground truth. Argenti’s striking prescription: leaders should be willing to delay AI-at-scale until the data is in order (months or years) — “a very unpopular concept these days.” Data readiness is itself a prioritisation signal for which use cases to tackle. The mechanism is the garbage-in/garbage-out problem made worse: AI “makes garbage output look plausible,” so the usual GIGO penalty is amplified. This is the wiki’s clearest CIO-level statement that the data layer, not the model, is the binding constraint on enterprise adoption.
• Obsess over evaluations and benchmarks. Real organisational processes resemble the Garbage Can model (chaotic, nonlinear), not the clean step-by-step SOPs firms codify. Goldman’s approach to applying AI to firmwide processes (client onboarding worked example): first codify what good looks like (process-quality metrics + experienced-operator decisions), then build evals comparing agent outputs to desired outcomes, with feedback loops so the AI self-improves. Outcome-based, not step-by-step (tell Maps the destination, not every turn). This is the enterprise-adoption face of the agent-harness contracts/evals layer.
• Demand radical targets to force genuine adoption. “If you want your developers to change habits, ask them to be 3× more productive, not 20%”; aim for “90% reduction of manual touchpoints, not 20%.” Incremental targets yield optimisation-of-the-old-process (the micro-productivity-trap); radical targets force the rethink. Even reaching halfway proves the team went through real transformation, not just speed-up.

Argenti joins the wiki’s CIO/executive-vantage adoption anchors (OpenAI Ask→Assist→Automate; Microsoft 30/60/90) — distinctively contributing the data-readiness precondition the others under-weight.

The June-2026 leadership-altitude triple: maturity arc, data-strategy gate, and the CIO-as-conductor

Three sources from the 11–12 June 2026 batch add executive-altitude adoption discipline:

• A staged maturity arc with exit criteria ( CIO Symposium). Monica Caldas’s IT-ops deployment is the wiki’s cleanest crawl-walk-run field account: personal-productivity assistance → identify which workflow pieces to reimagine → deploy micro-agents (“not one agent that does everything”) with clear OKRs and explicit entry/exit criteria through the maturity arc, plus a “trust fabric and governance.” George Westerman’s automate-first-then-put-humans-in-the-right-places and rebuild-processes-around-outcomes is the same anti-micro-productivity-trap discipline at CIO altitude. Ramesh Razdan’s “car keys to a new driver” (small→medium→large trust) is the adoption-pacing metaphor.
• Data strategy as the adoption gate ( Sydney). The keynote reframes adoption as a data-readiness problem: most value is blocked upstream (Gartner: 80% of data-governance initiatives fail by 2027; 99% invest in data but only 29% see value). The prescription is business-problem-first, not gather-clean-lake (“$50M, 5 years, one fired CIO, no value”); identify a problem → get the data that solves it right → repeat. The quotable adoption test: “can an agent consume this without a human translating?” And minimum-viable-governance (guardrails not roadblocks; open-by-default) as the way to avoid the lock-everything-down failure that drives shadow IT.
• The CIO shifts from owner to conductor ( AWS Sydney). The role re-definition: “from owner of the stack to conductor of the stack.” Plus a concrete Monday-morning sequence — pick one workflow (not one AI strategy), classify it USE/COMPOSE/BUILD, then work through talent/structure/governance. This corroborates the Allen adoption playbook (Model A is dead; map workflows to tiers) with a second AWS strategist.

The adoption-vs-capability lag + the workforce-strategy CEO playbook ( Emerson, Kropp et al. 2026)

BCG’s labor report supplies the diffusion half of the adoption story: a multiyear lag between automation potential and realized impact — “economic impact often lags model capability” because it depends on application maturity, workflow redesign, legacy integration, and human capital to deploy it. Industries with high automation potential (financial services, legal) are not the ones with high scaled adoption (technology/software lead); diffusion varies by sector, company size, and integration-talent availability, with larger enterprises moving faster. BCG names forward-deployed engineers / systems integrators / project managers — embedded with business teams to translate AI into working solutions — as both an emerging new-job category and the key bottleneck on adoption (supply ≪ demand). This is the supply-of-implementation-capacity constraint the wiki’s vendor sources (Beutler’s embedded engineering, Allen’s pod+platform) describe from the inside.

BCG’s four CEO starting points are an executive-readership adoption playbook: (1) embed workforce strategy into competitive strategy (it can’t sit downstream of automation; avoid reactive headline/peer-driven cuts that ignore your own automatable/augmentable mix); (2) refocus automation on redesign, not cost (new domain-specific KPIs; see micro-productivity-trap); (3) put upskilling/reskilling/redeployment at the centre with per-segment playbooks; (4) shape the AI narrative — sequencing matters, leading with the most-substitutable roles demoralizes and erodes the will to upskill. Pairs with the Beutler / Microsoft / Argenti CEO-vantage anchors above.

The disclosure-culture constraint: trust beats governance ( HBR June 2026)

A sharp human-factor addition to the maturity story: even where individual adoption is high, the collective gain can stay near zero because employees hide the workflows they discover — ai-knowledge-hiding. The binding driver is organizational trust, operating through psychological safety, not formal AI policy or approved tooling. In the authors’ 604-respondent survey of daily-AI users, 30.3% intentionally withheld AI knowledge; the lowest-trust quartile was ~4× more likely to hide than the highest (47% vs 14%), and having an AI policy or sanctioned tools, alone, predicted nothing. A cited Stanford study of 51 enterprise deployments found 77% of the hardest adoption challenges were non-technical — the same human-maturity gap the MIT CISR maturity work and the “trust gap” anchor describe, here viewed from the disclosure angle. The actionable inversion: capturing AI’s gains depends less on raising adoption than on making disclosure safe and rewarding (stop taxing efficiency gains; reward adopted-workflow “multipliers”; legitimize experimentation via “side quests”). Connects directly to the micro-productivity-trap (the individual-level efficiency-tax that motivates hiding).

”Token capital” + AI-is-the-future-of-the-firm: the platform-CEO refounding mandate ( Possible, June 2026)

Nadella (Microsoft, June 2026) supplies the wiki’s strongest platform-CEO statement that AI adoption is a firm-redefinition problem, not a procurement one:

• Token capital as a balance-sheet-invisible asset. “This economy is going to be shaped by both human capital and… token capital.” A firm’s tacit knowledge — “the unique ways you operate, pass judgment, have taste” — is now extractable by models “through human trajectories” and encodable as weights/context/skills. The adoption mandate is to compound that into something “you own, you control, you created” — and not leak it, because “if you leak it, it’s a one-way door.” Nadella’s leakage mechanism (model companies “setting up gyms with rewards, employing employees who worked at your company previously”) makes the data-as-moat claim concrete at the firm-IP level — the CEO-altitude version of the AWS data-strategy your-data-is-your-context thesis.
• The refounding mandate: “AI is not a technology, it’s the future of the firm.” Unlike PC/cloud/mobile (“have an IT department deal with vendors and reduce cost”), this requires the CEO to answer “what’s your token capital?” concretely — and Nadella is “perplexed” that most non-tech CEOs “still haven’t woken up,” doing a press release and “pointing to eight agents they built.” This is the sharpest articulation in the corpus of the adoption-vs-capability lag: the durable move is building a firm-owned hill-climbing machine (“welcome the models in, they hill-climb inside a machine you control, your data is your context, you collect the traces, you don’t let it leak”), where the new IP is “getting clear about the evals and objectives you care about.” Directly convergent with [[2026-06-18-ramaswamy-mckinsey-every-company-software-company|Ramaswamy’s every company a software company]] and [[2026-06-18-dumra-mit-smr-dbs-everyone-an-innovator|Dumra’s everyone an innovator]] at the same CEO altitude.
• Sovereign AI for companies, not just countries. Nadella reframes sovereignty around comparative advantage (Ricardo): the best sovereignty is preserving the advantage embodied in your thriving companies, not firewalls/data-residency. The failure modes are symmetric — “go off frontier and that makes no sense; but depend on one frontier model and you’re not sovereign either” — resolved by “using models to hill-climb on your own, one firm at a time.” A useful extension of the concept’s adoption-strategy layer up to the national-economy scale.

The Dutch national-adoption datapoint: ahead of EU average, behind the leaders ( Rabobank, June 2026)

RaboResearch (June 2026) adds the wiki’s first Netherlands-specific enterprise-adoption reading, from a CBS enterprise survey (firms >10 employees, AI use by business function). The headline: Dutch IT and business-services firms use AI above the EU average in every business function, but below the EU leader in every function — “lopen voor op het Europees gemiddelde, maar behoren niet tot de kopgroep.” (E.g. for business services, administratieve processen NL 25% vs EU-leader 33% vs EU-avg 14%; marketing NL 16% vs 31% vs 10%.) This is a middle-of-the-pack-but-ahead-of-mean national position — consistent with the global high-adoption/low-maturity gap this page tracks, now with a country-level qualifier. RaboResearch frames the strategic stake as the same one Nadella and BCG name: firms that integrate AI into processes and skills “kunnen hun productiviteit structureel verhogen” while laggards “zullen terrein verliezen” — and it names three enabling conditions (strategic ownership; a reliable/governed data-IT-process base embedded in workflows; governance/trust) that line up with the dynamic-capabilities sensing/seizing/transforming microfoundations and with responsible-ai. The labour-side and task-level-productivity halves of the same study live on ai-employment-effects and micro-productivity-trap.

Debates and supersession

• High adoption vs. low maturity. 78% adoption + 1% mature (per AI Index) and 28%+34% in Stages 1–2 + only 7% Stage 4 (per MIT CISR) describe an organization-wide scramble in early innings. Sources interpreting 78% as “AI is mainstream” are technically correct but misleading about depth. See ai-maturity-measurement-comparison.
• Methodology divergence between sources. The AI Index uses McKinsey’s binary “use in ≥1 function” instrument; MIT CISR uses a triangulated AI-effectiveness score across operations, CX, and ecosystem support, banded into 4 stages. Different definitions of “mature” — 1% (AI Index, GenAI specifically, C-suite self-report) vs. 7% (MIT CISR, Stage 4, score-based). ai-maturity-measurement-comparison tracks the cross-walk.
• Replacement vs. augmentation. AI Index 2025 notes the workforce-reduction expectation is declining, hinting that early evidence supports augmentation (Jevons-paradox style demand expansion) over replacement. Open question: does that hold beyond 2024?
• Equalizing effect persistence. Robust in early studies. Open question: as AI tools mature, do high-skill workers eventually catch up by leveraging more sophisticated workflows? Or does the effect deepen as agents handle more of what high-skill humans currently do?
• Methodology drift. The AI Index 78% number comes from McKinsey’s instrument; whether year-over-year comparisons are apples-to-apples depends on McKinsey holding the survey design steady. The MIT CISR Stage distribution comes from the 2022 Future Ready Survey, predating the GenAI explosion — so the 28/34/31/7 numbers should be read as baseline, not current.
• Where the financial inflection is. MIT CISR locates it specifically at Stage 2 → Stage 3 (pilots → scaled ways of working). AI Index 2025’s function-level financial-impact data shows modest gains across all functions but doesn’t directly speak to between-stage gradients. The MIT CISR claim is a sharper, more actionable framing.

Related concepts

• generative-ai — the technology driving most of the 2024 adoption jump
• responsible-ai — the risk-management discipline orgs are scrambling to build alongside adoption
• foundation-models — what enterprise AI is increasingly built on top of
• ai-benchmarks — context for the capability claims that drive procurement decisions
• ai-knowledge-hiding — why high individual adoption need not produce collective gains: employees hide discovered workflows when organizational trust / psychological safety is low (Anicich & Brouwers 2026)
• founder-led-sales — the seller-side mirror: the adoption gap (AI interest that doesn’t convert to value) seen from the founder’s chair, where it reads as “mistaking attention for traction” (Rubinstein & Onyemah 2026)

“The year of increased accountability” — the editorial framing for 2026 ( O’Reilly, January 2026)

Julie Baron’s annual Signals for 2026 outlook gives the wiki its single best editorial-framing headline for the 2026 adoption-curve inflection: “Expect enterprises to shift focus from experimentation to measurable business outcomes and sustainable AI costs.” This is the Dutt et al. “experimentation-to-transformation” reframe compressed into a single phrase, written for an O’Reilly Radar trade-press audience three months before the HBR essay.

Baron also names the strategic-crisis framing explicitly: “Most companies have moved past simple AI experiments but are now facing a strategic crisis. Their existing product playbooks (sizing markets, roadmapping, UX) weren’t designed for AI-native products. Organizations must develop clear frameworks for building a portfolio of differentiated AI products, managing new risks, and creating sustainable value.” The wiki’s reading: 2026 is the year measurable-outcome obligations finally bite — the micro-productivity-trap becomes operationally untenable, and the buy framework becomes a forced choice.

USE / COMPOSE / BUILD as the AWS-vendor strategic-decision frame ( AWS London Exec Forum, 21 May 2026)

Jonathan Allen (AWS Executive in Residence) at the AWS London Executive Forum 2026 names the USE / COMPOSE / BUILD economic-decision framework as his headline strategic-altitude prescription for enterprise agentic-AI adoption — and reports that ~80% of his AWS customers are currently landing at COMPOSE (composing with frontier-model APIs — Haiku / Sonnet / Opus 4.7 — rather than training their own models). The framework is the agentic-system-economics overlay on Jassy’s three-layer AI stack from May 2025 (infrastructure / orchestration / applications) — same Amazon-vendor lens at 12-month delta. Allen’s framing of the deployment-failure mode mirrors the micro-productivity-trap: “AI bolted on is going to fail… pilots stay as demoed, never connected to actual business decision flows.” The empirical anchor Allen builds the prescription against is the MIT NANDA 95%-of-AI-pilots-fail report (Apr 2026, not yet ingested) — see the source page for the named failure modes (workflow misalignment, wrong targets, missing integration) and the 5%-pattern counter-prescription (workflow-level focus, right team composition, integrated into business decision flows from day one).

Brooklyn Solutions’ customer testimony in the same talk supplies a vendor-customer-side ratification of the workflow-level-focus prescription: their frequency × variability use-case discovery workshop method — sitting with SMEs from every department, identifying tasks that are “mundane, high-frequency, not that variable, that happen day in, day out” — is the operational complement to Allen’s strategic frame. Brooklyn’s “momentum beats perfection” closing aphorism is explicitly not permission to flood disposable apps — their 4-phase progression (basic use-cases → conversational Ask Brooklyn → agentic → multi-agent) is iterative quality compounding, the Brooklyn counter to the disposable-applications failure mode Allen names.

Interaction-design as the lever — the four-parameter framework + the negative-untailored finding (Krakowski et al. 2025 (Management Science))

Where most adoption frameworks in this concept page operate at the deployment-decision layer (where to use AI, which capabilities to deploy, which stage to be in), Krakowski et al. move one layer down to the within-deployment interaction-design layer. Their N=72 Nordic pharma sales experts DiD field experiment shows: holding the AI system constant, the configuration of work procedure, decision authority, training, and incentives — tailored to individual cognitive style per Kirton 1976 — is what determines whether the deployment is value-creating or value-destroying. Untailored deployment causes market share to gradually decline vs a stable legacy-IT baseline, with utilization in the untailored condition also gradually decreasing as sales experts withdraw from the system. The wiki’s first academic-RCT-grade evidence that deployment alone is not the lever; interaction-design is. Pairs structurally with Kropp et al.’s finding (the same year) that the framing of AI in the org (tool vs employee) also matters at the interaction-design layer — both papers argue the apparently-administrative organisational-design choices around AI deployment are causally consequential.

The anti-anthropomorphizing-AI prescription at consultant altitude ( BCG Henderson Institute, HBR May 2026)

BCG Henderson Institute’s RCT (N=1,261 HR/finance managers/directors/executives in US/Canada/EU) adds an explicit empirical anchor on how the framing of AI in enterprise org charts changes adoption outcomes: organisations that frame AI as employees (“AI on the org chart,” named AI teammates) experience accountability diffusion, escalation inflation, quality decline, and identity uncertainty without meaningfully increasing adoption intent. The five-point prescription — redefine workflows + name new human role expectations; make accountability explicit and personal; capability-building plan for managers of agents; don’t constrain agents into 1-for-1 roles (single agent across many workflows, multiple agents reshape one job); make deliberate choices about how human work evolves — is the wiki’s most concrete advisory-altitude operational prescription on agentic-AI organisational design. Population baseline from the experiment: 31% of participating orgs already frame AI as a teammate/employee; 23% have AI agents listed on org/work charts, across tech, healthcare, financial services, retail, and professional services — the practice is mainstream by mid-2026 even though the empirical evidence shows it is counter-productive.

Vendor-side productization of agentic adoption (May 2026 Radar Trends digest)

The May 2026 digest’s framing line — “AI is becoming operational” — names the vendor-side version of the same inflection: shared team agents as enterprise process automation instead of language-game demos. Five vendor-side moves toward harness-as-a-service in a single month (Anthropic Managed Agents + Claude Code routines + OpenAI workspace agents + OpenAI Agents SDK open-source + Amazon Bedrock AgentCore agent registry + Cursor 3 as orchestrator) signal that the substrate for enterprise agent deployment is being commoditised by the foundation-model vendors themselves. Direct implication for this concept page: the adoption-curve bottleneck is shifting from “can we get a working agent into production” (largely solved by HaaS) to “can we govern the agents we now have running in production” — the governance-and-evaluation layer becomes the binding constraint.