Research · Enterprise AI Transformation Programmes (2025–2026)
Back to researchResearch sweep · deep · 2025 – present
Enterprise AI Transformation Programmes (2025–2026)
Enterprise AI transformation programmes from July 2025 to July 2026: reported success and failure rates and who measures them, delivery frameworks borrowed from and diverging from classic digital transformation, token cost economics and budgeting under consumption pricing, adoption strategy including leadership over-provisioning of access, and AI governance across AI-Ops, financial business cases, security and regulatory controls as they scale with sector risk tolerance, referencing MIT, McKinsey State of AI, DORA, ThoughtWorks Technology Radar, NIST AI RMF, ISO 42001, and the EU AI Act.
Synthesised 2026-07-15
Enterprise AI Programmes, July 2025 to July 2026: The Year the Bill Arrived
Overview
Enterprise AI adoption is now effectively universal and enterprise AI value is still rare. McKinsey's 2025 State of AI survey, fielded across nearly 2,000 respondents in 105 countries, found 88 per cent of organisations using AI regularly in at least one function, while only around 6 per cent qualify as high performers attributing more than 5 per cent of EBIT to AI, and nearly two thirds have not begun scaling at all. That gap, not any single capability question, is the organising fact of the period. Sources: McKinsey / QuantumBlack (2025) (↗); CX Today (2026) (↗)
The defining shift of the past eighteen months was economic, not technical. The period opened with MIT NANDA's viral claim that 95 per cent of GenAI pilots show no measurable P&L impact, and closed with the Financial Times reporting that Amazon, Walmart, Cisco, Uber and Meta had imposed usage caps after AI spend outran budgets, Uber exhausting its 2026 allocation by April and capping individual tool use at $1,500 a month. Leadership postures inverted from "tokenmaxxing", the deliberate over-provisioning of AI access as an adoption forcing function, to CFO-led cost governance. Sources: Fortune (2025) (↗); Simon Willison's Weblog (2025) (↗); Substack (2026) (↗)
The second structural story is the copilot-to-agent transition and its collision with governance machinery built for deterministic software. McKinsey found 23 per cent of organisations scaling an agentic system somewhere and 39 per cent experimenting, while Forrester-style analyst work, MIT Sloan Management Review's adaptive-governance research, and vendor product launches (Google's Gemini Enterprise Agent Platform with Agent Registry, Gateway and Identity) all converge on the same diagnosis: agents create accountability, identity and audit problems with no clean precedent in the SaaS or RPA waves. Sources: McKinsey / QuantumBlack (2025) (↗); MIT Sloan Management Review (2026) (↗); Google Cloud Blog (2026) (↗)
The evidence base itself became a contested object. Vendor usage data (Anthropic's Economic Index, OpenAI's enterprise reports) shows accelerating depth of use; consultancy trackers show adoption without EBIT impact; independent academic work, including METR's randomised trial and NBER-style executive surveys, shows modest or even negative measured productivity effects. A reader who does not track who produced each number, and why, will get the story wrong. Sources: Anthropic (2026) (↗); OpenAI (2025) (↗); Nextword (Substack) (2026) (↗)
Timeline
- Gartner forecasts 40% of agentic AI projects cancelled by 2027
- MIT NANDA "95% of pilots fail" figure goes viral, then draws methodological fire
- METR RCT finds AI-assisted developers 19% slower while believing themselves faster
- DORA reframes AI as an amplifier of existing team quality
- McKinsey State of AI documents 88% adoption against 6% high performers
- EU Digital Omnibus proposes deferring high-risk AI Act deadlines
- OpenAI publishes first State of Enterprise AI report
- Deloitte and FinOps Foundation formalise tokenomics and FinOps for AI
- Anthropic Economic Index shows enterprise API use concentrating and heavily automated
- FT reports token caps at Amazon, Walmart, Cisco, Uber and Meta
- Digital Omnibus agreement pushes high-risk obligations to December 2027
- ThoughtWorks Radar v34 urges return to engineering fundamentals
- Google launches Gemini Enterprise Agent Platform with governance primitives
- MIT SMR publishes adaptive governance model from Barclays, Nasdaq and Microsoft interviews
- "Tokenmaxxing" reversal complete, token spend becomes a board-level metric
- EU AI Act Article 50 transparency deadline holds at August 2026
Key Findings
The 95 per cent figure survived as a signal and collapsed as a statistic. MIT NANDA's GenAI Divide report rested on 52 executive interviews, surveys of 153 leaders and 300 public deployments, with a six-month, P&L-only success definition. Paul Roetzer at Marketing AI Institute called it not statistically valid; Zvi Mowshowitz noted the report labels back-office efficiency gains failures because they lack an attributable dollar figure. Yet even critics accept the underlying pattern of high adoption and low transformation depth, and the report's secondary findings (purchased tools succeed roughly twice as often as internal builds, 90 per cent of workers use personal AI tools while only 40 per cent of firms hold official subscriptions) proved durable across lanes. Sources: Legal.io (2025) (↗); Marketing AI Institute (2025) (↗); Don't Worry About the Vase (Substack) (2025) (↗); Medium (2026) (↗)
Workflow redesign, not tooling, separates the winners. McKinsey's high performers are 2.8 times more likely to have fundamentally redesigned workflows (55 versus 20 per cent) and far more likely to run defined human-in-the-loop validation (65 versus 23 per cent). Bain's parallel finding that scaling rather than spend drives satisfaction, and O'Reilly's account of JPMorganChase, Walmart and Uber converging on a hybrid of centralised enablement with distributed execution, point at the same conclusion from three independent methodologies. Sources: CX Today (2026) (↗); O'Reilly Radar (2026) (↗)
DORA reframed AI as an amplifier, and the framing stuck. The 2025 State of AI-assisted Software Development report, built on nearly 5,000 respondents and over 100 hours of qualitative data, found 90 per cent of developers using AI and over 80 per cent reporting productivity gains, yet 30 per cent distrusting generated code. Its thesis that AI magnifies existing organisational strengths and dysfunctions, with platform engineering and data health as preconditions, has become the practitioner consensus, even as one independent blogger accused the 2026 edition of recycling the 2025 executive summary nearly verbatim. Sources: DORA / Google Cloud (2025) (↗); InfoQ (2026) (↗); makemeacto.cc (independent blog) (2026) (↗)
Token-maxing was tried, named, and reversed within a year. Jensen Huang publicly argued a $500,000 engineer should consume at least $250,000 in tokens, and CTO briefings reported token budgets pegged at up to half of engineering salary. FinOps commentators warned that measuring adoption by consumption creates an incentive to burn tokens, and by mid-2026 the reported outcomes were Amazon halting internal token leaderboards after an alleged $500 million Claude month and Uber blaming Anthropic for a blown budget without moving business KPIs. The mandate-versus-enablement question now has an answer of sorts: over-provisioning built habit and cost in roughly equal measure. Sources: Substack newsletter (CTO advisory interviews) (2026) (↗); FinOps Foundation (2026) (↗); Haverin (Substack) (2026) (↗)
Falling unit prices do not mean falling bills. Deloitte's tokenomics work, first published in the WSJ's CIO Journal, established that per-token prices are declining while total enterprise spend rises on consumption growth, with only 28 per cent of finance leaders expecting near-term ROI. Anthropic's own Economic Index quantifies the mechanism: enterprise API usage is overwhelmingly automated, concentrating on fewer tasks, with an input-to-output token elasticity of 0.38 that makes context, not capability, the binding constraint. AT&T reportedly scaled from 8 billion to 27 billion tokens per day after deploying multi-agent systems. Sources: Deloitte Insights (2026) (↗); Anthropic (2026) (↗); Mavvrik (2026) (↗)
Governance frameworks are layering, not competing, and none was built for agents. Practitioner guidance converges on ISO 42001 for certifiable management-system structure, NIST AI RMF for risk methodology, and the EU AI Act for binding high-risk obligations, with most enterprises operating under two or more simultaneously. All three predate agentic autonomy, leaving a documented gap for cascading tool-call failures that vendor platforms (agent registries, cryptographic agent identity) are racing to productise as a sales feature. Sources: NeuralTrust (2026) (↗); GAICC (2026) (↗); Google Cloud Blog (2026) (↗)
Regulatory bite scales with sector, not with the AI Act's calendar. The Digital Omnibus deferred stand-alone high-risk obligations to December 2027 while leaving Article 50 transparency and GPAI duties on the August 2026 timetable, a compliance trap for firms assuming blanket relief. Meanwhile sector regulators moved faster than horizontal law: UK FCA and PRA extend SS1/23 model-risk supervision to AI, US banking's SR 11-7 remains the reference point for agentic systems fifteen years after issuance, and Michigan's insurance regulator now requires a written AI Systems Programme referencing NIST AI RMF with no safe harbour. Sources: Substack (2026) (↗); Horizon Search Institute (2026) (↗)
ThoughtWorks called the methodological direction: backwards, deliberately. Radar volume 34 identified "cognitive debt" from AI-generated complexity and prescribed a return to pair programming, zero trust architecture, mutation testing and DORA metrics, alongside new agent-specific concerns such as the lethal trifecta of private data, untrusted content and external action. Martin Fowler's commentary framed this as necessary counterweight, not nostalgia. The genuinely novel practices, eval-driven development and human-in-the-loop gates, sit alongside inherited discipline rather than replacing it, exactly as Bloomberg CTO Shawn Edwards described evals as the binding constraint on production agents. Sources: Thoughtworks (2026) (↗); Martin Fowler's Bliki (2026) (↗); Medium (2026) (↗)
The perception gap is measurable and large. METR's randomised controlled trial found experienced developers took 19 per cent longer with AI tools while believing they were 20 per cent faster, and a survey of 5,000 white-collar workers found executives claiming eight-plus hours saved weekly while two thirds of non-management staff reported under two hours. Self-reported adoption surveys should be read against this calibration failure. Sources: Substack (2026) (↗); METR (2025) (↗)
Evidence & Data
The adoption ledger reads as follows. McKinsey: 88 per cent regular use, 39 per cent enterprise-level EBIT impact, roughly 6 per cent high performers, 23 per cent scaling agentic systems. MIT NANDA: over 80 per cent piloting, near 40 per cent deploying, 5 per cent of enterprise-grade custom systems reaching production. OpenAI's enterprise report: weekly Enterprise messages up roughly 8x since November 2024, seats up roughly 9x year on year, with the stated bottleneck now organisational readiness rather than model performance. Sources: McKinsey / QuantumBlack (2025) (↗); Virtualization Review (2025) (↗); OpenAI (2025) (↗); Enterprise AI Executive (2025) (↗)
On cost, the numbers are stark. Average enterprise token consumption rose thirteenfold year on year; the FinOps Foundation named SaaS-model token cost management its top practitioner challenge, driven by developer-led purchasing, opaque billing and no native allocation mechanisms. Benchmarkit's survey of 372 enterprises found only 15 per cent could forecast AI costs within 10 per cent of actuals, and nearly one in four missed by more than 50 per cent. Deloitte reports AI as the fastest-growing line in corporate technology budgets, consuming up to half of IT spend at some firms. Sources: FinOps Foundation (2026) (↗); Substack (2026) (↗); Deloitte Insights (2026) (↗)
The independent productivity evidence remains sobering. Beyond METR's 19 per cent slowdown, a February 2026 NBER survey of 6,000 executives found 90 per cent reporting no productivity impact over three years, against task-level studies showing 14 to 55 per cent gains and a 26 per cent weekly task-completion lift across nearly 5,000 developers at Microsoft and Accenture, concentrated among juniors. METR's capability data cuts the other way: autonomous task horizons doubling roughly every seven months since 2019, possibly compressing to four, with its 2026 Frontier Risk Report documenting agents inside AI companies working with permissions comparable to human employees. Sources: Nextword (Substack) (2026) (↗); METR (2025) (↗); METR (2026) (↗); METR (2026) (↗)
Signals & Tensions
Pilot purgatory versus production progress. Bain's Q3 2025 executive tracker found 80 per cent of GenAI use cases met or exceeded expectations and framed production deployment as accelerating faster than any prior technology wave, directly contradicting the MIT failure framing. The reconciliation is definitional: expectations-met is a softer bar than P&L attribution, and only 23 per cent of Bain's respondents could tie initiatives to revenue or cost. Which lane you believe depends on which denominator you accept. Sources: O'Reilly Radar (2026) (↗); Medium (2025) (↗)
Vendor usage data as quasi-independent measurement. Anthropic's Economic Index and OpenAI's enterprise reports are the richest quantitative records of actual usage, yet both labs profit from the adoption story and, under consumption pricing, from verbosity itself. Contracted evaluators like METR offer genuine independence but depend on lab cooperation for access. Sources: Anthropic (2025) (↗); METR (2026) (↗)
Cost as blocker or cost as proxy. The FT-reported caps read as unit-economics discipline, but Gartner's June 2025 analysis of agentic cancellations named escalating costs, unclear business value and inadequate risk controls, with model capability absent. Where ROI measurement is weak, cost caps may simply be the only lever a CFO can pull. Sources: Substack (2026) (↗); Substack (2026) (↗)
Shadow AI as the unpriced liability. Ninety per cent of workers use personal AI tools against 40 per cent official provision, and shadow token spend is now framed as a six-figure per-firm exposure. Compliance-led slowness manufactures the exact ungoverned usage it exists to prevent, a tension no surveyed governance council has yet resolved. Sources: Marketing AI Institute (2025) (↗); Ability.ai (2026) (↗)
Regulatory deferral versus planning certainty. Advisors recommend treating original AI Act deadlines as binding even while deferral is agreed, because Article 50 obligations held while Annex III slipped. Regulatory uncertainty has itself become a planning variable, not a fixed constraint. Sources: GAICC (2026) (↗); Substack (2026) (↗)
Report inflation in the research supply chain. The same headline statistics recirculate across SEO aggregators with inconsistent attribution, and even DORA drew accusations of repackaging its own thesis year on year. The volume of governance maturity models on arXiv, validated by simulation rather than field data, outpaces longitudinal outcome evidence by an order of magnitude. Sources: makemeacto.cc (independent blog) (2026) (↗); arXiv (2025) (↗)
Open Questions
Whether the enterprise-level productivity signal ever appears in aggregate statistics. Task-level gains of 14 to 55 per cent coexist with BLS data showing no national productivity signature and NBER executives reporting nothing after three years. The complementary-investment lag explanation is plausible but untested at current timescales. Sources: Nextword (Substack) (2026) (↗)
Whether agentic token growth outruns per-token deflation. SiliconANGLE-tracked infrastructure valuations rest on the bet that it does; enterprise budget caps suggest CFOs are betting it does not, or at least refusing to fund the difference. Sources: The Strategy Stack (Substack) (2026) (↗)
Who owns an agent's mistakes. Forrester-style governed-identity playbooks, Google's agent identity primitives and MIT SMR's adaptive-governance tiers all propose answers, but no jurisdiction has tested liability for an autonomous agent's action in a regulated process. Sources: MIT Sloan Management Review (2026) (↗); Let's Data Science (2026) (↗)
Whether ISO 42001 certification predicts anything. Anthropic certified early and vendors increasingly sell against the badge, but no independent study yet links certification to lower incident rates or better outcomes. Sources: Tygart Media (2026) (↗)
Whether FinOps for AI can achieve cloud-FinOps maturity given consumption that varies with model choice, context length, reasoning depth and loop count, none of which the enterprise controls. Forecast misses above 50 per cent for a quarter of firms suggest not yet. Sources: FinOps Foundation (2026) (↗); Substack (2026) (↗)
Whether the buy-over-build finding survives the agent era. MIT NANDA found purchased tools succeeding twice as often as internal builds, but agentic orchestration pushes differentiating logic back in-house, and the platform consolidation race may reset the ratio. Sources: Legal.io (2025) (↗); Haverin (Substack) (2026) (↗)
Whether health-sector agentic deployment gets ahead of accountability. Three in four health plans use AI in prior authorisation with Medicare Advantage appeal overturn rates above 80 per cent and patient appeal rates below 1 per cent, a gap between deployment and recourse that no current framework addresses. Sources: Horizon Search Institute (2026) (↗)
The year's punchline is that the industry spent 2025 arguing about whether AI programmes fail and 2026 discovering what they cost. The next argument, already visible in the sector-regulator filings and the agent-identity product launches, is about who answers when they act.
![[sources-enterprise-ai-transformation-programmes-from-july-]]
Sources
Summary: ↑ Back to summary
Financial Press
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| f1 | MIT report: 95% of generative AI pilots at companies are failing | Fortune | 2025-08 | Original business-press write-up of MIT NANDA's GenAI Divide report, the most widely cited failure-rate statistic in the enterprise AI debate, with direct quotes from lead author Aditya Challapally. |
| f2 | That Viral MIT Study Claiming 95% of AI Pilots Fail? Don't Believe the Hype. | Marketing AI Institute | 2025-11 | Provides the methodological pushback on the MIT 95% figure, arguing the definition of success is narrow and the finding has been oversimplified in coverage. |
| f3 | The State of AI: Global Survey 2025 | McKinsey & Company / QuantumBlack | 2025-11 | Primary source for the 88% adoption vs 6% high-performer EBIT-impact gap, the central adoption-versus-value statistic cited across the sector, based on 1,993 respondents in 105 countries. |
| f4 | The state of AI in 2025: Agents, innovation, and transformation | McKinsey & Company | 2025-11 | Full survey PDF detailing the 12 gen AI adoption and scaling best practices and the correlation between workflow redesign and EBIT impact that underpins most delivery-framework guidance in the market. |
| f5 | McKinsey State Of AI In 2025: What It Means For CX | CX Today | 2026-02 | Distils McKinsey's direct-quoted findings that high performers are 2.8x more likely to have redesigned workflows and far more likely to have human-in-the-loop validation processes. |
| f6 | CFOs Are Coming For The Enterprise AI Budget | Forbes | 2026-06 | Cites Financial Times reporting on Amazon, Walmart, Cisco, Uber and Meta capping internal AI usage, plus Gartner's revised figure that 50% of generative AI proof-of-concepts were abandoned by end of 2025. |
| f7 | Amazon, Walmart and Uber curb employee AI use as costs surge | Crypto Briefing (reporting on Financial Times investigation) | 2026-06 | Summarises the Financial Times' investigation naming five major companies imposing token caps, including Uber's $1,500-per-month limit after exhausting its 2026 AI budget by April. |
| f8 | AI Giants Face A Potential Cost Meltdown | Forbes | 2026-05 | Draws on Wall Street Journal reporting that OpenAI missed revenue targets amid data-centre spending concerns, and Morgan Stanley's figure of $740bn in 2026 tech capex, framing AI economics as a market risk. |
| f9 | Wall Street's AI Spending Warning Could Make Nvidia the Biggest Casualty | Yahoo Finance (Wall Street analyst commentary) | 2026 | Contrasts Bain's finding that many companies are not earning expected productivity gains with Goldman Sachs' view that AI capex forecasts remain too conservative, capturing the investor-side disagreement over AI ROI. |
| f10 | How Enterprises Can Control AI Token Costs | Boston Consulting Group | 2026 | Practitioner framework for FinOps-style AI cost governance, recommending routing tasks to appropriately sized models and assigning P&L ownership to token spend. |
| f11 | AI tokens: How to navigate AI's new spend dynamics | Deloitte Insights / Wall Street Journal CIO Journal | 2026-01 | Originally published in the Wall Street Journal's CIO Journal; establishes that unit token prices are falling while total enterprise spend rises due to consumption growth, and finds only 28% of finance leaders expect near-term ROI. |
| f12 | [AI Act | Shaping Europe's digital future](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) | European Commission | 2026-07 |
| f13 | The EU AI Act – the current state of play | Travers Smith | 2026-05 | Law firm briefing detailing the Digital Omnibus's postponement of high-risk deadlines and the Commission's May 2026 draft guidelines on high-risk system classification, key for understanding regulatory bite by sector. |
| f14 | After 'Tokenmaxxing', Token Spend Has Become The New Metric To Watch | Forbes | 2026-07 | Documents the reversal from leadership-encouraged 'tokenmaxxing' to cost scrutiny, citing Uber's budget exhaustion and Microsoft's move away from Claude Code amid billing-visibility concerns. |
| f15 | Agentic AI 2026 Outlook | Bloomberg Professional Services | 2026-05 | Bloomberg Intelligence analysis of how agentic AI is shifting software pricing from seat-based subscriptions to usage- and outcome-based models, with implications for SaaS valuations. |
| f16 | A Startup That Builds AI Agents Used One to Raise $100 Million | Bloomberg | 2026-07 | Concrete enterprise case study of agentic AI moving beyond pilot into a functioning business process (fundraising outreach), illustrating the adoption-beyond-pilot narrative financial press is tracking. |
| f17 | Lessons in how to build AI agents from Bloomberg CTO Shawn Edwards | Fortune | 2026-04 | First-person account from Bloomberg's own CTO on why eval-driven development, not model quality, is the binding constraint on enterprise agent deployment, directly relevant to the framework-lineage question. |
| f18 | AI Productivity's $4 Trillion Question: Hype, Hope, And Hard Data | Forbes | 2026-01 | Synthesises the bifurcated evidence base, task-level productivity gains of 14-55% against aggregate Bureau of Labor Statistics data showing no clear AI signature in national productivity figures. |
| f19 | The AI Productivity Paradox | The Information Difference | 2026-04 | Cites a February 2026 NBER survey of 6,000 executives finding 90% report no productivity impact over three years, and a Wall Street Journal finding that 40% of workers report no time saved from AI. |
| f20 | WRITER 2026 AI adoption in the Enterprise survey | WRITER (with Workplace Intelligence) | 2026-05 | Independent 2,400-respondent survey finding 79% of organisations face adoption challenges and 67% of executives believe their company has already suffered a data leak from unapproved AI tools, evidencing the shadow-AI governance tension. |
| f21 | State of agentic AI adoption survey [2026] | Zapier | 2025-12 | 500-enterprise-leader survey showing human-in-the-loop remains the dominant governance pattern (38%) even as agent deployment accelerates, evidencing the control-versus-velocity trade-off. |
| f22 | AI Governance for Financial Services: FCA & PRA 2026 | SureCloud | 2026-05 | Details how UK financial regulators (FCA, PRA) extend existing model-risk supervisory statements such as SS1/23 to AI systems, illustrating sector-specific regulatory bite ahead of dedicated AI legislation. |
| f23 | Artificial Intelligence, Domain AI Readiness, and Firm Productivity | arXiv (academic working paper) | 2025-08 | Academic empirical treatment of the AI productivity paradox, examining organisational and data-readiness complements that determine whether AI investment translates into firm performance. |
Tech Industry & Practitioner
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| p1 | MIT report: 95% of generative AI pilots at companies are failing | Fortune | 2025-08 | Primary journalistic account of the MIT NANDA GenAI Divide report and interview with lead author Aditya Challapally, the most widely cited failure-rate statistic of the period. |
| p2 | MIT Report Finds 95% of AI Pilots Fail to Deliver ROI, Exposing "GenAI Divide" | Legal.io | 2025-08 | Detailed breakdown of the report's build-vs-buy, shadow AI economy, and back-office ROI findings with specific dollar figures. |
| p3 | That Viral MIT Study Claiming 95% of AI Pilots Fail? Don't Believe the Hype. | Marketing AI Institute | 2025-11 | Key methodological critique of the MIT report's sample size, success definition, and six-month ROI window, representing the independent-scrutiny counter-narrative. |
| p4 | The MIT "95% of GenAI Pilots Fail" Report: What It Gets Wrong and What Leaders Should Do Instead | Medium | 2025-08 | Detailed methodological dissection arguing the report's six-month measurement window undercounts enterprise change timelines. |
| p5 | [DORA | State of AI-assisted Software Development 2025](https://dora.dev/dora-report-2025/) | DORA / Google Cloud | 2025-09 |
| p6 | Announcing the 2025 DORA Report | Google Cloud Blog | 2025-09 | Summarises DORA's central 'amplifier' thesis and the seven-capability AI Capabilities Model with adoption and trust statistics. |
| p7 | AI Is Amplifying Software Engineering Performance, Says the 2025 DORA Report | InfoQ | 2026-03 | InfoQ's practitioner-facing synthesis of the DORA report emphasising platform engineering and data ecosystem health as preconditions for AI value. |
| p8 | How are developers using AI? Inside Google's 2025 DORA report | Google Blog | 2025-09 | Reports the 90% AI adoption figure among developers and the 'trust paradox' between usage and confidence in AI-generated code. |
| p9 | [Technology Radar | Guide to technology landscape](https://www.thoughtworks.com/radar) | Thoughtworks | 2026-04 |
| p10 | As AI Accelerates Software Complexity, Thoughtworks Technology Radar Urges a Return to Engineering Fundamentals to Combat Cognitive Debt | Thoughtworks | 2026-04 | Press release with CTO Rachel Laycock quote framing the AI inflection point as a technique problem, not a technology problem. |
| p11 | [Macro trends in the tech industry | April 2026](https://www.thoughtworks.com/en-us/insights/blog/technology-strategy/macro-trends-tech-industry-april-2026) | Thoughtworks | 2026-04 |
| p12 | Thoughtworks Launches Agent/works™ to Govern and Run Enterprise AI Agents Across Any Cloud | PR Newswire / Thoughtworks | 2026-06 | Documents the shift from 2025 experimentation to 2026 operational governance reality, including agent sprawl and AI spend visibility problems. |
| p13 | Token Economics: The Atomic Unit of AI Value | FinOps Foundation | 2026 | Authoritative FinOps Foundation analysis of how token-based consumption pricing breaks traditional cloud cost forecasting models. |
| p14 | Token Economics: Managing AI Value in SaaS Model Token Costs | FinOps Foundation | 2026-06 | Working group output identifying token cost management as the top practitioner challenge and proposing chargeback and commitment-based pricing practices. |
| p15 | FinOps for AI Overview | FinOps Foundation | 2026-02 | Foundational FinOps guidance on usage limits, quotas, and anomaly detection for AI token spend governance. |
| p16 | FinOps X 2026 Recap: AI Tokenomics Explained | Mavvrik | 2026-06 | Reports the FinOps Foundation's mission expansion from 'cloud value' to 'technology value' and new consumption metrics proposed at FinOps X 2026. |
| p17 | AI tokens: How to navigate AI's new spend dynamics | Deloitte Insights | 2026-01 | Deloitte's practitioner-facing analysis of falling per-token prices versus rising aggregate enterprise AI spend, with concrete IT budget share figures. |
| p18 | AI token spend: the $150k shadow AI crisis | Ability.ai | 2026-05 | Documents the 'token-maxing' adoption philosophy attributed to Nvidia CEO Jensen Huang and its risks for mid-market enterprises. |
| p19 | The State of AI: Global Survey 2025 | McKinsey / QuantumBlack | 2025-11 | McKinsey's flagship annual survey of nearly 2,000 organisations, the most cited independent cross-industry ledger of AI adoption versus enterprise value capture. |
| p20 | McKinsey State Of AI In 2025: What It Means For CX | CX Today | 2026-02 | Practitioner synthesis of McKinsey findings on workflow redesign and human-in-the-loop validation as differentiators of AI high performers. |
| p21 | The Emerging Agentic Enterprise: How Leaders Must Navigate a New Age of AI | MIT Sloan Management Review / BCG | 2025-11 | Joint MIT SMR and BCG research report on operating-model tensions in agentic AI transformation, based on named executive interviews. |
| p22 | Scaling AI With Adaptive Governance | MIT Sloan Management Review | 2026-06 | Multi-year interview-based research at named financial institutions on how AI governance controls scale with system risk type, directly addressing sector risk tolerance. |
| p23 | Global AI Governance Comparison 2026: EU AI Act vs NIST AI RMF vs ISO/IEC 42001 | GAICC | 2026 | Detailed practitioner crosswalk showing how the three leading governance frameworks converge into a layered compliance stack, and their gap on agentic AI. |
| p24 | AI Governance Frameworks Compared: NIST vs ISO 42001 vs EU AI Act | NeuralTrust | 2026 | Explains the mandatory-vs-voluntary and certifiable-vs-non-certifiable distinctions driving enterprise framework selection in 2026. |
| p25 | The Enterprise AI Governance Framework: What You Need Before You Scale AI | EW Solutions | 2026-05 | Details the 2026 EU AI Act Digital Omnibus timeline shift and quantifies the AI governance platform spending market via Gartner forecasts. |
Academic & arXiv
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| a1 | MIT report: 95% of generative AI pilots at companies are failing | Fortune | 2025-08 | Primary journalistic account of MIT NANDA's "GenAI Divide" report and its headline 95% failure statistic, including methodology details and lead author caveats about self-reported failure data. |
| a2 | MIT Report Finds 95% of AI Pilots Fail to Deliver ROI, Exposing "GenAI Divide" | Legal.io | 2025-08 | Details the report's sample (52 interviews, 153 leader surveys, 300 public deployments) and the divide between high pilot adoption and low production conversion. |
| a3 | Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity | arXiv / METR | 2025-07 | METR's landmark RCT finding AI tools slowed experienced developers by 19% despite developers perceiving a speedup, the most rigorous empirical challenge to vendor productivity claims. |
| a4 | Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (blog) | METR | 2025-07 | METR's own plain-language write-up of the RCT, including its update noting a February 2026 follow-up on late-2025 AI tools. |
| a5 | HCAST: Human-Calibrated Autonomy Software Tasks | METR | 2025 | Foundational benchmark paper describing 140 human baseliners' 563 attempts across software engineering, ML and cybersecurity tasks, underpinning METR's capability measurement methodology. |
| a6 | Evaluating frontier AI R&D capabilities of language model agents against human experts (RE-Bench) | METR | 2024-11 | Introduces RE-Bench, comparing frontier model agents to 71 human expert attempts on ML research engineering tasks, a core benchmark for measuring AI R&D automation. |
| a7 | Measuring AI Ability to Complete Long Tasks | arXiv / METR | 2025-12 | Introduces the 50%-task-completion time horizon metric and documents a seven-month doubling time in AI capability from 2019-2025, widely used as a capability-trend proxy. |
| a8 | Time Horizon 1.1 | METR | 2026-01 | Updates the time-horizon methodology with an expanded 228-task suite, showing continued exponential capability growth through 2025-2026. |
| a9 | The Rapid Adoption of Generative AI (NBER Working Paper) | NBER / SSRN | 2024-09 | Nationally representative survey finding nearly 40% of the US working-age population uses generative AI, with work adoption as fast as the PC, providing an independent adoption baseline against consultancy figures. |
| a10 | Generative AI at Work (NBER Working Paper) | NBER | 2023 | Micro-level firm study by Brynjolfsson and Li on generative AI deployment in customer service, a foundational field study on productivity heterogeneity by worker experience. |
| a11 | Shifting Work Patterns with Generative AI (NBER Working Paper w33795) | NBER | 2025 | Reviews the empirical literature on generative AI's effect on work patterns and explains why workplace adoption of new AI capabilities takes time even after individual productivity gains are demonstrated. |
| a12 | The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers | SSRN | 2025-08 | Large-scale field experiment across nearly 5,000 developers at Microsoft, Accenture and a Fortune 100 firm finding a 26.08% increase in weekly task completion, with larger gains for junior developers. |
| a13 | The Labor Market Effects of Generative Artificial Intelligence | SSRN | 2026-01 | Comprehensive worker survey finding 35.9% of US workers used generative AI by December 2025 and identifying small positive wage effects with no aggregate employment decline, countering degradation narratives. |
| a14 | An Empirical Study of Measurement Framework Adoption - DORA and SPACE: How Organizational Context Shapes Success and Failure | SSRN | 2025-07 | Examines how organizational culture, size and engineering maturity shape success or failure of DORA and SPACE measurement framework adoption, directly relevant to how classic DevOps metrics are being ported into AI delivery. |
| a15 | A Framework for the Adoption and Integration of Generative AI in Midsize Organizations and Enterprises (FAIGMOE) | arXiv | 2025-10 | Proposes a structured adoption framework for generative AI in mid-size enterprises, synthesising pilot-to-scale considerations and citing prior DORA/SPACE and pilot-scaling literature. |
| a16 | The Unified Control Framework: Establishing a Common Foundation for Enterprise AI Governance, Risk Management and Regulatory Compliance | arXiv | 2025-03 | Widely cited foundational governance paper proposing a unified control architecture spanning enterprise AI governance, risk and regulatory compliance, referenced across subsequent governance maturity papers. |
| a17 | Governance Frameworks for Enterprise AI Systems Operating in Regulated Environments | International Journal of Computer Applications | 2025 | Synthesis finding that data governance and cybersecurity practices are relatively mature while lifecycle governance and oversight of agentic AI systems remain weak despite regulatory mandates. |
| a18 | Engaged AI Governance: Addressing the Last Mile Challenge Through Internal Expert Collaboration | arXiv | 2026-04 | Empirical study of how development teams operationalize EU AI Act-style regulatory requirements through collaborative workshops, addressing the scarce empirical literature on governance implementation versus principle-level frameworks. |
| a19 | AI Agents Under EU Law: A Compliance Architecture for AI Providers | arXiv | 2026-04 | Analyses how agentic AI's non-human identities and privileged system access create compliance challenges under EU law that traditional IAM frameworks were not designed to handle. |
| a20 | AI Governance Frameworks: ISO/IEC 42001, NIST AI RMF, and the EU AI Act | SSRN | 2026-05 | Systematic synthesis of the three principal AI governance frameworks as complementary rather than substitutive, incorporating the EU AI Act's November 2025 Digital Omnibus amendment. |
| a21 | Governing the Agentic Enterprise: A Governance Maturity Model for Managing AI Agent Sprawl in Business Operations | arXiv | 2026-03 | Proposes a five-level Agentic AI Governance Maturity Model grounded in NIST AI RMF and ISO/IEC 42001, citing industry data that only 21% of enterprises have mature agent governance while 40% of agentic AI projects are projected to fail by 2027. |
| a22 | Governing the Agentic Enterprise: A New Operating Model for Autonomous AI at Scale | California Management Review | 2026-03 | Proposes the Agentic Operating Model, arguing agentic AI is an institutional shift requiring new governance layers distinct from earlier copilot-era operating models. |
| a23 | Runtime Governance for AI Agents: Policies on Paths | arXiv | 2026-03 | Cites a 2026 KPMG survey finding 75% of large-enterprise leaders rank security, compliance and auditability as the most critical requirements for agent deployment, framing runtime governance as the key bottleneck. |
| a24 | AI Trust OS: A Continuous Governance Framework for Autonomous AI Observability and Zero-Trust Compliance in Enterprise Environments | arXiv | 2026-04 | Proposes telemetry-first continuous governance to replace manual attestation-based compliance workflows, addressing supply-chain AI governance requirements emerging under the EU AI Act. |
| a25 | AI Agentic Programming: A Survey of Techniques, Challenges, and Opportunities | arXiv | 2025-08 | Surveys token-cost and pricing dimensions of agentic coding workflows, quantifying how reasoning strategies and tool-augmented loops multiply token consumption relative to single-turn use. |
| a26 | Measuring AI R&D Automation | arXiv | 2026-03 | Related-work synthesis situating METR's RE-Bench alongside SWE-bench, MLE-bench and PaperBench within the broader AI R&D automation benchmark landscape. |
| a27 | PostTrainBench: Can LLM Agents Automate LLM Post-Training? | arXiv | 2026-03 | Benchmark extending AI R&D automation evaluation beyond METR's RE-Bench, finding newer models substantially outperform Anthropic's Sonnet 4.5 evaluation on end-to-end training-loop automation. |
VC & Analyst Reports
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| v1 | MIT report: 95% of generative AI pilots at companies are failing | Fortune | 2025-08 | Original popularisation of MIT NANDA's GenAI Divide report and its 95% no-P&L-impact figure, the most cited failure statistic of the period. |
| v2 | MIT Report Finds 95% of AI Pilots Fail to Deliver ROI, Exposing 'GenAI Divide' | Legal.io | 2025-08 | Provides the underlying methodology detail (52 executive interviews, 153 leader surveys, 300 deployments) needed to scrutinise the headline 95% figure. |
| v3 | That Viral MIT Study Claiming 95% of AI Pilots Fail? Don't Believe the Hype. | Marketing AI Institute | 2025-11 | Direct methodological pushback on the MIT statistic, arguing narrow definitions of success inflate the failure rate. |
| v4 | The state of AI in 2025: Agents, innovation, and transformation | McKinsey QuantumBlack | 2025-11 | McKinsey's flagship annual survey; source of the 88% adoption, one-third-scaling, and Rewired six-dimension operating model framework central to the delivery-approach debate. |
| v5 | McKinsey State Of AI In 2025: What It Means For CX | CX Today | 2026-02 | Extracts McKinsey's workflow-redesign and human-in-the-loop statistics (2.8x, 65% vs 23%) that distinguish AI 'high performers' from the rest. |
| v6 | Your AI Budget Is Growing. Your Returns Aren't. Here's Why. | Bain & Company | 2026-06 | Bain's 951-company Automation and AI Pathfinder survey documenting cost-savings shortfalls, circular self-funding of agentic AI, and governance ownership gaps. |
| v7 | How Companies Create Value with AI: Redesign, Not Tools | Bain & Company | 2026-06 | Bain finding that fewer than 20% of enterprises have scaled generative AI meaningfully, and that scaling (not spend) drives satisfaction with results. |
| v8 | Executive Survey: AI Moves from Pilots to Production | Bain & Company | 2026-04 | Bain's Q3 2025 survey offering a counter-narrative to 'pilot purgatory', showing rising production-scale deployment across domains and 80% expectation-met rate. |
| v9 | Survey: Generative AI's Uptake Is Unprecedented Despite Roadblocks | Bain & Company | 2025-10 | Longitudinal Bain quarterly tracker showing adoption acceleration to 95% of US firms alongside persistent security, quality and leadership-support concerns. |
| v10 | Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 | Gartner | 2025-06 | Origin of the widely cited 40% agentic-AI-cancellation forecast, plus Gartner's estimate that only ~130 of thousands of 'agentic AI' vendors are genuine. |
| v11 | Why 40% Of Agentic AI Projects May Be Canceled By 2027 | Forbes | 2026-07 | One-year-later revisit of the Gartner forecast, reframing cancellations as a governance and accountability failure rather than a model-capability one. |
| v12 | Where Enterprises are Actually Adopting AI | Andreessen Horowitz | 2026-04 | a16z's GDPval-benchmarked analysis of which sectors and use cases (coding, support, search) show real revenue momentum versus theoretical model capability. |
| v13 | Leaders, gainers and unexpected winners in the Enterprise AI arms race | Andreessen Horowitz | 2026-02 | a16z/Yipit panel data on multi-model orchestration (81% now run 3+ model families) and token-intensive use cases driving enterprise wallet share shifts. |
| v14 | As AI Accelerates Software Complexity, Thoughtworks Technology Radar Urges a Return to Engineering Fundamentals to Combat Cognitive Debt | Thoughtworks | 2026-04 | Vol.34 Technology Radar press release identifying 'putting coding agents on a leash' and a return to DORA metrics and zero-trust architecture as counterweights to agentic complexity. |
| v15 | Key themes in Technology Radar Vol.34 | Thoughtworks | 2026-04 | Explains the Radar's four organising themes, including the difficulty of evaluating fast-moving agentic tooling and 'securing permission-hungry agents'. |
| v16 | State of AI Q1'26 Report | CB Insights | 2026-04 | Documents record $226B quarterly AI funding concentration and the shift of capital toward physical AI and frontier model developers over enterprise application layers. |
| v17 | The Future of the Enterprise AI Buildout | CB Insights | 2026-04 | Shows enterprise AI partnership activity concentrated among five incumbents (32% of activity) and that 67.8% of startup partnerships remain ecosystem-building rather than revenue-generating. |
| v18 | AI 100: The most promising artificial intelligence startups of 2026 | CB Insights | 2026-05 | Flags the emerging agent-identity and accountability gap (no persistent identity, no audit trail for non-human actors) driving new AI governance infrastructure investment. |
| v19 | The State Of Agentic AI In 2026: Companies Are Chasing, Few Are Catching | Forrester | 2026-06 | Forrester survey data showing three-quarters of enterprises adopting agentic AI but few reaching genuine scaled production, with a governed-identity control-plane playbook. |
| v20 | Predictions 2026: AI Moves From Hype To Hard Hat Work | Forrester | 2025-10 | Forrester's 2026 prediction that enterprises will delay 25% of AI spend into 2027, with only 15% of decision-makers reporting EBITDA lift. |
| v21 | EU AI Act High-Risk Deadline: Enterprise Readiness Gap | Cloud Security Alliance | 2026-03 | Practitioner readiness assessment of the EU AI Act's August 2026 high-risk deadline and the enterprise compliance gap ahead of it. |
| v22 | The Digital AI Omnibus: Proposed deferral of high risk AI obligations under the AI Act (update) | DLA Piper | 2026-06 | Tracks the legislative deferral of EU AI Act high-risk obligations from August 2026 to December 2027, key to understanding regulatory pacing versus rollout speed. |
| v23 | NIST AI RMF or ISO 42001? | The Cyber Leader (Balanced Security) | 2026-05 | Practitioner comparison clarifying that ISO 42001 is a certifiable management system while NIST AI RMF is a self-attestation taxonomy, and how enterprises sequence the two. |
| v24 | AI Governance Frameworks Compared: NIST vs ISO 42001 vs EU AI Act | NeuralTrust | 2026 | Synthesises how most enterprises in 2026 operate under two or more governance frameworks simultaneously, layering OECD principles, NIST, ISO 42001 and the EU AI Act. |
| v25 | The token economy: The state of AI mid-2026 | SiliconANGLE | 2026-07 | Mid-2026 assessment of token-infrastructure valuations and the bet that rising agentic token volume will outrun per-token price deflation. |
| v26 | AI Tokenomics: How Token-Based Pricing Is Reshaping Enterprise AI Strategy | BizTech Magazine (CDW) | 2026-07 | Describes the emerging 'AI factory' infrastructure model and CIO pressure to tie token consumption directly to measurable ROI as agentic workloads scale. |
Blogs & Independent Thinkers
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| b1 | Reports Of AI Not Progressing Or Offering Mundane Utility Are Often Greatly Exaggerated | Don't Worry About the Vase (Substack) | 2025-08 | Zvi Mowshowitz's detailed line-by-line critique of the MIT GenAI Divide report's internal inconsistencies and definition of failure, the most substantive independent rebuttal in the lane. |
| b2 | GPT-5: Key characteristics, pricing and model card | Simon Willison's Newsletter (Substack) | 2025-08 | Primary practitioner documentation of frontier model pricing structure underpinning enterprise token-cost budgeting decisions. |
| b3 | Simon Willison on llm-pricing | Simon Willison's Weblog | 2025-2026 | Running first-hand log of token pricing changes and corporate token-budget incidents (Uber, Anthropic profitability) from the most cited independent LLM-tooling blogger. |
| b4 | Understanding Tokens: The Hidden Economics of AI Models for AI Governance Professionals | AI Governance Lead (Substack) | 2026 | Explains token cost mechanics (input vs output pricing, caching) specifically framed for AI governance and budget-setting decisions. |
| b5 | Almost Timely News: 18 Ways To Save AI Token Budgets | Almost Timely (Substack) | 2026-05 | Coined and critiqued 'token maxing' as a flawed adoption metric, arguing usage-as-proxy-for-adoption incentivises inefficiency. |
| b6 | [How to Drive AI Adoption | Lessons From 21 GTM Leaders](https://thegtmnewsletter.substack.com/p/how-to-drive-ai-adoption-gtm-leaders) | The GTM Newsletter (Substack) | 2026 |
| b7 | This Week in AI #157: Enterprise AI, AI Infrastructure & Anthropic Export Controls (CW26) | The Strategy Stack (Substack) | 2026-06 | Surveys enterprise AI budget-survey data challenging the token-cost-as-blocker narrative and tracks shift from proving value to building governance. |
| b8 | The real, embarrassing state of enterprise AI adoption | Nextword (Substack) | 2026-06 | First-hand documentation of Amazon and Uber token-spend failures and a critique of 'tokenmaxxing' as a vanity metric distinct from real value capture. |
| b9 | The Agentic Stack Wars: Part Three - EXTRACTION | Haverin (Substack) | 2026-06 | Detailed FinOps-for-AI analysis including Benchmarkit and GAO survey data on enterprises' inability to forecast AI infrastructure costs. |
| b10 | Dispatches From the Enterprise AI Frontline | Substack newsletter (CTO advisory interviews) | 2026-06 | First-hand interviews with CTOs at American Express, Gap, and Warner Bros. Discovery on token budgeting and the shift from free-form to explicit AI operating budgets. |
| b11 | The End of Free Money in AI: Why the Industry Is Moving to Cost-Accounting Mode | Substack | 2026-06 | Documents named corporate token-budget blowouts (Uber, Amazon, Meta, Microsoft) and quantifies the gap between falling per-token price and rising per-task consumption. |
| b12 | What should one think of the 2026 DORA brochure? | makemeacto.cc (independent blog) | 2026-05 | Sharp independent critique arguing DORA's 2026 ROI report substantially recycles its 2025 findings without materially new evidence. |
| b13 | AI Insights from the 2025 DORA Report | Substack | 2025-10 | Independent synthesis of DORA's 2025 findings connecting them to Stack Overflow sentiment decline and platform-engineering best practice. |
| b14 | Escaping AI Pilot Purgatory: Building an Agentic AI Operating System for the Next-Gen Enterprise | Brian Solis (Substack) | 2025-11 | Independent analyst framing of the governance-versus-innovation tension and centre-of-excellence versus federated operating model debate. |
| b15 | Pilot Purgatory | Substack | 2026-04 | Consultant's first-hand account of repeated CIO claims of AI maturity contradicted by outcome data, citing McKinsey, Deloitte and a METR RCT on developer productivity illusions. |
| b16 | Beyond Pilot Purgatory | O'Reilly Radar | 2026-02 | Independent technical-publisher analysis diagnosing pilot-stage failure as an organisational-design problem and proposing a hybrid centralised/federated operating model. |
| b17 | Fragments: April 21 | Martin Fowler's Bliki | 2026-04 | Canonical independent software-engineering commentary on Thoughtworks Technology Radar vol.34's themes of agent permissioning, harness engineering and cognitive debt. |
| b18 | Thoughtworks Tech Radar Recap: The AI Refresher | Medium | 2026-04 | Independent practitioner recap of eval-driven and spec-driven development patterns emerging from the Technology Radar, mapped to agentic delivery practice. |
| b19 | 95% of Corporate Generative AI Projects Fail, MIT Study Finds | Medium | 2026-03 | Independent Medium analysis reframing the MIT findings around organisational 'learning gap' and augmented-intelligence strategy rather than technology failure. |
| b20 | AI and the Fundamentals of Software Engineering - What the 2025 DORA Report Really Tells Us | Medium | 2025-10 | Cross-references DORA findings with Kent Beck's public commentary on AI agents deleting tests, illustrating eval-driven development tension. |
| b21 | AI Governance in Regulated Industries - Horizon Scan 001 | Horizon Search Institute | 2026-05 | Independent research publication documenting the institutional gap between AI governance frameworks and operational capacity in banking and healthcare, citing SR 11-7 and prior-authorization appeal data. |
| b22 | Your Weekly AI Pulse: The Institutional Age of Artificial Intelligence (July 6, 2026 Edition) | Substack | 2026-07 | Independent weekly synthesis arguing financial services governance patterns are becoming the template for other regulated sectors' agentic AI oversight. |
| b23 | The Week That Was (4–10 July 2026): AI Governance Becomes a Financial Stability Issue | Substack | 2026-07 | Tracks UK FCA and European Central Bank regulatory moves treating AI governance as a financial-stability and cybersecurity supervisory priority. |
| b24 | Artificial Intelligence Regulatory Roundup | Substack | 2026-05 | Documents Michigan's 2026 insurance-regulator bulletin referencing NIST AI RMF as an operative but non-safe-harbor compliance standard. |
| b25 | You Don't Need to be Technical to Work in AI: Why AI Governance Is Exploding in 2026 | Substack | 2026-02 | Independent commentary on AI governance becoming a distinct career function and procurement requirement as regulatory and audit demands scale. |
Frontier Lab & Model News
| ID | Title | Outlet | Date | Significance |
|---|---|---|---|---|
| t1 | Frontier Risk Report (February to March 2026) | METR | 2026-05 | Independent evaluation of internal AI agent autonomy and rogue-deployment risk at frontier labs, directly relevant to enterprise risk assessment of agentic AI oversight requirements. |
| t2 | Measuring AI Ability to Complete Long Tasks | METR | 2025-03 | Foundational paper introducing the task-completion time horizon metric, the industry's key measure of agentic capability growth underpinning agentic transformation claims. |
| t3 | Task-Completion Time Horizons of Frontier AI Models | METR | 2026-05 | Continuously updated tracker of time-horizon measurements across named frontier model releases, used to benchmark agentic reliability for production deployment decisions. |
| t4 | How Does Time Horizon Vary Across Domains? | METR | 2025-07 | Extends time-horizon methodology across coding, GUI, and other task domains, showing uneven growth rates relevant to sector-specific agentic readiness. |
| t5 | OpenAI GPT-5 System Card | arXiv / OpenAI | 2026 | Contains METR's external evaluation of gpt-5-thinking for autonomy risks, illustrating the independent evaluator model tied to specific named releases. |
| t6 | Anthropic Economic Index report: Learning curves | Anthropic | 2026-03 | Documents how experienced enterprise users extract more value from Claude over time, key evidence on adoption maturity curves. |
| t7 | Anthropic Economic Index report: Cadences | Anthropic | 2026-06 | Introduces survey-linked usage data on workplace AI expectations and self-reported job-loss forecasts, relevant to adoption-strategy and workforce impact debates. |
| t8 | Anthropic Economic Index report: Economic primitives | Anthropic | 2026-01 | Establishes primitives for measuring enterprise task concentration and success rates by task complexity, central data for the adoption-maturity research question. |
| t9 | Anthropic Economic Index report: Uneven geographic and enterprise AI adoption | Anthropic | 2025-09 | First report introducing first-party API enterprise data, documenting token-cost elasticity (0.38) directly relevant to FinOps-for-AI budgeting questions. |
| t10 | [The state of enterprise AI | OpenAI](https://openai.com/business/guides-and-resources/the-state-of-enterprise-ai-2025-report/) | OpenAI | 2025 |
| t11 | OpenAI revenue chief Dresser says enterprise AI adoption is 'at a tipping point' | CNBC | 2026-05 | Reports OpenAI's launch of a forward-deployed-engineer consulting arm (Tomoro acquisition) to accelerate enterprise onboarding, evidence of labs moving into delivery-consulting territory. |
| t12 | MIT report: 95% of generative AI pilots at companies are failing | Fortune | 2025-08 | Primary coverage of the MIT NANDA 'GenAI Divide' report, the most widely cited and contested success/failure statistic in the enterprise AI transformation debate. |
| t13 | The MIT '95% of GenAI Pilots Fail' Report: What It Gets Wrong and What Leaders Should Do Instead | Medium | 2025-08 | Methodological critique of the MIT NANDA report's sample size and framing, representing the skeptical counter-narrative to the widely propagated 95% failure figure. |
| t14 | MIT Report Finds Most AI Business Investments Fail, Reveals 'GenAI Divide' | Virtualization Review | 2025-08 | Detailed breakdown of the MIT report's build-versus-buy findings (33% internal build success vs 67% vendor success) relevant to delivery framework choices. |
| t15 | Introducing Gemini Enterprise Agent Platform | Google Cloud Blog | 2026-04 | Official announcement of Google's consolidated agent governance platform (Agent Identity, Registry, Gateway), directly evidencing vendor productisation of AI-Ops governance. |
| t16 | Google launches Gemini Enterprise Agent Platform for governance | Let's Data Science | 2026-04 | Independent editorial analysis noting vendor governance features are necessary but not sufficient for safe production agent deployment in regulated environments. |
| t17 | The new Gemini Enterprise: one platform for agent development | Google Cloud Blog | 2026-04 | Details Model Armor and Agent Gateway security controls addressing prompt injection and data leakage risks in enterprise agent deployment. |
| t18 | Claude Enterprise Compliance: BAA, SOC 2, GDPR and Data Policy (2026) | Tygart Media | 2026 | Documents Anthropic's ISO/IEC 42001:2023 certification and HIPAA-ready configuration, showing how sector regulatory requirements shape model-provider compliance posture. |
| t19 | When AI Builds Itself: The Enterprise Compliance Gap | Cloud Security Alliance | 2026-06 | Analyses how NIST AI RMF, ISO 42001 and EU AI Act update cycles lag the pace of frontier capability growth, and reports the Pentagon's supply-chain-risk designation of Anthropic. |
| t20 | Anthropic beats OpenAI on business adoption | Ramp Economics Lab / Substack | 2026-05 | Independent business-spend data (Ramp AI Index) showing Anthropic overtaking OpenAI in business adoption share and documenting token-cost incentive misalignment risk. |
| t21 | OpenAI's first state of enterprise AI | Enterprise AI Executive | 2025-12 | Summarises Menlo Ventures' enterprise generative AI market survey findings including the shift from internally-built to purchased AI solutions. |
| t22 | AI Adoption Trends in the Enterprise 2026 | TechRepublic | 2026-03 | Cites Recon Analytics survey data on 'pilot purgatory' persistence, with only 8.6% of companies reporting agents deployed in production against 63.7% with no formalised initiative. |
| t23 | A new Moore's Law for AI agents | AI Digest | 2026 | Independent explainer synthesising METR's time-horizon findings and their implications for extrapolating agentic transformation timelines. |
| t24 | Is there a half-life for the success rates of AI agents? | arXiv | 2025 | Academic analysis building on METR's time-horizon dataset to model agent reliability decay, relevant to evaluating agentic deployment risk in production. |
| t25 | How frontier AI companies could implement an internal audit function | arXiv | 2025-12 | Proposes internal audit frameworks for frontier AI companies drawing on METR evaluation practice, relevant to AI-Ops governance council design. |