Enterprise AI Transformation Programmes (2025–2026)

Narrative

The MIT NANDA "GenAI Divide" report's headline that 95% of generative AI pilots show no measurable P&L impact drew immediate and sustained scrutiny from independent AI writers. Zvi Mowshowitz, writing on his widely-read Substack, dissected the report's internal inconsistencies directly, noting the study's own data shows purchased tools succeed "twice the success rate of" internally built ones and questioning the "zero return" framing given that the report calls the 95% of projects without "measurable P&L impact" failures, even though back-office ROI is hard to prove within a P&L window, and doubling coder productivity has real value even without a directly attributable dollar figure. Marketing AI Institute's Paul Roetzer made a parallel critique, arguing the study's methodology used a narrow six-month, P&L-only success definition and rested on just 52 interviews that the report itself admits are only "directionally accurate based on individual interviews rather than official company reporting". Even sources sympathetic to the report's broader thesis, such as a 2026 industry landscape blog, concede that "the figure has been debated, critics note its six-month P&L definition ignores efficiency and CX gains, but even sceptics accept the underlying signal: High Adoption, Low Transformation".

Independent commentary through 2026 has shifted attention from pilot failure rates toward a more granular story about token economics and cost governance. Simon Willison's Newsletter and TILs blog documented the collision between falling per-token prices and exploding per-task consumption in near real time, noting cases like Uber, where he observed that the company's coding-tool budget problem stemmed from a policy set "before anyone could have predicted how popular token-burning coding agents were about to become". Other Substack writers quantified this pattern more starkly: one investor-newsletter piece reported that "Amazon, supposedly, blew $500m in one month on Claude without seeing much ROI and had to stop token leaderboards" while "Uber blew its 2026 AI budget in 4 months and blamed Anthropic for it, without moving any of their business KPIs", while a separate FinOps-focused Substack cited survey data that "Benchmarkit's 2025 survey of 372 enterprise organizations found that only 15 percent of companies could forecast AI costs within 10 percent of actual spend," with "nearly one in four" missing "by more than 50 percent". A widely-discussed newsletter piece coined "tokenmaxxing" for the practice of gamifying token consumption as an adoption proxy, warning that "organizations are measuring AI adoption by how much the users are using... often called token maxing... it's a terrible measure. Because if you think about it, if you are holding people accountable to how many tokens they use, they have a natural incentive to use as much as possible". Confirming the trend at leadership level, an enterprise-CTO-briefing Substack reported "Jensen Huang has publicly proposed a token budget of up to 50% of an engineer's annual salary" and that "the CTO is becoming the CFO of tokens".

On delivery frameworks, independent bloggers treat the DORA and ThoughtWorks lineages as the credible bridge between classic DevOps/agile practice and agentic AI, while pushing back on report inflation. One critical Substack post accused Google Cloud's DORA team of recycling its own prior findings, observing that the 2026 report's core message closely mirrors the 2025 report's: "Artificial Intelligence (AI) serves as a powerful amplifier in software development. It magnifies the strengths of high-performing organizations and the dysfunctions of struggling ones," compared with the 2025 report's near-identical framing that "AI's primary role in software development is that of an amplifier". Thoughtworks' own Technology Radar volume 34, discussed on Martin Fowler's blog, argued the industry response should be a return to fundamentals rather than a new methodology, noting the radar's finding that teams are "returning to many established techniques, from pair programming to zero trust architecture, and from mutation testing to DORA metrics" as "not nostalgia, but a necessary counterweight to the speed at which AI tools can generate complexity". Independent consultants writing on "pilot purgatory" converge on organisational rather than technical explanations: one consultant's Substack cited a randomised trial finding that a survey of 5,000 white-collar workers found over 40% of executives claimed AI saved them eight-plus hours weekly while two-thirds of non-management staff reported saving under two hours, and a METR randomised controlled trial "found that AI tools made them 19% slower - while the developers themselves believed they were 20% faster". O'Reilly Radar's independent analysis reached a similar structural diagnosis, arguing that centre-of-excellence models "invariably" become bottlenecked ivory towers and that the fix companies like JPMorganChase, Walmart and Uber found was a "third way," an "outcome-oriented hybrid architecture" combining "centralized enablement with distributed execution, aggressive governance with operational autonomy, and technical excellence with a relentless focus on business value".

On governance and sector risk, independent research blogs increasingly note that existing regulatory scaffolding predates agentic AI and is being stretched to fit it. A May 2026 independent research publication observed that in financial services, "Goldman Sachs has embedded Anthropic engineers to co-develop autonomous compliance agents while its model validation function still operates on quarterly cycles," and that "three in four health plans now use AI in prior authorization, with appeal overturn rates above 80% in Medicare Advantage and patient appeal rates below 1%", framing this as an accountability gap rather than a framework gap. The same source noted that decade-old US banking model-risk rules remain the reference point even for agentic systems, since "SR 11-7, issued jointly by the Office of the Comptroller of the Currency and the Federal Reserve Board in 2011... established the supervisory expectation that banks maintain strong model governance," with its "foundational principles" still described by risk professionals in February 2026 as "conceptually robust" even as implementation tools strain under agentic complexity. A regulatory-tracking Substack flagged early state-level enforcement signals, noting that Michigan's insurance regulator issued a 2026 bulletin explicitly referencing NIST's framework, requiring firms to maintain a written AI Systems Program while making clear that Michigan's bulletin "explicitly references the NIST AI RMF as an appropriate framework" but "does not provide a formal safe harbor," meaning "AI made the decision" is not a defense for adverse consumer outcomes.