Software Project Outcomes Over Time, and Is AI Improving Them

Narrative

Financial-press coverage of software delivery outcomes clusters around three storylines: the durability of grim baseline statistics, high-profile public-sector cancellations, and a sharpening Wall Street debate over whether AI coding tools are actually paying off. The foundational quantitative source remains McKinsey's Oxford collaboration on 5,400 large IT projects, which found that on average large IT projects run 45 percent over budget and 7 percent over time, while delivering 56 percent less value than predicted, with software projects carrying the highest risk of cost and schedule overruns. That single 2012 dataset still anchors nearly every subsequent citation of "large IT project" failure rates across consultancies and trade press, which is itself a methodological concern: independent academic scrutiny of the Standish CHAOS figures (Eveleens and Verhoef, cited in the peer-reviewed literature) found that Standish's successful and challenged project results are indeed meaningless for benchmarking, even though the directional findings, that large projects fail far more than small ones, have been corroborated by separate McKinsey, BCG and PMI surveys.

Reuters' August 2025 investigation into the Pentagon's cancellation of two nearly finished Navy and Air Force HR software systems, built by Accenture and Oracle over 12 years for more than 800 million dollars, is the clearest financial-press case study of large-scale public-sector software failure in the current period. Reuters reported that both systems had cleared multiple independent technical reviews and an April status update had described the Air Force project as "on track," yet Pentagon leadership issued a "strategic pause" before scrapping the work in favour of new bids to Salesforce and Palantir, prompting Democratic lawmakers to call the reversal a "costly do-over." This episode illustrates a recurring theme in the public-sector literature: cancellations driven by governance and procurement politics rather than technical failure, with sunk costs approaching completion.

On the AI-assisted development question, Bloomberg Businessweek's February 2026 feature "AI Coding Agents Like Claude Code Are Fueling a Productivity Panic in Tech" captures the shift in mood a year after "vibe coding" entered the vocabulary: the promise of easier development instead touched off a high-pressure race to build software at any cost. This sits alongside Bloomberg's earnings coverage showing Wall Street turning skeptical of the hundreds of billions of dollars Big Tech is spending on AI, with investors demanding evidence of monetisation from Microsoft, Meta, Alphabet and Amazon before rewarding further capital expenditure. The rigorous instrumented evidence remains thin and contested: METR's randomized controlled trial of 16 experienced open-source developers found that allowing AI actually increased completion time by 19 percent even though developers had forecast a 24 percent speed-up beforehand, a result METR itself later cautioned may be an unreliable signal given participant self-selection bias in follow-up experiments. Google's 2025 DORA report, surveying nearly 5,000 technology professionals, found AI adoption surged to 90 percent but concluded AI does not fix a team, it amplifies what is already there, meaning strong teams improve further while struggling teams see existing problems intensified.

Enterprise commissioning coverage shows the low-code and AI-coding markets both scaling fast but without independent evidence resolving the build-versus-buy debate. Gartner's Magic Quadrant methodology and market-sizing estimates (cited across trade press rather than primary financial outlets) put the AI code-assistant market at roughly 3 to 3.5 billion dollars in 2025, while low-code platform forecasts from Gartner and IDC project the category will account for the majority of new enterprise applications by 2026. Reporting on enterprise AI cost governance, including CIO.com's coverage of AI cost overruns and Indeed Hiring Lab data (reported via Yahoo Finance) showing software developer job postings ticking up since Claude Code's 2025 release, point to a market still in the price-discovery phase rather than one with settled, audited productivity outcomes.