The CIO's AI Library
The research behind the CIO's AI Playbook. Each entry gives a plain-English summary, the publication date, why it is worth a CIO's time, who should read it, who wrote it, the ways it is most often misquoted, and which playbook steps it shapes. Every number was checked against the original source.
Curated by Pavel Veselov · 11 essential reads · Last reviewed
The big surveys and the counter-data. Read these to know the numbers your board will quote, and how each one is usually misquoted.
McKinsey & Company (QuantumBlack) · Survey ·
The annual benchmark: AI use keeps spreading, but enterprise-level profit impact has barely moved in two years.
For: CIOs and AI leads preparing a board update · CFOs building or challenging an AI business case · CEOs who want a second look at their AI plans
Summary, why read it, about the authorBoston Consulting Group · Survey ·
Almost half of companies now get some value from AI. The new bottleneck is governing agents, not proving ROI.
For: CIOs and AI leads designing an agent governance model · CIOs who suspect they do not see the full AI budget · Risk, security and internal audit leaders
Summary, why read it, about the authorMIT Media Lab, Project NANDA · Report ·
The source of the famous “95% of GenAI pilots fail” headline, and worth reading for what it actually says.
For: CIOs fielding board questions about AI failure rates · Procurement leaders deciding between buying and building · Anyone who has quoted the 95% figure
Summary, why read it, about the authorAndreessen Horowitz (a16z) · Article ·
Revenue data on where enterprise AI is real: coding first, then legal, customer support and search.
For: CIOs choosing a first portfolio of AI use cases · Customer-support and BPO leaders · Strategy teams sizing enterprise AI markets
Summary, why read it, about the authorStrategy pieces on the gap between individual productivity and enterprise results, and between a great AI output and an organization that can use it.
Harvard Business Review · Article ·
Names the “micro-productivity trap”: everyone gets faster at tasks, and the P&L doesn't move.
For: CIOs with many AI pilots and no P&L story · COOs and business-unit leaders told to “use more AI” · Transformation and strategy teams
Summary, why read it, about the authorHarvard Business Review (January–February 2026 issue) · Article ·
GM's AI-designed seat bracket was 40% lighter and 20% stronger, and it never made it into production.
For: CIOs building a multi-year AI roadmap · Strategy and innovation leaders · Manufacturing and supply-chain executives evaluating AI-designed products
Summary, why read it, about the authorHow to govern, staff and architect AI agents: who owns them, how they learn your judgment, and when not to build one at all.
MIT Sloan Management Review and BCG · Report ·
Agentic AI is “a tool that learns and a worker that is owned”, and that breaks how companies manage both.
For: CIOs writing an agent strategy · CHROs and COOs who will share ownership of agents · Executives deciding whether to retrofit or redesign processes
Summary, why read it, about the authorHarvard Business Review · Article ·
Everyone has access to the same models. The companies scaling agents are the ones that wrote down how their best people decide.
For: CIOs and AI leads moving agents from pilot to production · Operations and HR leaders who will co-own digital labor · Domain experts asked to “document their process” for an AI team
Summary, why read it, about the authorAnthropic (Engineering blog) · Guide ·
The engineering baseline everyone borrows: start simple, prefer workflows, and invest in tools more than prompts.
For: CIOs and engineering leaders setting agent architecture standards · Product managers scoping “agent” features · Non-technical executives who want to know what an agent actually is
Summary, why read it, about the authorTwo practitioner columns that turn the ideas above into measurements: who quietly stopped using your AI tools, and what one completed transaction really costs.
CIO.com · Column ·
Adoption dashboards show who showed up. Cohort retention shows which AI tools people quietly stopped using.
For: CIOs with enterprise AI licenses up for renewal · IT procurement and vendor managers · Owners of internal AI platforms
Summary, why read it, about the authorCIO.com · Column ·
An AI workflow cut turnaround by 90% and cost about fifteen times more per document than the clerk it replaced.
For: CIOs presenting AI business cases · CFOs and FP&A teams reviewing AI spend · Innovation teams running pilots on central budgets
Summary, why read it, about the authorStart with the two large surveys, McKinsey's State of AI 2026 and BCG's Applied AI Index 2026, to understand where value is and isn't showing up. Then read HBR's piece on the micro-productivity trap for why pilots don't reach the P&L, and Anthropic's Building Effective AI Agents for the engineering baseline. Finish with the two CIO.com columns on abandonment and cloud cost, which turn the ideas into metrics you can run next week.
It comes from MIT Project NANDA's July 2025 report and is usually misquoted. The 95% refers to all organizations in the sample, most of which never ran a custom pilot, failing to get measurable P&L impact from task-specific generative AI. Among organizations that did pilot, roughly one in four succeeded. The report is preliminary, self-published and based on 52 interviews and 153 survey responses, so treat it as a signal, not a census.
BCG's Applied AI Index 2026 lists six controls: governed data and memory access, human-in-the-loop for high-risk decisions, testing with shut-off and rollback, clear hand-off interfaces, a named owner for every agent, and ongoing security. Only 5% of companies have all six, and those report three times the agentic value of companies with one. MIT SMR and BCG add an “HR for agents” lifecycle covering onboarding, evaluation and retirement.
Track two things adoption dashboards miss. First, cohort retention: of the people who activated a tool in a given month, how many are still active at weeks 1, 4 and 12. Second, fully loaded cost per completed transaction against the manual baseline, including model fees, compute and the human time spent on exceptions. Speed and seat counts alone can hide a tool that is abandoned or more expensive than the work it replaced.