A board that hears “AI is paying off” has learned very little it can act on. Kyndryl, a technology services provider, released its third annual Readiness Report on November 10, 2026, and it says 64% of leaders report positive returns on AI investments, up 10 points from 2025. A return claimed for a whole program can’t tell anyone which initiative to stop, and that gap is where money goes unexamined.
Basil C. Puglisi reads the finding as a governance question that finance teams are well placed to answer. The stop decision is part of the return, and it needs a named owner and per-initiative numbers. Every figure below comes from a vendor survey, and the limits section says how far each one can carry the argument.
Kyndryl’s 64% and its 21% belong in one sentence
The same release carries a second set of numbers. Kyndryl’s survey of 3,700 senior business leaders across 21 countries found that about a quarter (24%) saw an AI investment fail and 21% canceled a major AI initiative. Kyndryl also reports that just 27% have formally updated roles and responsibilities around AI, and 29% have fully implemented change management.
Put the 64% and the 21% side by side and the story changes. Most leaders say AI returns money, and about one in five canceled a major initiative. Puglisi’s position is that a canceled initiative can be governance working, provided a named person decided it on evidence and wrote down why. Without that record, later reviewers can’t judge the cancellation or apply its lessons elsewhere.
Kyndryl’s release also reports that average readiness across external risks fell seven points to 24%, and it frames the report as a widening divide between AI returns and readiness to manage the risks that come with them. That figure measures risk readiness and offers no direct evidence on stopping.
A program-level ROI figure can’t show what to stop
CloudZero, a vendor of AI and cloud cost tools, published a November 6, 2026 analysis of its survey of 252 senior finance leaders. CloudZero reports that 23% can both allocate AI spend systematically and measure its return rigorously. The company’s page also says only 3% of that group failed to secure an AI investment, against 19% of everyone else.
The finding most relevant to a stop decision sits in the middle of the page. Across the full sample, CloudZero says 47% report that their board asks for ROI proof they cannot provide. Among companies claiming board-defensible ROI who see only a total or a rough category split, the figure is 56%. CloudZero adds that those companies killed or paused fewer initiatives than weak-allocation peers claiming no return, 9% against 21%.
CloudZero summarizes the distinction this way: “ROI measurement tells a board the program returned something. AI cost allocation tells them which part of it did.” CloudZero presents the lower kill rate as a modest benefit of claiming ROI. Puglisi reads it differently: without per-initiative allocation, a board can’t tell whether fewer stops reflect strong performance or missing evidence.
Dashboards and cost alerts are machine checking machine
Spend dashboards and anomaly alerts are one common tool response to the allocation gap. Puglisi treats these as useful and incomplete. In his Enterprise AI ROI paper, published April 2, 2026, he separates two modes of working with AI. Factory Quality means the machine checks the machine, which that paper ties to Responsible AI. Handmade Quality means a named human holds authority at defined checkpoints, which the paper ties to AI Governance.
An alert that fires when spend jumps is Factory Quality. It tells someone to look, and it can’t decide. In Puglisi’s definition, Checkpoint-Based Governance (CBG) puts a named human at the decision point with authority to accept, modify, or reject, and with a record of who decided, on what evidence, why, and what dissent ran against it. For a running AI initiative, Puglisi translates those three outcomes into continue, modify, or stop.
The same paper applies the Factics method, which moves from Fact to Tactic to KPI, and it argues that AI deserves capital governance. Its section 3 is titled “Your Pilots Are Not Failing. They Were Never Built to Succeed.” Its first section asks that every AI initiative above a defined materiality threshold carry a named capital owner, with authority during deployment to redirect or terminate.
Five steps for a leadership and finance team
Puglisi would advise a client to take these steps together and assign each one an owner before the next budget review.
- Allocate AI spend to each initiative. A total, or a rough category split, is the condition CloudZero’s weak-allocation group describes. Finance and the business sponsor agree on the allocation method before the next budget cycle.
- Write stop and change thresholds at approval. Each initiative carries a written level of cost or outcome at which it changes or ends. The threshold exists before anyone has a reason to defend the project.
- Name one capital owner per initiative. That person signs continue, modify, or stop at each quarterly review. A committee can advise, but one name signs.
- Record canceled initiatives and where the money went. Report them to the board beside the returns, so a stop reads as a decision with an outcome.
- Tie the decision record to a roles update. Kyndryl says only 27% have formally updated roles and responsibilities around AI, and that those who also fully implemented change management are 8 points more likely to report positive ROI. A decision record means little if no job description gives the signer the authority.
Four KPIs a board can read
Each step produces a number an executive can put on a dashboard. These are proposed measures, offered as a starting point for testing, and none is an established benchmark.
- Share of AI spend allocated to a named initiative.
- Share of initiatives with written stop thresholds and a named owner.
- Time from a missed threshold to a signed decision.
- Share of stopped-initiative budget redeployed within a quarter.
The third measure matters most. A long gap between a missed threshold and a signature suggests the checkpoint exists on paper only.
What the evidence can and can’t carry
Both surveys come from vendors with a commercial interest in the subject. Kyndryl sells technology services, and CloudZero sells cost tools. The returns are self-reported, and the Kyndryl release doesn’t define “positive returns.” CloudZero’s sample of 252 is small, and its subgroups are smaller. Both surveys show association, not cause, including Kyndryl’s 8-point comparison. Companies that allocate well may simply be better managed.
Puglisi’s own framework is also a proposal. It’s one contribution to a live discussion, offered for others to test or replace. Finance teams with a different approach to stop decisions should compare their results against these four KPIs.
Once returns are claimed, the next argument is over what gets stopped, and that argument needs a person’s name on it. Watch whether companies that report rising returns also begin reporting what they canceled and who signed the cancellation. If that disclosure stays absent, the 64% will stay hard to interpret.
Sources
- CloudZero. (2026, November 6). AI ROI measurement: What the 23% do differently. https://www.cloudzero.com/blog/ai-roi-measurement/
- Kyndryl. (2026, November 10). Kyndryl report: AI returns rise as enterprise readiness falls [Press release]. https://www.kyndryl.com/us/en/about-us/news/2026/11/global-enterprise-ai-readiness-gap
- Puglisi, B. C. (n.d.). Checkpoint-Based Governance: Presence is not authority. basilpuglisi.com. https://basilpuglisi.com/cbg/
- Puglisi, B. C. (2026, April 2). Enterprise AI ROI: What seven landmark reports found, what they missed, and five decisions worth making now. basilpuglisi.com. https://basilpuglisi.com/enterprise-ai-roi-seven-landmark-reports-five-decisions/
Questions readers ask
What does it mean to say the stop decision is part of the AI return?
A return figure only guides a board if someone can act on it. Ending an initiative that misses its thresholds can free budget for better uses, after exit costs, so the decision to stop changes the program’s net return. Basil C. Puglisi treats that decision as part of ROI and asks for a named owner.
Why can’t a program-level ROI claim tell a board what to stop?
A program-wide return combines results across initiatives and can hide the differences between them. CloudZero reports that 56% of companies claiming board-defensible ROI, while seeing only a total or rough category split, say their board asks for proof they cannot provide. Without per-initiative allocation, nobody can see which piece is not returning.
What is Checkpoint-Based Governance for an AI initiative?
Checkpoint-Based Governance, or CBG, places a named human at defined decision points with authority to accept, modify, or reject, plus a record of who decided, on what evidence, and why. For a running initiative, Puglisi translates those outcomes into continue, modify, or stop, signed by one capital owner each quarter.
Are automated dashboards and cost alerts enough?
They are useful and insufficient. Alerts and dashboards are machine checking machine, which Basil C. Puglisi’s framework calls Responsible AI. They flag a problem and cannot decide. A named capital owner who signs continue, modify, or stop at a checkpoint supplies the decision and the accountability.
How reliable are the Kyndryl and CloudZero figures?
Treat them as directional. Both are vendor surveys with self-reported answers, and CloudZero’s sample is 252 senior finance leaders. Kyndryl’s release does not define positive returns. The results show associations, not cause, so test the proposed KPIs against your own initiative data.
#AIg
Leave a Reply