Jon W. Hansen, FCIPS · Procurement Insights | Hansen Models™ · October 2026
At its IT Symposium/Xpo in Orlando this week, Gartner opened with a sobering set of findings. According to its 2026 C-Suite AI Survey, only one in five AI use cases delivers positive returns. The average organization increased its number of AI use cases by more than 400 percent in a single year, yet most of those projects still fail to pay back. And 86 percent of CIOs report that their AI risk is growing faster than their AI value, by four times on average.
Gartner Distinguished VP Analysts Daryl Plummer and Kristin Moyer responded with a new framework, the Value-Risk Flywheel, and a new concept, the AI Central Bank. Both are worth understanding. But the numbers deserve attention first, because they describe a pattern this archive has been recording for nearly two decades, and it is good to see it gaining mainstream attention.
The failure rate did not change
One in five use cases delivering returns means roughly four in five do not. That is the failure rate this archive has tracked across every technology era, from ERP to e-procurement to cloud and now AI. In March, I wrote that the procurement technology failure rate had survived every era, because each new technology was deployed into organizations whose real operations had never been examined. Gartner has now measured the same pattern for AI.
Its own phrase captures the reason: activity without value is simply bad momentum. A 400 percent increase in use cases is activity. It is what was counted. The outcome is what the one-in-five figure measures.
The risk is compounding, as predicted
The keynote names the risks it says continue to compound across the enterprise: shadow AI, dark AI, technical debt, process debt and cybersecurity exposure. That is close to the Compounding Technology Shadow Wave™ this archive described in May: shadow tools layering on top of each other with every technology era, each layer adding to the gap between the operation as documented and the operation as it actually runs.
It is one thing to say risk is compounding. It is another to say where it comes from. It comes from the parts of the operation nobody has mapped.
Believing versus knowing
The most telling figures in the keynote sit side by side without being connected. Gartner reports that 58 percent of functional leaders say their AI initiatives deliver positive value. It also reports that only one in five AI use cases actually delivers positive returns.
That gap is the difference between believing and knowing. More than half of leaders believe AI is paying off. The returns confirm it for one use case in five. The distance between those two numbers is exactly where organizations make their next investment decisions.
One caution applies to all of these figures: they are survey results. “Risk is growing four times faster than value” is what CIOs report, not a measurement of risk itself. That makes the perception gap more important, not less.
Where the remedy can go further
The Value-Risk Flywheel asks leaders to stop treating AI as a collection of projects and to manage it as a compounding system, building capabilities that make each success easier to repeat. That is sound advice. The AI Central Bank proposes a common control plane for accountability, technology integration, cost and shared intelligence, which addresses a real problem: accountability becomes harder to see as AI spreads across people, agents and platforms.
But both start from the AI portfolio: the use cases, the platforms, the agents. Neither starts from where the operation actually fails and traces back from there. That is the difference between governing the agent and governing the operation the agent acts in.
A central bank can only govern what is on its ledger. Shadow AI and process debt are, by definition, the things that are not. They are the workarounds, the unsanctioned tools and the unrecorded handoffs that never appear in the official picture. And the shadow often starts at the top: in a March 2026 survey of 2,001 employees in the United Kingdom and the United States, 65 percent of senior decision-makers use unapproved AI tools, versus 31 percent unauthorized use by the people they manage. A control plane built on the documented operation will see the documented operation clearly, and miss the compounding risk Gartner is warning about.
The keynote also frames governance as drag: good drag protects value, bad drag creates bureaucracy. The distinction is useful. But governance does not have to be friction. A firm fence, built on outcome evidence rather than rules alone, is what allows AI to keep operating safely as the operation changes. It is firm about the purpose, the outcome threshold, the authority boundary and the evidence, and flexible in its explanation of how the operation works.
People at the center
The keynote’s closing message is the one I agree with most. Gartner calls people the first responders to AI failures, the last line of defense against dark AI, and the source of new ways to create value, and asks CIOs to make “people to the rescue” their mantra.
That lands in the same week Harvard Business Review argued for moving decision authority away from people and toward AI systems, on the grounds that human overrides are mostly self-protection. Both cannot be right in every case. When people override an AI system, some are protecting themselves. Others are seeing something the system cannot. Telling the two apart is the work that comes before deciding who holds authority. I explored that distinction in my response to the HBR article.
I would go one step further. People are most valuable not after the system fails, but alongside it from the start, seeing what it cannot.
The takeaway
Gartner has measured the pattern: four in five AI use cases fail to pay back, and risk is compounding faster than value. That convergence is welcome. When two independent records, built in different ways, arrive at the same pattern, each makes the other harder to dismiss. The archive adds the why: technology deployed into operations nobody has traced, with the real risk sitting in the parts the official picture never shows.
The remedy is not more AI activity, and it is not only better governance of the AI portfolio. It is starting from the operation, tracing back from where it fails, and building firm fences that learn from outcomes through loopback learning mechanisms. The role people play in that case is that of a synthesized critical-thinking partner who sees what the system cannot.
Sources and related reading:
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Related
Gartner’s New AI Numbers Match What the Procurement Insights Archive Has Tracked for Years
Posted on October 22, 2026
0
Jon W. Hansen, FCIPS · Procurement Insights | Hansen Models™ · October 2026
At its IT Symposium/Xpo in Orlando this week, Gartner opened with a sobering set of findings. According to its 2026 C-Suite AI Survey, only one in five AI use cases delivers positive returns. The average organization increased its number of AI use cases by more than 400 percent in a single year, yet most of those projects still fail to pay back. And 86 percent of CIOs report that their AI risk is growing faster than their AI value, by four times on average.
Gartner Distinguished VP Analysts Daryl Plummer and Kristin Moyer responded with a new framework, the Value-Risk Flywheel, and a new concept, the AI Central Bank. Both are worth understanding. But the numbers deserve attention first, because they describe a pattern this archive has been recording for nearly two decades, and it is good to see it gaining mainstream attention.
The failure rate did not change
One in five use cases delivering returns means roughly four in five do not. That is the failure rate this archive has tracked across every technology era, from ERP to e-procurement to cloud and now AI. In March, I wrote that the procurement technology failure rate had survived every era, because each new technology was deployed into organizations whose real operations had never been examined. Gartner has now measured the same pattern for AI.
Its own phrase captures the reason: activity without value is simply bad momentum. A 400 percent increase in use cases is activity. It is what was counted. The outcome is what the one-in-five figure measures.
The risk is compounding, as predicted
The keynote names the risks it says continue to compound across the enterprise: shadow AI, dark AI, technical debt, process debt and cybersecurity exposure. That is close to the Compounding Technology Shadow Wave™ this archive described in May: shadow tools layering on top of each other with every technology era, each layer adding to the gap between the operation as documented and the operation as it actually runs.
It is one thing to say risk is compounding. It is another to say where it comes from. It comes from the parts of the operation nobody has mapped.
Believing versus knowing
The most telling figures in the keynote sit side by side without being connected. Gartner reports that 58 percent of functional leaders say their AI initiatives deliver positive value. It also reports that only one in five AI use cases actually delivers positive returns.
That gap is the difference between believing and knowing. More than half of leaders believe AI is paying off. The returns confirm it for one use case in five. The distance between those two numbers is exactly where organizations make their next investment decisions.
One caution applies to all of these figures: they are survey results. “Risk is growing four times faster than value” is what CIOs report, not a measurement of risk itself. That makes the perception gap more important, not less.
Where the remedy can go further
The Value-Risk Flywheel asks leaders to stop treating AI as a collection of projects and to manage it as a compounding system, building capabilities that make each success easier to repeat. That is sound advice. The AI Central Bank proposes a common control plane for accountability, technology integration, cost and shared intelligence, which addresses a real problem: accountability becomes harder to see as AI spreads across people, agents and platforms.
But both start from the AI portfolio: the use cases, the platforms, the agents. Neither starts from where the operation actually fails and traces back from there. That is the difference between governing the agent and governing the operation the agent acts in.
A central bank can only govern what is on its ledger. Shadow AI and process debt are, by definition, the things that are not. They are the workarounds, the unsanctioned tools and the unrecorded handoffs that never appear in the official picture. And the shadow often starts at the top: in a March 2026 survey of 2,001 employees in the United Kingdom and the United States, 65 percent of senior decision-makers use unapproved AI tools, versus 31 percent unauthorized use by the people they manage. A control plane built on the documented operation will see the documented operation clearly, and miss the compounding risk Gartner is warning about.
The keynote also frames governance as drag: good drag protects value, bad drag creates bureaucracy. The distinction is useful. But governance does not have to be friction. A firm fence, built on outcome evidence rather than rules alone, is what allows AI to keep operating safely as the operation changes. It is firm about the purpose, the outcome threshold, the authority boundary and the evidence, and flexible in its explanation of how the operation works.
People at the center
The keynote’s closing message is the one I agree with most. Gartner calls people the first responders to AI failures, the last line of defense against dark AI, and the source of new ways to create value, and asks CIOs to make “people to the rescue” their mantra.
That lands in the same week Harvard Business Review argued for moving decision authority away from people and toward AI systems, on the grounds that human overrides are mostly self-protection. Both cannot be right in every case. When people override an AI system, some are protecting themselves. Others are seeing something the system cannot. Telling the two apart is the work that comes before deciding who holds authority. I explored that distinction in my response to the HBR article.
I would go one step further. People are most valuable not after the system fails, but alongside it from the start, seeing what it cannot.
The takeaway
Gartner has measured the pattern: four in five AI use cases fail to pay back, and risk is compounding faster than value. That convergence is welcome. When two independent records, built in different ways, arrive at the same pattern, each makes the other harder to dismiss. The archive adds the why: technology deployed into operations nobody has traced, with the real risk sitting in the parts the official picture never shows.
The remedy is not more AI activity, and it is not only better governance of the AI portfolio. It is starting from the operation, tracing back from where it fails, and building firm fences that learn from outcomes through loopback learning mechanisms. The role people play in that case is that of a synthesized critical-thinking partner who sees what the system cannot.
Sources and related reading:
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Share this:
Like this:
Related