The AI Bottleneck Has Moved. Have You?
In February 2025, enterprises told the International Data Corporation (IDC) what was blocking their AI programs: financial risk and uncertain ROI, disorganized data, and missing in-house expertise. Money, data, skills. Three significant input problems.
Eleven months later, the same researchers asked again. Adoption had doubled, and the blockers had changed: responsible AI, governance, and the design of human-AI collaboration within workflows.
Nothing about the technology got harder to implement in eleven months. The constraints moved.
From Inputs to Operating Models
We can see the change from 2025 to 2026. IDC’s CIO Playbook, commissioned by Lenovo, surveyed 2,920 IT and business decision makers in February 2025 and 3,120 in January 2026. While both editions of the CIO Playbook were sponsored by Lenovo, the data sets are large enough that they represent a serious look at industry trends.
In the 2025 edition, the top three barriers among organizations not yet adopting AI clustered within a percentage point of each other: financial risk and uncertain ROI at 33.3%, insufficient or disorganized data at 33%, and lack of in-house AI expertise at 32.5%. The top success factors named that year were data sovereignty, compliance, and data quality. Every item is an input you can buy, clean, or hire.
By the 2026 edition, the ranking of critical success factors had reordered. Employee training and upskilling came first, scalable infrastructure came second, and effective human-AI collaboration for optimized workflows came third. Two of the top three are questions about how your organization works, not who it has or what it owns.
The two editions ask slightly different questions of slightly different populations, but that demonstrates a shift away from procurement problems towards operating problems.
Among the top AI trust concerns enterprises reported in 2026, poor data quality ranks fifth. Last. Behind lack of responsible AI, behind knowledge and application of responsible AI, behind poor data security, and behind shadow AI deployments. That starts to raise interesting questions about the current state of AI implementation.
The Deployment Trap Is Now Measurable
I have written about the Deployment Trap before: the assumption that AI is something you install rather than something you decide. The conventional diagnosis for a failed pilot is fairly predictable: the model wasn’t accurate enough, the data wasn’t clean enough, or the use case wasn’t narrow enough. And as a result, the prescribed treatment is always more of the same: better models, cleaner pipelines, tighter scope.
What the 2026 data shows is that the deployment trap is now visible directly in the numbers.
Adoption has moved. The share of organizations piloting or systematically deploying AI doubled year over year, reaching 60%. Investment has moved with it: 96% of organizations plan to increase AI spending in the next twelve months, at an average growth rate of 13%, and AI agent development jumped from the tenth-ranked investment priority to the second in a single year.
Readiness has not moved with either of them.
Only 21% of organizations have deployed agentic AI at scale. Fifty-five percent are exploring, piloting, or running limited deployments. And 39% say they need more than twelve months simply to be ready for scaled implementation. Not to finish. To be ready to start.
Set that beside the governance statistics and the picture becomes even clearer. Only 27% of organizations report a comprehensive AI governance framework. Fifty-six percent are still developing policies. Twenty-seven percent governed, twenty-one percent deployed at scale.
The most quietly damning figure in the study is actually neither of those. Across AI projects that actually reach production, average adoption among the target users is 48%. The pilot survived, the deployment happened, and half the people it was built for don’t use it. No AI model alone is going to improve that number.
If you lead in aerospace, defense, or automotive, the manufacturing data should shock you. Forty percent of manufacturers need more than twelve months to be ready for agentic AI. More than a quarter have no plans to adopt it at all. And while global focus on agentic AI is growing 52% year over year, in manufacturing it is growing 9%. The sector with the most complex workflows is preparing the slowest, and potentially the farthest behind.
The Framework is Catching up to the Problem
On June 23, 2026, Scaled Agile released AI-Native SAFe. Andrew Sales, the organization’s Chief Methodologist, framed the release around a single claim: “The bottleneck has moved. The challenge for organizations is no longer whether they can build something in the time available. Rather, it is keeping up with the need to validate whether what they are building is safe, secure, and valuable.”
I talked about that line in July, when I wrote about the two AI decisions a Chief AI Officer cannot make for you, and now I’m quoting it again for a different reason. In July it was a hypothesis from a framework body. Now, set against two years of survey data, it is a description of something that already happened.
Scaled Agile’s own material on the new AI Value Architect role puts it more bluntly than I would have: most AI initiatives fail not because the technology doesn’t work, but because no one understood the path from prototype to production.
That is a call for improving your operating model. We have to stop treating AI as a technology deployment problem. A deployment problem lives in the technology department. An operating model problem lives in the leadership team, and you can’t delegate your way out of it, or solve it with another vendor contract.
What AI-Native SAFe Actually Demands of Leaders
Three operating model changes that will drive the success of AI adoption.
A named role that owns AI outcomes inside a value stream.
The AI Value Architect is not a data scientist with a title change or a product owner with an AI suffix. It is a dedicated role on the Agile Release Train, accountable for coaching teams on responsible AI use, connecting business objectives to technical implementation, managing the associated risks, and measuring whether any of it worked. Scaled Agile positions it as the systems view that connects strategy to execution.
The role exists because AI agents require a different kind of human ownership than traditional software. Software does what it is told, but AI agents infer, and that difference creates risk, and risk needs to be managed by someone.
Validation becomes a first-class activity, not a phase-gate at the end.
If the constraint is confirming that what you built is safe, accurate, and organizationally sanctioned, then validation capacity is the thing you plan for, staff, and measure. Most enterprises still treat it as a review step that happens after the interesting work is finished. IDC's own read of the 2026 data is very clear: unlike generative AI, agentic AI demands deeper process redesign to realize value, and workflows must be standardized, monitored, and in some cases completely restructured.
Smaller teams, shorter cycles, and a governance cadence that matches the risk.
Traditional agile rhythms were designed for human teams producing human outputs. AI agents generate errors, bias, and downstream process damage faster than any human team can. AI-Native SAFe restructures around smaller AI-augmented teams working in shorter, more iterative cycles. The operating review cycle has to match the risk profile, not the pre-AI planning calendar.
What This Looks Like in Practice
A Fortune 500 financial services organization I worked with spent months building an AI agent for document review. The model performed, the integration worked, and at deployment, they discovered that no one held authority over exception handling.
The agent went live. The exceptions built up. And since nobody owned them, everybody said someone else should, and the risk compounded while the org chart was debated. The pilot was suspended four weeks later and the failure was recorded internally as a data quality issue.
It was not a data quality issue. It was an unassigned decision right, and it had been unassigned since the day the project was funded.
Remember where data quality ranked in IDC’s 2026 trust concerns. Fifth, behind four different governance and responsibility failures. The excuse enterprises reach for when a pilot dies is the concern practitioners worry about least. That gap is not an accident, “data quality issue” is the only root cause that doesn’t blame anyone in the room.
Now compare that to a cyberphysical manufacturing organization I supported, inside a business unit of 12,000 people. Their leaders were spread across more than a dozen functional verticals with no standardized ways of working between them, and the churn showed up as constant renegotiation of communication and priorities.
We started with the question rather than the technology: where is the current process actually breaking? The answer was that there was no single clean source of truth for enterprise standards and governance. The internal repositories and solution intent were massive, hard to navigate, and inconsistent with each other. The verticals weren’t disagreeing on principle. They were reading different documents.
Only then did we build. We deployed AI agents that validated the data, identified inconsistencies, and powered a search that surfaced every relevant source while highlighting where those sources conflicted. And we paired the agents with a single individual, what SAFe now calls an AI Value Architect, explicitly empowered to decide which standards could be kept and which had to be escalated to the senior leadership team for resolution.
That pairing is the whole engagement. The agents found the conflicts. A human with real authority resolved them.
We delivered a simplified, consistent source of truth in two months. Across the first Planning Interval after implementation, the eight to twelve week cycle SAFe teams plan and commit against, integration conflicts between verticals dropped by more than 60%.
None of that required manufacturing to stop being manufacturing. It is the sector the 2026 data says is preparing slowest, and the constraint there was never the complexity of the workflows. It was that nobody had been given the authority to simplify them.
Three Decisions Before the Next Pilot
If you are accountable for an AI transformation, there are three decisions that belong to you before the next pilot launches.
Define the workflow before you define the technology.
What is the current state, step by step? Where does it break? What does success look like in process terms, expressed as a number? This is the work most enterprises skip. The question is not "what can AI do?" It is "what should AI do, in this workflow, at this point, with these constraints?" One is a capabilities question, the other is an operating model question. Only the second produces a pilot worth running.
Assign the AI Value Architect, or its functional equivalent.
Someone has to own the intersection of outcome accountability, risk, and governance for agent work. Not the CDO setting enterprise data strategy. Not the CAIO managing vendor relationships. A person with operational authority over specific value streams, accountable for whether the agents inside those streams are safe, accurate, and governed. If you cannot name that person today, you have 27%-of-the-market governance, which is to say you have none.
Design the human-AI handoff before the pilot starts.
Where does autonomous action stop and human judgment begin? What is the threshold, who holds it, and what is the escalation path when the agent is uncertain or wrong? Define it as an operational specification, not a philosophical position: the agent handles intake up to this threshold, anything above it routes to this role within this response time. I have never seen an enterprise agent pilot that built this specification before deployment. I have seen many that tried to retrofit it after the first failure.
Are You Ready to Receive What You Build?
Four key measures from the 2026 data:
That gap is not a technology problem. It is the difference between what enterprises have bought and what they have built the discipline to operate.
In 2025, the honest answer to “what is stopping us?” was budget, data, and skills. Those were real constraints and the industry spent a year clearing them. What is left is the part that we never really want to talk about: who decides, who is accountable, what happens when the agent is wrong, and how you would know?
The question for your next steering committee is not which pilot to run next. It is whether your organization is ready to receive what you build.
Most are not. The ones that say so out loud are the ones that get to 21%.
Sources
- IDC CIO Playbook 2026: The Race for Enterprise AI (IDC survey commissioned by Lenovo, January 2026, n=3,120)
- IDC Global CIO Playbook 2025: It’s Time for AI-nomics (IDC Custom Solutions infographic sponsored by Lenovo, February 2025, n=2,920, doc #WW242508IG)
- Scaled Agile Releases AI-Native SAFe (June 23, 2026)
- Scaled Agile Framework, AI Value Architect
- Scaled Agile Framework, AI-Native SAFe Is the Operating Model for the AI-Native Organization