In the first week of June, Deloitte updated its "State of AI in the Enterprise" report. On the same day, KPMG released a six-month tracking of its "AI + Manufacturing" special initiative. Add to that Bain's repeatedly forwarded AI budget survey from June 1 — three reports, each independently researched, all saying the same thing.
Enterprises' investment in AI is turning into a "faith top-up": budgets rise year after year, confidence is brimming, but the numbers on the ledger remain unchanged.
Three Reports, Three Dimensions, One Conclusion
First, spread out the data.
Deloitte surveyed 3,235 executives across 24 countries worldwide. 66% of companies said AI helped them improve efficiency — that's the good news. But scroll down, and only 34% of companies have truly restructured their businesses with AI. In other words, two-thirds of companies are still using AI to optimize existing processes, rather than using AI to rethink "how this business should actually be done."
The more striking numbers are hidden on the governance page: only 20% of enterprises have a mature autonomous AI agent governance model. One-fifth. And Deloitte predicts that the use of Agentic AI will increase significantly in the next two years—models are running, but rules haven't kept up.
KPMG has been closely monitoring the manufacturing sector for half a year. How is the implementation of the eight-department "AI+Manufacturing" special action progressing since its policy document was issued last December? 49% of manufacturing companies say they have already realized commercial value from AI. That sounds promising. But the report contains a set of numbers that make you scratch your head: 83% of companies believe they have built a solid AI data foundation, while 76% simultaneously admit that data is their biggest risk.
83% vs 76%: 83% of enterprises consider their data systems to be complete, while 76% admit that data is the biggest bottleneck — KPMG calls this a "misalignment between confidence and capability".
These two numbers together prompted KPMG to give a very accurate description: "a mismatch between confidence and capability". In simple terms, many companies feel their data is well-managed when filling out surveys, but when it comes time to actually run models, they discover that systems aren't connected, master data hasn't been cleaned, and real-time data streams are broken on some server in the middle.
Bain's report directly crunches the numbers. Among 951 large enterprises with annual revenues exceeding $100 million, 90% are still increasing their AI budgets, but only 4% have saved more than 30% in costs. 40% of companies haven't even saved 10%. Bain's headline is straightforward, which translates to: "Your AI budget is rising, but your returns aren't. Here's why."
90% of enterprises with increasing AI budgets save over 30% in costs: 4% of enterprises
Three reports, three sets of data, all point to the same issue: between strategic confidence and execution capability lies a rather wide gap.
It's not that the technology is lacking, but the organization hasn't kept up
There is a detail in the Deloitte report that I have read over several times: In 2026, more enterprises (42%) believe their AI strategy readiness is "relatively high"—higher than in previous years. But at the same time, enterprises' readiness at the operational level in terms of infrastructure, data, risk, and talent has actually declined.
What do you mean? It means the leadership thinks "we are ready," but the middle and frontline levels think "what are you ready for?"
Deloitte calls this a "disconnect between strategic and operational readiness." To put it another way: it's when the PPT is ready, but the production line isn't.
Bain's report confirms the same assessment. They found that 44% of companies are using "expected AI savings" to support new AI investments—first assuming how much money AI can save, then using that assumption to apply for the next budget. Once this cycle is established, no one dares to stop and ask: Did the previous investment actually save any money?
Bain report original statement: The technology works, but the value hasn't arrived. The root of the problem is not the model, but the organization.
KPMG's manufacturing data has concretized this "organizational issue."
| Scene | Landing effect | Instructions |
|---|---|---|
| AI quality inspection of automotive parts | Defect detection rate 92% → 99.7% | The effect is significant, but when scaling to the third production line, we encountered inconsistent data standards. |
| AI energy consumption optimization for chemical enterprises | Unit energy consumption decreased by 8.3% | Effective for a single site, but differences in process parameters when replicating across factories require model reconstruction. |
| Electronic Manufacturing AI Scheduling | Changeover time reduced by 42% | Most effective, but relies on digitizing the experienced master's knowledge — this step is the slowest |
What do these cases have in common? The technology has succeeded in individual points, but the bottleneck for large-scale replication lies not in model accuracy, but in data governance, whether the experience of veteran workers can be structured, and whether systems across different factories can be aligned.
89% of manufacturing executives believe that "managing AI agents" will be the most essential workplace skill within five years. I think this number is more informative when read in reverse—today, only a very small number of companies actually have people who know how to manage AI.
Three Operational Suggestions for CIOs
After reading the three reports, I summarize three judgments that are directly useful for enterprise digital leaders:
First, answer the question "Did it save money?" before talking about "transformation." Bain's data is very clear—only 4% of companies saved more than 30%. If you haven't reached this threshold, don't rush to tell the story of "AI-driven business restructuring." First, run a specific, quantifiable cost-saving scenario. How much labor did intelligent quality inspection save? How much line changeover time did scheduling reduce? Nail this number to the wall before applying for the next budget.
Second, data governance is not about "finishing it first, then applying AI," but about "paying off debts while building." KPMG's report shows 83% claim data is OK, but 76% say data is a bottleneck — indicating that most companies are not engaging in true data governance, but rather data bookkeeping. Issues such as unstandardized master data, disconnected real-time data streams, and inconsistent historical data formats will not automatically disappear just because a large model is purchased. For every AI scenario implemented, fixing a piece of the data pipeline along the way is the only realistic path.
Third, agent governance must be established now. Deloitte says only 20% of enterprises have mature agent governance models. This means 80% of enterprises are putting AI agents into business processes without rules to constrain them. This is not an exaggeration—IDC previously predicted that by 2028, there will be 1.3 billion AI agents running globally, and 88% of enterprises have already encountered agent security incidents. When it comes to governance, waiting until an incident occurs to establish it is too late.
Let me say something very straightforward: In 2026, the dividing line for enterprise AI is not "whether to use AI," but "whether you dare to clearly state how much money was spent and how much was saved." Three reports have already made it clear for you—most companies still don't dare to.
