Almost every large company already uses AI. Almost none of them scale it without stumbling. A new Cloudera survey shows where the stumble happens: there is no shortage of models, there is a shortage of data foundations that AI can trust and the company can actually control.
The report is called The Great AI Re-Architecture. Below are the numbers that matter most and what we take from them for anyone who wants AI agents working inside their operation, not stuck in a pilot.
What the survey measured
The sample covers companies with at least 1,000 employees in 9 markets: the US, Canada and Brazil in the Americas; the UK, Spain and South Africa in EMEA; India, Japan and Singapore in APAC.
Keep in mind that Cloudera sells a hybrid data platform, so it has a stake in the conclusion. Even so, the numbers match what we see in the companies we work with.
Governance became the bottleneck
95% of companies delayed or canceled AI projects in the last 12 months over data governance, compliance or regulation. More than half (55%) lost six projects or more. It is not a lack of ideas or budget. It is not being able to guarantee who accesses what, under which rule, with which audit trail.
AI projects rarely die on the model. They die on the question “is this data allowed to go there?”
At the same time, 77% already use AI actively and 26% call themselves highly AI-driven. The simple-pilot days are over. The issue now is scale, and scale needs consistent access to governed, high-quality data.
Today’s data architecture can’t carry it
72% say their data architecture needs a significant overhaul to meet their AI goals, and 75% have already changed where and how they store data because of AI. The number one reason for the change is not performance or cost:
Look at the fifth item: a third want to depend less on a single cloud provider. It is the same logic as not tying your operation to a single AI model.
Governance built for people, not for agents
The most interesting line in the report is this one: companies had good governance when only humans accessed data, by hand. When thousands of agents start serving employees, customers and partners, governance needs another dimension.
And data doesn’t sit still. 97% move data between environments every month, and 31% every day, across public cloud, private cloud, on-prem servers, SaaS and the edge. Every move is a chance for the rule to get lost on the way.
In practice, every agent needs an owner, clear permissions, a human approval step where it matters and a log of everything it did. Policy has to travel with the data instead of living in a compliance document. That is where GDPR, CCPA and sector rules stop being a checklist and become system design.
Coming home, and a hybrid future
66% moved AI workloads from public cloud back to on-premises or private cloud in the last year. That is not an exit from the cloud. It is putting each workload where it runs best, weighing performance, cost, latency, governance and availability.
No single environment wins. Over the next two years, 29% plan to spend more on cloud, 25% want a hybrid-first approach, 24% will spend more on-prem and 22% at the edge. The report’s conclusion is blunt: the companies that move fastest will be the ones that bring AI to their data, wherever it lives, without losing control.
Tokens are a new cost line
84% saw infrastructure costs rise because of AI. And the report puts token consumption next to compute and storage units as one more cost variable.
That changes the conversation with finance. An agent that runs all day consumes tokens all day. Without a per-agent cap, the right model for each task and a monthly cost report in dollars, the bill shows up before the results do.
Every region is in the same boat
The three regions are heading to the same place along slightly different paths. All of them passed 50% of companies losing six or more projects to governance.
In the Americas, 81% agree that AI makes data governance more complex and harder to maintain, against 65% in EMEA and 61% in APAC. EMEA, meanwhile, reports the deepest adoption: 47% use AI widely across core operations.
What to do about it
The survey is about large companies, but the lesson holds at any size: a good agent goes to the data, works inside the company’s rules and shows what it did and what it cost.
- Map where your data lives. Cloud, on-prem servers, spreadsheets, SaaS tools, and how often data moves between them. That map unblocks the project.
- Give every agent an owner and limits. What it can read, what it can do, where it stops and asks for approval.
- Log everything. Every agent action goes into a history you can audit.
- Cap the cost. Token limits per agent, the right model for each task and the cost in the monthly report.
- Avoid lock-in. Swappable components, so you can change models or clouds without starting over.
- The AI Analyst builds that map, finds the work that repeats and sets the metrics before anything is built.
- The Forward Deployed Engineer works inside your operation and runs the agents in your own environment, with permissions, logging and human approval at the right checkpoints.
Finance, legal, customer service, procurement and logistics are just examples. If a task repeats, it can be automated, as long as the data stays under control.
AI is already here. The companies that get their data house in order will scale first.
Enjoyed this article?
Get the next one before it hits the blog.
A short email per issue with what actually matters about AI agents at work. No spam.