Last month, I hosted a dinner in New York for a group of CIOs and CTOs. The discussion opened with a debate over frontier and open-source models. It ended with a table of technology leaders agreeing that their data, more than any model, will decide what they get from AI.
A word about the group. Last year, we started the Unisys CIO & CTO Community, an invitation-only circle of senior technology leaders who meet regularly, in person and virtually, to compare notes with their peers. Members set the agenda. We keep the groups small and the format conversational, with no presentations, and we follow the Chatham House Rule. So, I'll share what we discussed without naming anyone.
The group's conclusion is one I share. Along the way, we covered the state of enterprise data, where AI processing should run, security, and how to avoid locking into a single model.
Confidence is running ahead of results
Start with the state of enterprise data. Members described data that is still fragmented and unstandardized. One noted that enterprise data is often scarce and hard to access, with much of it sitting outside the enterprise or with independent partners.
Our research tells a similar story. In the Unisys AI & Cloud Insights Report 2026, 90% of organizations say they have the architecture to support data-driven decisions. Yet the share exceeding their operational efficiency goals fell to 65%, down from 80% last year. When we asked what limits their ability to scale agentic AI safely, the top answer was lack of visibility over where data and AI workloads reside (58%).
Leaders are confident in the foundation. Proof that the foundation is working is harder to find.
Last year, I wrote that poor data gives you wrong answers faster and on a greater scale. Agents raise those stakes. A generative AI tool hands a person a flawed draft to review. An agent can act on flawed data before anyone looks at it. If you can't say where your data lives, you can't say with confidence what your agents are acting on. Knowing where the data lives is the first test. Knowing what it contains, including sensitive information in places it was never meant to be, is the next.
Bring the AI to the data
The traditional instinct is to consolidate, pulling data out of the systems where it lives and moving it to a central platform before putting it to work. With AI, that instinct deserves a second look. Every move adds cost and latency, creates another copy to govern, and can strip away the context that made the data trustworthy in the first place.
Members offered two examples of bringing the AI to the data. The first was anchoring AI processing to a stable, well-structured on-premises source rather than migrating the data elsewhere. The second was relocating compute physically closer to a real-time data feed, purely to reduce latency. Neither is exotic. Both start with knowing exactly where the data lives and why it's there.
This is where hybrid infrastructure matters. In our research, regionalizing AI workloads is the most common response to data sovereignty requirements (48%). As those requirements tighten, flexibility counts. Organizations that can move workloads across on-premises, regional and cloud environments as the rules change have options. Those tied to a single provider or jurisdiction have far fewer.
Design for flexibility
On the model debate itself, most members agreed that open-source and open-weight models now handle most use cases well, with frontier models best reserved for the hardest, highest-stakes problems that open tools can't yet solve.
The group also flagged a related concern. Model capabilities and market leadership shift every few weeks. Committing an entire architecture to a single model or vendor is an increasing strategic risk. Several members also noted that internal politics and leadership preferences still shape these decisions as much as workload requirements do.
Their answer was practical:
- Justify each model choice by what the workload needs.
- Put a vendor-neutral layer between your infrastructure and your models, where guardrails and integration logic live.
- Design for change, so a better model can replace the current one without a rebuild.
Our research points in the same direction. More than half of organizations (57%) cite concentration among hyperscaler cloud providers as a barrier to safely scaling agentic AI.
This connects directly to our work. Orchestrating AI across a hybrid environment, with the right guardrails and the freedom to change models as the market moves, is the work we focus on with clients.
Guardrails come first
Security was the liveliest topic. Several members described the same pattern: an early push to give everyone AI tools and build fluency, followed by a pullback once they saw where real exposure was happening. People discussed the risk of sensitive data slipping out through creative workarounds and routine, everyday use of tools. A determined employee, or simply an unaware one, can expose sensitive data even with strong policies in place. Data can also expose itself. Sensitive information sitting in a field never meant to hold it can surface in a model’s response months later, so finding it before an agent does is part of the job.
The consensus was to spend on guardrails and security before the next round of AI investment. I made a related point last month in The AI governance question every leadership team should be asking: governance has to move at the same speed as adoption. Funding guardrails first does not mean holding back adoption. Access and rules should go out together. With nearly three times as many organizations reporting a breach in the past 12 months as in our previous survey (49% compared with 17%), this is not the place to economize.
Questions to bring to your leadership team
- Can we map where the data our AI systems depend on lives today, including data held by partners?
- For our priority AI use cases, would it be cheaper and safer to move the compute or move the data?
- If a better model came out next month, how long would it take us to switch?
- Are guardrails and security funded ahead of our next AI expansion?
For any leadership team weighing its next AI investment, I'd start where this group ended up: get clear on where your data lives and how it's governed. The model decision becomes much easier after that.
If your team is working through where AI should run and how to keep your options open, let's talk. We help organizations orchestrate AI across hybrid environments, with the governance and security to scale it.
