Imagining what can be done with AI isn’t the hard part. Everything around it is.
In my experience, what truly holds organizations back is the tangle of legacy systems, regulatory exposure, security gaps, siloed data, and the growing weight of “shadow AI” popping up across business units. Decision-makers are asking questions like:
- How do we operationalize AI without breaking what already works?
- How do we build an ecosystem strong enough to handle it?
The irony is, enterprises aren’t just asking questions or thinking about AI. Most are heavily investing in it. IDC reports that AI and automation initiatives are now the most protected line items in technology budgets, even during reductions.
For many leaders right now, AI feels less like a revolution and more like a tug of war. Business teams push for rapid innovation, IT raises alarms about risk, security is tightening controls, legal is flagging compliance concerns… meanwhile, token spend is spiraling and no one has full visibility into who’s building what.
The tension may be discouraging, but I see an opportunity here: it means your organization has reached the point where AI must move from experimentation to ecosystem design.
Here are a few hard-won lessons from helping organizations scale AI past POCs in a way that is safe, scalable, and genuinely useful.
The problem with AI POCs
That exciting Proof-of-Concept your AI Innovation Team put together? It was successful precisely because it could afford to ignore the types of constraints that underpin your whole business – legacy tech, security, legal, governance…
It’s why OpenClaw went viral so quickly, while being completely problematic for any industry that has to operate in the real world.
Dive deeper: AI Governance Lessons from OpenClaw and Moltbook
POCs work because devs are deliberately going to avoid the hard parts – they won’t bother with the oldest legacy system or any security and legal constraints. ‑to‑point setup collapses under real‑world complexity.
Imagine if you released a successful AI assistant on a hospital website, for example, only to discover it was ingesting lab results from a legacy file transfer system that never enforced patient‑level access controls. Now, you’re exposing protected health information.
If you struggle to factor agentic AI into your current architecture, you’re not alone. IDC research shows that 45% of AI projects in the last two years have delivered measurable business outcomes. 1 In other words, more than half have not seen the AI impact they hoped for.
Fragmentation is a silent killer of AI scaling efforts
Many enterprises struggle to realize ROI from AI because they’re trying to scale the technology on ecosystems that were never designed to support autonomous systems.
I recently had the pleasure of discussing this topic with Shari Lava, who leads IDC’s AI & Automation research team. As many organizations look to scale from POCs to implement AI throughout their environments, they realize it looks more like what she calls “islands of automation,” rather than a unified ecosystem.
To unite these islands, some organizations build “bridges.” Often, they are custom-coded integrations designed to move data between disparate applications or different environments.
These bridges help in the short term, but they are rarely built for long-term scalability, visibility, or control.
We saw this in one multi-national company we work with: after some successful AI proof-of-concept initiatives – driven by a newly formed AI team – they hit a roadblock in scaling because their integration landscape was too siloed.
One team was managing EDI, another was responsible for MFT, another for APIs, and iPaaS, and so forth. Because each group owned only its piece of the puzzle, no one had end-to-end visibility or a unified strategy. AI agents needed access to data and processes spread across all these systems, but the teams were not aligned, the technologies were not interoperable, and governance was inconsistent.
The problem is that many vendors today can only expertly address one slice of the problem, as opposed to offering a holistic, multi-pattern integration foundation to truly operationalize AI safely and at scale.
A unified integration layer is the only way to give AI secure, consistent access to data and processes. And especially when it comes to agentic AI, the consequences of fragmentation extend far beyond inefficiency.
AI control is a critical security and regulatory consideration
“Garbage in, garbage out” becomes even more relevant in the era of agentic AI.
AI models need consistent, controlled, real-time data to deliver value. Autonomous systems may amplify integration weaknesses, rather than compensate for them.
We’re not just talking about hallucinations or unhelpful chatbots. In an agentic architecture, you’re no longer managing a few thousand human identities, but millions of autonomous agents.
What’s more, compliance with local and industry regulations doesn’t allow most companies to “move fast and break things” to innovate.
For example: in Europe, companies now must consult their Works Council before deploying AI tools that could affect working conditions. This means an AI rollout can be legally halted even after the technology is ready, if organizational and regulatory approvals weren’t built into the plan.
Enterprises we speak to around the world are concerned about things like:
- AI Act reporting
- data minimization
- regulated data flows
- digital sovereignty
- content filtering
- operational guardrails
- auditing agentic decisions
The imperative to securely unify data so it can be controlled across on-premises, cloud, and external partner ecosystems is more urgent than ever.
You don’t need to modernize the legacy. You need to modernize the flows.
At this point, a common fear I hear from CIOs is: “Does scaling AI mean we have to rebuild all our systems?”
Absolutely not.
Scaling AI doesn’t require tearing out the systems that quietly run the business; rather, we need to look at the paths AI must travel to reach your data safely.
Your core MFT system might be 40 years old, rich with data and expertise. AI needs access to it, and all the rest: processes and information scattered across EDI, APIs, events, partners, databases, cloud apps, and mainframes.
Flows like EDI transactions and file transfers are also rich sources of operational insight – where AI can detect anomalies, enrich data, and trigger actions, provided those movements are visible, controlled, and properly orchestrated.
Rewriting legacy systems is slow, risky, and unnecessary. What we advise instead is to bring an abstraction layer on top of it, which then becomes the connective tissue between:
- users
- LLMs and agentic workflows
- every data source and backend system
It lets you scale AI without destabilizing anything that already works. And, as evoked earlier, this is where multi-pattern integration becomes a significant advantage: the ability to support and orchestrate diverse patterns within a unified framework, as opposed to a one-size-fits-all approach to enterprise integration.
The importance of a unified control layer in a fragmented world
In many enterprises, some of the most valuable AI signals are already in motion: embedded in EDI transactions, file transfers, or partner exchanges. The challenge is making those flows useful, secure, and AI‑ready. As organizations move toward agentic systems and AI‑driven automation, they are governing entire ecosystems of data, decisions, and actions, not isolated applications.
A unified control layer brings coherence to that complexity. It provides consistent access, shared governance, and end‑to‑end visibility across on‑premises, cloud, and partner environments. It enables AI to scale safely without slowing innovation or destabilizing what already works.
This is where Axway’s proven multi-pattern integration heritage becomes a strategic advantage. By unifying diverse integration styles under a common control framework, enterprises can operationalize AI at scale while maintaining security, compliance, and operational resilience.
Ultimately, AI‑enabled ecosystems are not about adopting more tools or chasing the latest model. They are about readiness, laying that sure foundation.
The leaders who will succeed with AI will be those who think in ecosystems, not tools – who can move fast while staying in control of their architecture.
Not sure where to start?
Get in touch with our experts for tailored guidance on turning AI ambition into measurable business outcomes.
1 Future Enterprise Resiliency & Spending Survey Wave 7, IDC, September 2025
2 Future Enterprise Resiliency & Spending Survey Wave 11, IDC, Dec 2024