AI keeps getting more capable. Deploying it inside an enterprise, however, is not necessarily getting any easier.

As companies move beyond experimentation and put AI into more parts of their businesses, they are running into a separate set of challenges. The questions increasingly center on which models should handle which tasks, whether the underlying data is good enough, who — and what — AI systems are allowed to access, and whether existing governance can keep up.

Several developments this week point to the same conclusion: the next phase of enterprise AI may depend less on access to the latest models and more on whether companies can actually manage them.

Multi-Model Operations Become the Norm

Enterprises are no longer simply choosing an AI model. They are running multiple models, each with different capabilities, costs and risks. That means someone has to decide which model handles which task — and when those decisions should change.

Payments and data company Deluxe, for example, has more than 50 AI agents, with a centralized gateway directing requests to different models. The company weighs factors such as quality, risk, speed and cost when deciding which models to use.

That is a markedly different challenge from picking a single AI provider. As enterprises add more models, model selection itself becomes an ongoing operational function — one that turns a periodic vendor choice into recurring work: routing requests, comparing outputs and revisiting those calls as quality, cost and risk profiles shift. (Related: Gemini 3.8 Live Transforms Conversational AI.)

Data Quality Still Undermines AI Projects

Managing the models is only part of the problem. AI projects are also hitting a familiar enterprise roadblock: poor or fragmented data.

In a recent Collibra survey, 72% of AI decision-makers said a poor data foundation was the root cause when enterprise AI initiatives fell short. The figure is a reminder that when AI initiatives stumble, the problem often predates the model: it sits in the data feeding it.

Governance Lags Behind Agentic AI

The operational questions do not stop at data. Companies may have AI governance policies in place, but their strategies may not account for agentic systems.

An EY report released on Tuesday found that nearly six in 10 respondents at organizations using agentic AI thought that no single group oversaw agents after deployment. Nearly half said their governance frameworks had not been updated to address agent-specific risks, while four in 10 lacked visibility into all the AI tools on their networks.

That is as much an operational gap as a governance one. Companies cannot effectively manage AI systems if they do not know what is running or who is responsible for overseeing it. The findings leave organizations with two immediate questions to answer: who owns agent oversight once agents are live, and how to build a complete picture of the AI tools already running on the network.

The Organizational Challenge

A more powerful model will solve none of these problems. They require companies to make decisions about architecture, data, ownership and oversight as AI becomes embedded in more of the business.

The hard part of enterprise AI may no longer be getting access to powerful technology. It is building an organization capable of operating it — and the indicators to watch are internal ones: data readiness, governance coverage and clear lines of ownership. (Related: Calls for AI slowdown raise new challenges for open-weight models.)

Also This Week in AI News

• AI Is Now Leading Driver of New Cybersecurity Spending: Seven in 10 CISOs now rank AI as their top priority for new cybersecurity spending, particularly for security automation and identity and access management.
• Altman, Amodei Talk AI Pacing at Dreamforce: The AI lab leaders used Salesforce's Dreamforce conference to discuss slowing the pace of frontier AI development as concerns grow that capabilities are advancing faster than safeguards.
• Gemini 3.8 Live Transforms Conversational AI: The Google model adds real-time reasoning, tool use and visual processing while maintaining natural voice interactions.
• Calls for AI Slowdown Raise New Challenges for Open-Weight Models: Calls to slow frontier AI development could create new challenges for open-weight models, potentially leaving enterprises with more responsibility for testing, monitoring and governance.
• AI Panic Is Giving CIOs a New Trust Problem: Rising anxiety around AI is creating a new challenge for CIOs as they try to build organizational confidence in enterprise deployments without downplaying legitimate concerns about the technology.
• 'We're Going to Stop Talking About AI,' and Other Industrial Predictions: Manufacturing companies expect AI to become less of a standalone initiative and more embedded in everyday operations. (Related: AI changes the ROI equation. Here's how some have found success.)
• AI Makes a Mess of the Tech Job Market: AI is reshaping the tech job market, with an analysis of nearly 50,000 engineering job postings finding shifting skill requirements and a growing number of specialized roles.

About the Author

Liz Hughes is an award-winning editorial and content strategist with decades of experience across newsroom publishing, digital media and executive communication. She most recently served as editor of AI Business and IoT World Today, leading editorial strategy, daily coverage and newsletters reaching more than 1 million readers each month. Her work focuses on AI and emerging technologies, with an emphasis on separating real-world adoption from hype, and she has covered generative AI, robotics, autonomous systems and enterprise technology. Hughes is also the founder of LTH Media, a strategic editorial practice, and continues to contribute editorial work to leading technology and business publications.

This article, written by contributing writer Liz Hughes, originally appeared on AI Business on September 18, 2026.