A growing divide among leading AI companies over how to secure increasingly powerful models is beginning to translate into challenges for enterprise IT, with implications for how organizations access, deploy, and govern AI systems.

The latest flashpoint came after Meta CEO Mark Zuckerberg called for neutral evaluators to independently test AI models, pushing back on calls from rivals to slow development or tighten coordination.

“trust and alignment are quickly becoming the most important capabilities that will differentiate agents and models. Any lab that doesn’t focus on alignment will fall behind,” Zuckerberg wrote in a post on X.

“Engaging independent evaluators and advisors is industry best practice,” he added, noting that Meta already does this in several areas.

His comments follow a series of public proposals from AI industry leaders including Dario Amodei, who argued for a more cautious pace of development, and Sam Altman, who called for collaboration on safety standards.

The debate has intensified amid disclosures from AI labs and policymakers on potential misuse of advanced systems. Anthropic has said it restricted attempts to use its Claude models in sensitive domains, while OpenAI has engaged with policymakers on AI-related risks, according to company statements and reports.

Enterprise concerns

While the debate is often framed as a choice between slowing innovation and strengthening oversight, analysts said enterprises should focus less on which approach prevails and more on the operational consequences already taking shape.

“Divergent safety approaches will make access to advanced AI models less predictable, rather than producing an industrywide slowdown,” said Sushovan Mukhopadhyay, director analyst at Gartner. Vendors are likely to apply different release schedules, regional availability, access tiers, and usage restrictions, he said, meaning enterprises could encounter similar capabilities “at different times and under materially different conditions.”

Mukhopadhyay said enterprises should plan for variability in access rather than assuming consistent availability across providers or geographies.

“I read this week as the point where frontier AI became a managed supply,” said Bhupendra Chopra, chief revenue officer at Kanerika. “For three years CIOs could assume the next model would simply show up. A frontier model now behaves more like a critical component from a supplier whose delivery dates depend partly on outside reviewers and export rules.”

Chopra added that “any AI roadmap built on a specific model arriving on a specific date is carrying supply risk it hasn’t priced.”

Security pressure builds regardless of slowdown

Analysts said slowing development alone is unlikely to materially change enterprise risk, particularly as open-source models proliferate.

“The biggest point isn’t the pause itself. It’s that the leaders of AI companies are agreeing on something,” said Nikhil Gupta, founder and CEO of ArmorCode.

Gupta said the threat landscape has already shifted. “Even if companies hit pause, open-source AI models are already out there,” he said. “I’m not convinced slowing down some companies meaningfully changes what adversaries can do.”

“Even if AI development slows down tomorrow, security must accelerate,” Gupta added. “The job of securing these systems has effectively gotten ten times harder.”

A new ‘AI assurance’ layer emerges

The focus on evaluation is driving what analysts described as an emerging “AI assurance” layer, where third parties assess models for safety and compliance.

“A distinct AI assurance layer is likely to emerge, but enterprises should not expect a single certification to establish that an AI system is safe,” Mukhopadhyay said. “Enterprise risk also depends on data, system instructions, tools, agents and deployment controls.”

Chopra said enterprises risk misinterpreting such evaluations. “Procurement teams may see a third-party evaluation and treat the model as vetted,” he said. “Within a year it becomes a checkbox.”

Instead, he said, enterprises will need to run their own validation. “CIOs who get ahead will test each model against their own data before it touches production.”

Fragmentation complicates multi-model strategies

For CIOs pursuing multi-vendor strategies, differing approaches across providers could introduce additional complexity.

“Fragmentation was already the default. Safety divergence deepens it,” Chopra said.

He said risk is most acute during transitions. “For an enterprise running several models, the exposure sits in the handoff,” he said. “When a model is delayed or replaced, the system can behave differently.”

“I’d rank untested model substitution above vendor lock-in,” Chopra said.

Gupta said open architectures will be important. “The framework needs to be open, not locked to any single vendor,” he said.

Mukhopadhyay added that enterprises should prepare for models becoming unavailable or restricted.

CIOs urged to build resilience

Analysts said enterprises will need to design AI strategies that can adapt to changes in availability, pricing, and governance.

“For critical applications, CIOs should separate application controls and business logic from the underlying model,” Mukhopadhyay said.

Chopra emphasized flexibility. “A routing layer between applications and model providers turns switching into configuration work,” he said, adding that contracts should cover deprecation timelines.

He also pointed to pricing implications. “Scarce access to the frontier starts to carry a premium,” Chopra said.

AWS is betting that AI agents need a different interface as they move beyond answering prompts and start working autonomously in the background.

The company has open-sourced Pizza Bot, a self-hosted application that gives users an inbox for managing work delegated to AI agents, with separate threads for ongoing tasks and a queue for work that is completed or needs human input, rather than keeping the management of agents restricted inside a conventional chat window.

The rationale, according to AWS, is that background agents do not always need a user’s attention while they work, and an inbox model will let users hand off longer-running tasks, return to them later, and see which jobs are complete or require intervention.

Under the hood

That approach is reflected in how the inbox organizes work with the help of an “All” tab that contains the history of each task or conversation, including the agent’s messages and work performed, the “Unread” tab that flags completed work that users have yet to review, and an “Action” tab that surfaces tasks paused while waiting for user input or approval.

The inbox interface also has a panel named Activity that shows users how an agent handled a particular task along with the transcript, AWS wrote in a blog post introducing Pizza Bot.

AWS’ inbox-oriented rationale also extends to Pizza Bot’s architecture.

It uses LangChain’s Deep Agents as the harness and LangGraph as the stateful runtime, with a combination of the two allowing an agent to checkpoint its progress as it works, preserving its messages, tool activity and current state so a task can be paused and resumed rather than being tied to a live chat session, the hyperscaler wrote.

Pizza Bot’s server sits on top of that stack, connecting the agent runtime with the user interface, skills, MCP servers and the model provider, with developers able to choose from Anthropic, OpenAI, Google Gemini and Amazon Bedrock, as well as local models through Ollama, it added.

As for its out-of-the-box capabilities, the application comes with skills for working with files, browsing the web and delegating tasks to specialist agents. Developers can also add existing Agent Skills and MCP servers to give the agents access to other tools and services, the hyperscaler wrote.

Enterprise integration could remain a hurdle

However, analysts were not too convinced that Pizza Bot’s out-of-the-box skills and support for existing tools will make its implementation easier for enterprises.

While the out-of-the-box capabilities and added support for existing skills saves enterprise teams the time to build the application, it leaves out the integration work required to make it work, said Bhupendra Chopra, chief revenue officer at IT consulting firm Kanerika.

“Integration is where most of the money in an enterprise agent deployment goes. A sales or finance team gets value from an agent when it can read and update CRM, email and ERP, and each of those systems needs a connector someone has to build, secure and maintain,” Chopra pointed out.

Pizza Bot’s integration could be further complicated by the application’s lack of support or any SLA, according to Manoj Chandra Jha, principal analyst at Nord-IQ Research. That means the integration and operational burden ultimately falls on enterprises, who would be responsible for running, securing and maintaining the open source software themselves, Jha said.

Rise in enterprise productivity likely

But for enterprises willing to take on that integration work, the inbox-oriented approach could improve productivity.

“It’s a meaningful shift because it changes the economics of delegating work to agents. A chat interface requires a person’s attention throughout the task, while an inbox brings them in only when their judgment is needed, much like how executives delegate work to their teams,” Chopra said.

“Coding agents have already demonstrated this model, where an engineer assigns an issue and reviews the resulting pull request. Pizza Bot extends that approach to tasks such as meeting preparation and follow-ups,” Chopra added.

Further, the analyst pointed out that the inbox-oriented approach could give enterprise teams a clearer window into how scheduled tasks perform.

“For scheduled tasks, having the results delivered as threads gives users a way to track what happened in the background and spot failures that might otherwise go unnoticed,” Chopra said.

Out of sight, out of mind

There are risks with this approach, however: “The inbox model can make bad work less visible. When somebody is watching an agent in a chat window, they can see it going off the rails,” said Phil Fersht, CEO of HFS Research.

In contrast, when hundreds of tasks are running quietly in the background, users may not see an agent making mistakes until the task is complete or an exception requires their attention, potentially making problems harder to catch early, Fersht said.

That reduced visibility can also create approval fatigue, according to Chopra.

“An agent sending back dozens of threads with multiple approval requests can condition users to approve them without reading them closely,” Chopra said.

There is also a risk in the delay between an agent preparing an action and a user approving it: “Information that was current when the agent paused may no longer be valid hours later, potentially resulting in an outdated CRM update or a meeting invite for a slot that is no longer available,” he said. “An assumption that might be caught in a live chat can go unchallenged as the agent continues building on it and consuming model resources. Scheduled runs can also add costs without anyone watching them in real time.”

Those disadvantages can be countered by designing processes with fewer, better-timed approval points, and verifying that the underlying data is still current before allowing an action to execute, he said.

Adoption likely to start with technical teams

Such tradeoffs are likely to shape where Pizza Bot gains traction.

“Adoption will likely be bottom-up, with individual technologists and small platform teams drawn to Pizza Bot’s control and standards-based design,” Jha said. “Risk-averse industries, especially ones in the regulated sectors, and business users with limited technical expertise, however, are likely to approach the application more cautiously,” he added.

That could leave Pizza Bot occupying a more targeted role in the enterprise, as a way for technical teams to experiment with asynchronous agents and new ways of delegating work, rather than an immediate replacement for the governed interfaces and managed services enterprises typically use for business-critical processes, he said.

This article first appeared on InfoWorld.