AI Is Moving Closer to the Decision. Control Is Becoming the Real Infrastructure

The signal

Four developments that look unrelated are pointing to the same shift.

An AI shopping agent has won an important legal round in its fight to act on behalf of users. Meta has released a smaller open-weight model designed to run on personal hardware. A new study found AI systems repeatedly recommending venues that had already closed. In Abu Dhabi, an AI-enabled robotic bronchoscopy platform is helping doctors navigate difficult parts of the lung while the clinician remains in control.

The common thread is not that AI is becoming more capable. That part is already obvious.

The more important change is that AI is moving closer to real decisions and real-world actions. Once that happens, the value of the system depends less on whether the model can generate a convincing answer and more on who controls permission, which data the system trusts, where the model runs and who remains accountable when something goes wrong.

That control layer is starting to look like infrastructure.

What the evidence shows

Permission is becoming part of the product

The Amazon and Perplexity dispute is one of the clearest examples of what happens when an AI agent stops being a chatbot and starts acting for a user.

On 4 August, the Ninth U.S. Circuit Court of Appeals overturned a preliminary injunction that had temporarily blocked Perplexity's Comet shopping agent from accessing Amazon on behalf of users. The court found Amazon was unlikely to succeed, at this stage, on its claim under the U.S. Computer Fraud and Abuse Act.

That is a meaningful ruling, but it is not a blanket right for every AI agent to enter every website.

Amazon is still fighting the underlying case. It can also use technical restrictions, terms of service and product design to make automated access difficult. Reporting from The Information shows Amazon has been tightening those controls against outside shopping agents from several major AI companies.

That leaves two layers of permission operating at once. A user can tell an agent to act. The platform being accessed can still decide how much automated activity it accepts.

For agentic commerce, that distinction matters as much as model capability.

Local AI shifts part of the control boundary

Meta's new Muse Glimmer model points to a different kind of control.

The model was introduced as an open-weight system designed to handle smaller agentic tasks on personal hardware rather than requiring a large remote data-centre deployment for every use case. That does not mean every UAE business can suddenly replace cloud AI with a laptop. Workload, hardware, security and support still matter.

But the direction is important.

When a useful model can run closer to the user or business, some decisions about cost, latency, customization and data handling can move away from a remote platform and back toward the organization using the model.

For UAE SMEs, that could eventually matter more than another benchmark win. Many small businesses do not need frontier-scale intelligence for every task. They need reliable automation for narrow jobs such as document handling, internal search, customer support or repetitive administration.

The smaller the infrastructure requirement becomes, the more practical it is for those businesses to decide where their data goes and how much of the workflow they want to control themselves.

Better reasoning does not fix stale reality

The venue recommendation study exposes another limit.

Researchers audited 2, 208 search-grounded responses from four production AI systems against a census of 4, 776 cafes, restaurants and bars across two markets. Permanently closed venues were still recommended 93 times.

The striking part is that outright fabrication was rare. The bigger problem was stale reality.

That distinction matters for any market where restaurants, shops, clinics, attractions and services change quickly. An AI system can reason perfectly over the information it has and still produce a bad recommendation if the underlying business data is old, incomplete or inconsistent.

For UAE venues, the implication is practical. Being accurately represented across websites, maps, booking platforms, menus and structured business profiles is becoming part of distribution. The customer may increasingly ask an AI system instead of opening ten tabs.

In that environment, data freshness is not a back-office housekeeping job. It becomes part of whether the business is visible and whether the recommendation is trustworthy.

In healthcare, assistance and authority remain separate

The Cleveland Clinic Abu Dhabi deployment shows the same control question in a much higher-stakes setting.

Johnson & Johnson said the first robotically-assisted bronchoscopy case in the Middle East was performed at Cleveland Clinic Abu Dhabi using its MONARCH platform. The case also used MONARCH QUEST, which adds AI-enabled navigation features intended to support the physician in reaching difficult lung targets.

The important detail is what the system does not do.

It does not replace the doctor. The clinician remains responsible for navigation, biopsy decisions and the medical procedure. The AI layer supports the task, but authority remains with the human team.

That is not a weakness in the technology. It is the operating model.

As AI moves deeper into healthcare, finance, commerce and other regulated environments, the most valuable systems may be the ones that make the boundary between machine assistance and human responsibility unusually clear.

Why it matters

These four signals point toward a more mature phase of AI adoption.

The first phase was about access. Who had the model?

The second phase was about capability. Which model was faster, cheaper or better at reasoning?

The next phase is increasingly about control.

For a UAE company, that can mean four very different things:

• Permission control: whether an external AI agent is allowed to act inside another company's platform.

• Execution control: whether a model can run locally or must send every task to a remote service.

• Data control: whether the information feeding the system is current enough to support a real-world decision.

• Accountability control: whether a human, company or regulated professional still owns the final decision.

These layers are easy to miss because none of them looks as exciting as a new model launch.

But they determine whether AI can move from a demo into a dependable business process.

A retailer does not only need an agent that can shop. It needs rules for what the agent is allowed to do.

An SME does not only need a smaller model. It needs to know whether running it locally actually improves cost, privacy or reliability for the specific workload.

A restaurant does not only need to appear in an AI answer. It needs the system to know whether the venue is open, what it serves and where the correct information lives.

A hospital does not only need AI navigation. It needs a clear clinical responsibility model around the technology.

This is why control is starting to resemble infrastructure. It sits underneath the visible AI experience and determines whether the experience can be trusted.

What could change the picture

The thesis is still early and several things could weaken or strengthen it.

The Amazon and Perplexity case is not finished. A later ruling on the merits, a settlement or new platform rules could change the legal boundary for user-directed agents.

Meta's Muse Glimmer also needs real deployment evidence. A model that can technically run on personal hardware is not automatically cheaper, safer or easier for a UAE SME to operate in production.

The venue study should be replicated in other markets. The findings are useful because the researchers compared AI recommendations with a complete market census, but the two audited markets were in Bali, not the UAE. A UAE-specific audit could produce a different visibility and staleness pattern.

Healthcare evidence will also matter. The Abu Dhabi bronchoscopy deployment establishes that the technology is in clinical use in the region. Longer-term evidence around procedure volumes, diagnostic performance and patient outcomes will determine how significant the deployment becomes.

What Pulse is watching next

Pulse is watching for the next evidence that control is moving from a product detail into a competitive advantage.

• New retailer or marketplace policies that explicitly define what outside AI agents may do.

• UAE deployments of smaller or open-weight models where local execution changes cost, privacy or workflow design in a measurable way.

• Evidence that UAE venues are changing how they maintain structured business information for AI-driven discovery.

• More healthcare deployments where AI assists a high-stakes procedure while responsibility remains clearly assigned to a clinician.

• Regulation or contractual standards that define who is accountable when an autonomous system acts on behalf of a user or business.

The biggest AI question is slowly changing.

It is no longer only: What can the model do?

It is becoming: Who decides what it is allowed to do, what it is allowed to trust, and when a human still has to say yes?