UAE companies rushing to deploy AI agents are hitting a data wall, and fixing it means another round of infrastructure spend

UAE companies are racing to put AI agents to work, but their data may not be ready for them.Executives deploying AI across banking and enterprise tech tell EnterpriseAM that fragmented customer records and legacy systems are now among the biggest barriers to taking agents beyond the pilot stage, and fixing that could mean another round of infrastructure spending before the payoff arrives.

The gap already shows up in the numbers: Nearly two-thirds of GCC organizations now use AI for specific business functions, but just 4% have embedded it companywide, according to Cooper Fitch data we covered this month. Only 34% of organizations in a separate Cloudera survey had fully integrated data sources across their environments.

The ambition isn’t the problem: “The real gap [...] is sitting one layer down,” Vice President for MEA, the Americas, and Asia Pacific at WebEngage, a customer data and agentic marketing platform, Hetarth Patel tells EnterpriseAM. UAE conglomerates and retailers may operate several apps and customer databases at once, but before an agent can act reliably across them, those separate views need to make sense as one. Why it matters: A person can spot a bad record and check it against another system. An agent simply acts on it, and faster. “There is a possibility that it will actually send a wrong offer to the wrong person,” Patel tells EnterpriseAM. The risk grows as agents get easier to buy: Dubai Chambers this week launched an AI hub offering members 22 ready-made agents for jobs including contract review and invoice matching.

Banks are feeling it first

Banks have decades of legacy technology attached to the same problem. Mashreq Chief AI Officer Xi Liang says older systems were built to process transactions, not support real-time AI interactions. Customer information may also sit across different platforms with different IDs, making it difficult to form a consistent view of the same person.

Some systems may only refresh every 10 minutes, and that’s a hurdle. “If you want to have a real-time conversational AI experience, then that’s not possible because the data is only as new as 10 minutes ago,” Liang says. That makes data governance, metadata, and integration “just as important” as the AI itself.

Cleaning that up can take time: AIQ, Adnoc’s AI joint venture with Presight, told EnterpriseAM earlier this month that its engineers spend an average of 100 days on-site mapping and consolidating customers’ data before training their AI systems. We covered the problem in Gulf oil and gas here.

The fix is a middle layer, not a rebuild

Nobody is ripping out core systems. Botteq Automation CEO Odeh Mahmoud says even relatively small companies can operate 15 or more applications across the business. Replacing systems that have been running for 10 or 20 years can be like “changing your DNA,” he says. The more practical route is “a middle layer of integrated data, clean data” that makes existing systems usable by AI. Patel says that layer needs to do two things. It has to reconcile customer identities across systems, and it has to give agents business context: who a customer is, what they’ve done, and what the company has already tried.

Banks are already building around that need: Mashreq is working on knowledge, context-management, and memory-management capabilities to help agents interpret and retain information. “I wouldn’t say it’s super mature yet,” Liang says, but she expects banks will ultimately need “proper context management [and] memory management” as deployments scale.

Gartner made the same broader point in July, warning that agentic AI efforts risk failing without an enterprise context layer that grounds models in the meaning and rules behind company data. So the next infrastructure bill may not be another model. It may be making everything the company already knows usable by one.

Start with the use case, experts say

The danger is spending heavily on data before proving the agent is worth it. “The biggest issue I see in the whole industry is finding the right use case,” Mahmoud says. Customer service, KYC, and document processing have been the clearer early wins because they tie directly to lower costs or round-the-clock service.

“The biggest issue I see in the whole industry is finding the right use case,” Mahmoud says. Customer-facing agents, banking KYC, and document processing have been among the clearer early applications because companies can tie them directly to round-the-clock service or lower operating costs.

That means working backward from a business outcome rather than forward from the technology. Patel says companies should pick something measurable — customer renewals or dormant-user reactivation, for example — then identify the data the agent actually needs and test its performance against a control group. “The pilots which scale [...] should have some P&L number, some outcome number and not a technology,” he says.

That could stop the “memory layer” from becoming another blanket infrastructure project. A renewal agent doesn’t need everything the company has ever stored. It needs the right information for that job.

The same logic applies to cost: Mashreq assesses what individual AI applications cost to build and run, including where the compute sits, before deciding how to deploy them. Liang says the bank also wants reusable components so later applications become progressively cheaper to build. The bank is already moving beyond experimentation in selected areas. “For this year, [the capital] was invested in building the foundation for agentic AI, and then next year […] we will industrialize and scale [the technology] across the bank,” Liang says, adding that Mashreq is already seeing measurable returns from some use cases.

The exposure is already there

Some of the risk starts before an agent is ever deployed: Mahmoud says employees are already putting sensitive company information into public AI tools without properly considering how that data may be handled. “The ignorance in this regard is very high in the market,” he says, arguing that data classification and security cannot be treated as an afterthought.

Which changes what the competitive advantage looks like: “The next advantage is not who has an AI,” Patel says. “Everyone will have [it].” What matters instead, he argues, is who has the cleanest customer and enterprise memory for that AI to draw upon. The next phase of the UAE’s AI race may therefore be fought somewhere much less exciting than the model layer: over whether companies can give agents enough reliable context to trust what they do next.