Companies in Singapore are moving to deploy AI agents that can carry out tasks across business systems, but many say their data infrastructure cannot provide the current, governed information those tools need to operate reliably.
Three-quarters of Singapore IT leaders surveyed by Confluent say they were deploying or piloting agentic AI. Yet, 73% say their agentic AI projects had stalled, and half say they had abandoned them, according to the company’s 2026 Data Streaming Report.
The report also reveals that 95% of respondents expected or experienced problems with data infrastructure and quality when scaling agentic AI. The same proportion cited legacy-system integration, while 93% cited concerns about large language model reliability.
The AI context gap
The challenge is not simply how much information a company holds. AI agents need access to information that is current, relevant to a task and subject to appropriate controls.
“It’s not just about giving models raw data. We actually need meaning or context behind it,” says Sean Falconer, Confluent’s vice president of Product, AI Products and Strategy, at a media briefing on the sidelines of the company’s Data Streaming World Tour -- Singapore.
See also: ByteDance, Alibaba pull AI companions as Beijing tightens rules
He describes the gap between what AI models need and what corporate systems can provide as the AI context gap.
An AI agent helping a passenger rebook a disrupted flight, for example, would need current flight availability, seat inventory and the customer’s latest booking details. Information that is incomplete or out of date may still produce an answer, but not one that reflects the airline’s current position.
The gap becomes more apparent when companies move from demonstrations to production. In a pilot, data can be simulated, and systems may not have to meet the same requirements for data quality, governance and trust, says Falconer.
See also: Investing in the anti-AI trade
Meanwhile, in production, companies need to know where information came from, who can access it and whether it is current.
Batch systems meet real-time demands
Many corporate data systems were built for reporting, forecasting and other forms of analysis. Information from operational systems is often collected in batches, then cleaned and prepared for use later.
Although that approach can support dashboards and longer-term analysis, it is less suited to AI agents that need to act on current information, such as in fraud monitoring, customer service and supply-chain operations.
Nearly four in five Singapore IT leaders say inadequate real-time data infrastructure is slowing their efforts to scale AI. Eighty-six per cent say continuous and up-to-date visibility across their business was a priority. The same proportion said effective management of data sovereignty was important, while 82% cited data provenance and tracking capabilities.
“We’re seeing business leaders directly tying revenue to speed and accuracy of data and freshness of data,” says Greg Taylor, Confluent’s SVP & GM for APAC, at the same media briefing.
The constraint is especially relevant for banks, where data is often held across older systems and any new technology must meet strict security and control requirements.
To stay ahead of the latest tech trends, click here for DigitalEdge Section
At a fireside chat during the event, Subhajeet Sur, director of enterprise architecture for data and AI at SMBC Group, shares that the bank had historically been “heavy into batch-based processing”. SMBC is now moving towards real-time data as it upgrades its cash-management systems and builds its AI architecture.
That work includes making core banking data available in real time and exploring applications such as customer onboarding. The aim is to give new services (including future AI tools) access to current banking information rather than data that has been processed and updated later. SMBC is using Confluent as part of those efforts, says Sur.
Cost adds pressure
The data challenge comes as companies assess the cost of using AI models. The amount of text a model processes and generates is measured in tokens, which carry a cost each time a company uses an AI service.
As companies move from experimentation to wider deployment, they are being pressed to show that such spending produces business value. “The free lunch era of tokens is coming to an end,” says Falconer, referring to a period when companies could encourage broad AI experimentation without the same pressure to justify the cost.
In response, companies are becoming more selective about the models they use for different tasks. They are also examining whether records of earlier AI interactions can be reused when similar work recurs, adds Falconer.
Confluent argues that data streaming can help manage both the quality and cost of AI deployment by supplying models with current, governed information rather than forcing them to work through large volumes of stale or poorly organised data. In its survey, 91% of Singapore respondents say data-streaming platforms could improve model reliability by helping ensure information is complete and up to date.
According to Taylor, streaming systems can also record how data moves through an organisation and who can access it. This is particularly relevant to regulated sectors that use AI agents with real-time data.
