As Singapore accelerates its national artificial intelligence (AI) ambitions, financial institutions face a challenge that extends beyond the capabilities of the technology itself. Simon Ma, Managing Director, Asia at Freshworks, highlights the importance of the systems underpinning enterprise AI, particularly in a financial sector shaped by legacy infrastructure, complex data environments and stringent security requirements.
“The hardest problem in enterprise AI is rarely getting a model to produce an answer. It is getting every system around that model to support it in production,” Ma explains. As Singapore advances its national AI agenda, the implications for financial technology leaders are becoming increasingly apparent.
Initiatives such as the National AI Impact Programme, which aims to scale adoption across 10,000 enterprises while upskilling 100,000 workers to become “AI Bilingual”, alongside IMDA’s AI for Enterprise Impact Playbook, highlight the importance of integration and infrastructure alongside talent and governance. According to Ma, for financial institutions, the next phase of enterprise AI will be defined not simply by what organisations can pilot, but by how effectively their underlying architectures can sustain deployments at scale.
AI Exposes the Seams in Enterprise Architecture
“Adoption rates across the financial sector are undeniably high,” Ma notes. Freshworks research indicates that 90% of financial services institutions globally are actively deploying AI, with nearly half planning to increase their commitments over the next 12 to 24 months.
However, adoption has also introduced considerable operational challenges. In Freshworks’ second Global Cost of Complexity report, 40% of financial services leaders described their AI-driven operational complexity as “very significant”. Meanwhile, 86% stated that AI initiatives had increased their teams’ day-to-day workloads, 31% reported substantial rework resulting from inaccurate outputs, and only 41% felt their deployments were adequately orchestrated across business systems and teams.
These findings point to a fundamental issue. “AI does not solve existing integration gaps. Instead, deploying intelligent agents into legacy environments quickly lays bare the underlying data silos and brittle point-to-point connections,” he added. Ma points out that a standalone pilot running on clean, isolated data often looks flawless. Once that same pilot connects to live core banking systems, fragmented customer records, and varying permission tiers, the underlying technical debt becomes an immediate operational blocker.
The Reality Behind the Build-Versus-Buy Instinct
As Ma notes, the preference for developing AI capabilities in-house remains strong in financial services. According to Freshworks’ data, 43% of financial organisations prefer building custom AI capabilities in-house, compared with a cross-industry average of 34%. The sector’s regulatory scrutiny, auditability requirements and data sovereignty rules make internal development seem like the safest wat to maintain control.
However, Ma warns that every internally built model and pipeline can become a long-term liability for an organisation’s technology infrastructure. Building an internal tool requires more than training an algorithm. Engineering teams must manage model observability, monitor model drift, scale infrastructure, maintain version control, apply security patches and integrate systems. As newer models are introduced, custom-built tools can become difficult-to-maintain silos.
For Ma, the decision requires a more considered approach than choosing between internal development and external solutions. “The critical decision is less about an absolute binary of building versus buying,” he writes. “Instead, it comes down to architectural intent: defining the specific customer or employee workflow that needs improvement before touching code, and choosing the path that introduces the least long-term operational friction.”
Governance Must Sit Within the Runtime Stack
Traditional banking applications operate on deterministic rules, where predictable inputs pass through fixed business logic to produce consistent outputs. AI, by contrast, introduces probabilistic models, non-linear reasoning, external knowledge retrieval and autonomous system actions.
According to Ma, this shift changes the fundamental nature of governance. Risk and compliance teams must be able to establish which data informed an automated credit or service decision, which model version performed the task and whose access credentials were used. They must also identify the downstream systems updated by an AI agent and ensure engineering teams can reconstruct the execution trace when unexpected exceptions occur.
For financial institutions, Ma argues, governance cannot be reduced to a manual policy review conducted after an application has been built. “Identity management, strict role-based access controls, data lineage, audit logging, and human-in-the-loop triggers must be designed directly into the service layer from day one,” he elaborates.
Why the Orchestration Layer Dictates Success
Orchestration is central to addressing these operational challenges. Ma explains that it goes beyond using point-to-point APIs to connect platforms; it is an architectural discipline that enables AI models, software and human operators to exchange context securely without increasing maintenance overhead.
While summarising emails or generating draft text requires relatively little architectural complexity, the demands increase when AI agents handle more complex workflows. These may involve retrieving sensitive customer histories, assessing them against fraud policies, updating account records across core systems and routing complex cases to a live specialist. “If context drops, or if permissions fail between those touchpoints, the entire workflow breaks down, regardless of how advanced the underlying foundation model might be,” Ma explains.
Moving Beyond Isolated Experiments
As Singapore’s financial leaders move beyond the pilot phase, Ma argues that competitive advantage will depend not on the number of disconnected AI use cases, but on the resilience and coherence of their underlying technology estates. “Complexity within enterprise IT is inevitable, but not all complexity is created equal,” he notes.
Ma distinguishes between complexity that supports essential business requirements and that which arises from fragmented technology deployments. As he outlines, “Some is essential to manage business logic and compliance. Some is legacy debt. The rest is self-inflicted, created by fragmented point solutions and unmanaged AI experiments deployed without an integration strategy.” For CIOs and CTOs in Singapore’s financial sector, Ma’s recommendation is to look beyond standalone AI models, invest in robust orchestration and build infrastructure that enables AI to integrate cleanly into daily operations.
