Organisations looking to adopt artificial intelligence (AI) agents will likely face challenges with high cost and complexity, but can address these by tapping reusable components and first understanding why they need the technology to begin with.
IDC has predicted that businesses will move from AI experimentation to "reinvention" this year, fuelled partly by the introduction of AI agents. The research firm adds that worldwide spending in AI will hit $227 billion in 2025, of which 67% will come from companies embedding AI capabilities into their core business operations.
In Asia-Pacific, 97% of business IT leaders already have implemented or plan to implement AI agents in the next couple of years, according to a study released by Salesforce MuleSoft, which polled 400 respondents in four markets across the region, including Singapore and Hong Kong.
Some 93% believe AI will boost their developers' productivity in the next three years, while this number is higher at 98% for those using AI agents.
While traditional AI models on predefined rules and data, agentic AI is an advanced form that possesses a degree of autonomy to solve complex multi-step problems, said Ady Meretz, Asia-Pacific president for Verint Systems, which specialises in customer experience automation.
Agentic AI can, in real-time, independently analyse challenges, initiate actions, and make decisions based on pre-defined goals and data insights, Meretz said in an email interview. The technology is a valuable asset for companies that want to optimise operations, enhance customer experience, and fuel business success.
He noted that AI agents can autonomously analyse challenges, make decisions, and take action without requiring human supervision or detailed instructions.
"Its ability to learn continuously through interactions and data consumption allows it to improve over time, therefore, identifying opportunities for growth and addressing potential loopholes or issues before they arise within a system or operation," Meretz said.
Data, though, is proving a challenge for organisations looking to leverage AI agents for these benefits.
The MuleSoft study reveals that 95% in Asia-Pacific cite integrating data across systems as a barrier in their AI adoption, with 93% noting that data silos are creating challenges in their organisation. This figure is higher at 98% for those using AI agents, compared to 89% of those that are not leveraging AI gents.
On average, just 27% of applications are connected within organisations, which the study states will impact the accuracy and efficacy of AI agents. Respondents in the region currently use 912 applications on average, with those that have deployed AI agents using a higher number of applications at 1,130 on average.
In addition, organisations that have deployed AI agents are using 24 AI models on average, compared to 15 AI models used by companies that have yet to leverage AI agents.
Without data connections, AI agents lack effective understanding
"AI agents are delivering on AI’s true promise where earlier solutions, like copilots, have fallen short," Andrew Comstock, MuleSoft's senior vice president and general manager, said in an email interview with FutureCIO. "Unlike past iterations, AI agents can understand context and autonomously execute tasks. This is driving massive productivity gains and allowing organisations to unlock new growth."
With Salesforce's Agentforce, for example, AI agents can pre-qualify leads so sales teams can focus on actual opportunities, Comstock said.
"By delegating repetitive tasks to AI agents, human employees can focus on innovation, strategy, and strengthening customer relationships. This creates opportunities for new revenue streams and higher profit margins," he added, echoing Meretz's comments.
However, to tap this, integration and APIs (application programming interfaces) are critical to building an agent-ready foundation, Comstock said, noting that agents depend on unified data -- powered by integration and automation -- to provide accurate responses and perform complex tasks.
AI agent outputs depend on connected data that facilitates a comprehensive understanding of context and nuances within user queries.
Comstock said these agents gather structured and unstructured data from diverse sources, including CRM, ERP, and human resources systems, as well as email, PDFs, and Slack, and use the data to make business decisions.
"AI agents are only as smart as the data they access and can only be as useful as the systems they are able to integrate with and act on," he said.
More control comes with added costs, complexity
Meretz noted that cost and complexity also have surfaced as key challenges, in particularly, as organisations decide if they should build their own AI models and tap in-house capabilities or develop applications with external vendors.
Building AI models in-house provides control and customisation, allowing companies to develop LLMs (large language models) tailored for their operations. However, it also requires heavy investment in AI talent and infrastructure as well as continuously refinement to keep pace with market changes and regulatory requirements, he said.
In addition, integrating AI agents with existing comes with added complexity, especially for companies and financial institutions that are dependent on legacy systems and tech stacks, he added.
Comstock noted that AI will change systems that are needed in the organisation, where old systems will be removed and new systems need to be implemented. A similar change happened with cloud computing more than two decades ago, he said.
"Building AI agents isn't just about deploying a model. It's about trust, governance, and scalability," he said. "Many businesses try a DIY approach, only to face hidden costs and prolonged implementation timelines."
He noted that developing reliable AI requires toxicity filters, trust layers, and strong data governance. This requires time, as companies struggle to train AI they can trust, he said.
He suggested that companies start with out-of-the-box, customisable third-party offerings that already are integrated with AI, data, automation, and security.
Meretz recommended that businesses adopt a hybrid approach and go with proven AI models or applications from a third-party vendor. This can help reduce the initial cost of building these models from scratch and allow them to make immediate gains with minimal disruption, he said.
Once these models are implemented, companies then can choose to build out their own LLMs and proprietary AI where needed and integrated them into the existing models, he added.
Clearing up common misconceptions about agentic AI
Above all, safeguards still must be in place to ensure organisations maintain control over AI-driven interactions, while benefiting from automation.
Meretz noted that a common misconception is that AI agents operate without controls and make unpredictable decisions.
"While agentic AI can independently process data and execute tasks, businesses [should] have access to robust tools and guardrails that prevent hallucinations and errors," he explained. These ensure AI agents run within predefined parameters, he said.
And rather than replace human roles, AI agents are most effective when they operate alongside customer service agents, augmenting -- not replacing -- human expertise. Meretz added that common missteps companies make are not fully understanding AI capabilities and having clear objectives.
For instance, they deploy AI agents without a target or business goal and they do not fully comprehend what AI agents can do. These can lead to inefficiencies and wasted resources, he said.
"Companies also underestimate the need for human oversight. AI agents require constant monitoring and finetuning to ensure they meet the business agenda and increase operational efficiency. A lack of human oversight may result in errors, compliance risks or misaligned outcomes," he cautioned.
Furthermore, employees should receive proper training on how to utilise AI agents or their organisation may not be able to fully maximise the potential of AI agents, he added.