Recent months have seen a flurry of product releases with generative AI technology. New tools can now do much more than sophistically generate human-like text or create realistic artifacts such as images, sound, video, or physics simulations. A new class of AI tools makes decisions and takes actions directly, often on behalf of their users. Such functionalities have emerged due to advances in AI task orchestration platforms coupled with a growing distillation of large language models (LLMs) into intelligent personal assistants able to help users do knowledge-based work more productively. Recent product releases and ongoing initiatives in intelligent automation can be expected to produce further efficiencies. Smaller enhancements continue to be introduced both in specialized AI systems for processes and workflows, as well as collaborative AI agents designed to work alongside humans in teams.
Productivity improvements from agentic AI tools may materialize faster and with higher impact than from earlier generations of AI. These capabilities can briefly be characterized and their use in Singaporean offices examined. Five major categories of agentic AI tools can be identified based on the sourcing of their main productivity boosts: AI task orchestration platforms; intelligent personal assistants for knowledge work; automations for workflow and process mining; collaborative AI agents for team coordination; and AI-driven decision support systems. While the number of vendor solutions available is large, investments in product development and research have yet to deliver the same breadth of functionality or specialized tooling associated with mainstream applications in document generation, image synthesis or machine translation.
Tool 1: AI Task Orchestration Platforms
Highly capable task orchestration platforms with AI agents supporting user interactions are emerging. The most widely deployed are commercial Task Orchestration Platforms (TOPs) capable of automating workflows across organizations. TOPs from vendors like ServiceNow, UiPath, and WorkFusion combine business process orchestration with process and task automations that can be connected with chatbot extensions. General-purpose language models with advanced memory and task orchestration capabilities included in products like Google Duplex and Microsoft Copilot, and those from Anthropic, OpenAI, and other research labs, are being directly integrated into business applications such as Microsoft Dynamics, Salesforce, and Zendesk. Parallel developments in the robotics domain are leading to the availability of capable agents for robotic process automation and physical task automation in limited domains, like autonomous driving.
TOPs, when leveraged in Singapore’s knowledge-based economy, hold the potential to significantly alleviate cognitive overheads in office settings. Augmenting knowledge workers with intelligent assistants that manage task orchestration and memory can substantially increase productivity as measured with output rates for workers with simple writing and analytical tasks. Nonetheless, debiasing and falsification checks remain crucial components for ensuring quality and preventing reputational loss in knowledge work. Understanding, estimating, and scheduling the effort required for communication and clarification by a human worker, along with the cognitive overload from managing those interactions, can assist in measuring quantitative returns on investment when deploying such AI assistants within a team environment.
Features and Capabilities
AI task orchestration platforms tackle complex productivity challenges through a combination of natural language processing, agentic automations, and integration with a broad set of enterprise systems. These tools take users’ requests expressed in natural language and attempt to fulfil those requests by finding appropriate automations or automating agents to perform each of the subtasks. For instance, an AI conversation with such a platform may start with the user asking to prepare a report on their finances for the past calendar quarter. Based on its understanding of such a request, the AI would imply the need to (i) obtain the relevant financial records, (ii) prepare a summary analysis, (iii) design a report template, and (iv) fill in the report. While some of these subtasks would ideally be performed by automating agents, others may require human judgment at the time of completion for example, filling in the summary analysis allowing for chatbots to be effective throughout the process even if they lack full automation capabilities.
Productivity improvements can come from direct benefits (e.g., higher-quality products in shorter cycles) and indirect benefits (e.g., service-level agreements being fulfilled more consistently, leading to increased client trust). In the case of AI task orchestration platforms, the largest benefits are expected to arise indirectly from the enhanced skill matching real or perceived between team members and the tasks they are involved in, along with more effective communication and coordination across teams. Suggested metrics for monitoring these productivity outcomes include the quality and cycle time of quantitative and creative deliveries (such as reports and presentations), cycle times of time-sensitive deliveries (e.g., crisis responses and RFPs), the quality of human resources allocated to critical projects, and basic measures of social performance across departments, clusters, and the firm as a whole.
Implications for Singapore Office Practices
Such platforms and capabilities are projected to significantly reshape how knowledge workers in Singapore perform their jobs. Task orchestration platforms automatically breakdown complex business objectives into granular tasks and dynamically assign, optimise, monitor, and remap task execution to various AI agents, software programs, bots, and/or people across the organisation. The agents can continuously learn from completed tasks and improve the efficiency, quality, and speed of subsequent task execution by themselves. Companies adopting these platforms are experiencing a dramatic reduction in needless context-switching, while free and up-to-date personal assistants such as ChatGPT help lighten cognitive loads by making brain dump articulations more natural and informal. Better still, the workers equipped with these capabilities and performing knowledge tasks are on average producing 48% more output.
Embracing these new capabilities, teams are starting to introduce lightweight processes for improved coordination with other teams. For instance, a user might announce and post the agenda of an approaching meeting, the participants use freeform generative writing to capture their inputs, and another person or a ChatGPT agent assembles the inputs in bullet form and upload the draft for review before the meeting. Some businesses are even experimenting with appointing a generative AI ChatGPT bot to play the role of a team member to facilitate further exploration of a topic beyond what human members alone could collectively pursue. However, early experiments have also surfaced a daunting question: How are teams to best communicate and govern such model-specific advice and recommendations?
Tool 2: Intelligent Personal Assistants for Knowledge Work
The second category of agentic AI tools helps knowledge workers accomplish tasks by performing activities on their behalf. These Intelligent Personal Assistants (IPAs) can fulfil, or guide the fulfilment of, requests communicated in natural language. They take an explicit role in the orchestration of activities, and play the part of an intermediary between external parties (the users’ organization’s environment) and information systems. For these tools, successful deployment in Singapore will likely require a comparatively extended investment phase, while productivity gains will follow within the typical corporate project lifecycle.
To date, controlled experiments have confirmed that IPAs can improve productivity for well-defined and repeatable tasks. Such gain is, however, hardly observable in practice. Typical scenarios for IPAs to augment knowledge work at scale still rely on a high degree of supervision. Internal tool usage data reflect this, since most engagement is performed by a marginal percentage of employees, supporting activities that much larger teams would complete without them. Corporate clients base their willingness to pay primarily on the low-cost outsourcing of tedious activities, combined with the intimidating pen-and-paper style that makes low-latency assistance truly valuable but also a barrier to adoption for knowledge workers’ superiors. Principal agents further accommodate the tools’ shortcoming by compensating the information loss arising from the resulting limited internal supervision.
Deployment Considerations in Singapore
Deployment of intelligent personal assistants for knowledge work may be complicated in Singapore by the heavy reliance of local companies, particularly banks and service providers, on the confidentiality of the information that employees deal with throughout the workday. These agents are designed to take action on behalf of the user, leveraging real-time knowledge of the specific data processed by the users and contextually applying this knowledge. To enable such automation, users must feed sensitive and proprietary information into the assistants, sharing data, board papers and related papers written by various stakeholders and uploaded prior to discussions in order for the assistant to understand these documents, pick out critical info, and task different groups to prepare follow-ups after the meetings. To do so, communication with the assistant takes place in natural language, i.e., voice, text, or other forms that imitate human conversation, following dialogue-centric call flows.
Such implicit sharing of information raises serious privacy and security issues. User-based systems have no level of oversight at information-sharing levels, and the organizations’ data protection policies become absent. Addressing the concerns of information theft, privacy, and confidentiality through policies on appropriate and responsible use similar to AI class public sector chatbots can be explored. Singapore banks and many other data-sensitive organizations in the region remain cautious on data confidentiality issues. To date, use of AI chatbots has been restricted largely to customer-facing systems only, limiting the potential for productivity benefits going forward.
Productivity Outcomes and Metrics
Consequently, the selection and deployment of intelligent PAs in Singapore must consider the expected productivity and Return on Investment (ROI). Dufi and Petersen argue that these aspects have received scant attention. While, to date, they have defined productivity as the ratio of the effort expended by the PA to the total effort expended by the person, the assertion made here is that it should be expressed as the ratio of the effort saved by the PA (in absolute terms) to the total effort expended by the person. Such an approach is in line with Allen’s recommendation to focus on the usefulness rather than just the usability of a technology for it to be readily adopted by users. It allows the use of a wider range of impact metrics for the assessment of PAs, including task completion time and completion quality, as Duf et al. aptly note.
The development of appropriate productivity measures is crucial to increase the acceptance of intelligent PAs and facilitate their integration into the corporate culture. A study assessing the acceptance of intelligent PAs within a large financial services company in Singapore found that respondents would prefer to use PA services if those services could provide a greater value proposition than cost. Favourable penetration rates among staff are linked to reduced effort as well as improved performance and outcomes, and such expectations serve as adoption motivators for employees.
Tool 3: Automations for Workflow and Process Mining
The productivity tool class referred to in the IT industry as “automation” programs or bot-like agents that take over repetitive tasks in business processes and workflows can also be leveraged by disciplines beyond IT. When automations execute rule-based processes such as customer onboarding and loan origination, they free up human employees for more cognitively demanding tasks, meet increasing digital-literacy requirements, and reduce human error. Automations reduce headcount and speed up processes, yielding fast, lucrative returns on investment (ROI) for organisations. Further, automations become even more useful when deployed with a process-mining solution, which maps business workflows, identifies process inefficiencies, makes automation suggestions, and tracks the success of automations. Automated “Decision Support Systems,” which recommend actions rather than executing them directly, can also fall under this productivity tool category.
If automations are to remain key productivity tools, their adoption must satisfy requirements based on legislation, firm culture, and the nature of the work being done. For example, banks in Singapore must comply with anti-money laundering requirements by performing Know-Your-Customer checks on new customers that mandate viewing and stating what Customer Due Diligence documents are being used, together with external face recognition matching. Automations might restrict checks for named individuals where risk rules indicate a higher risk for automated systems but an exemption is possible in the documentation and Approval workflow mapping. These automations are even more suited for use in less risky departments such as visual element approvals in marketing.
Integration with Local Compliance Requirements
Countries seeking to leverage the productivity benefits that come from the adoption of agentic AI tools will need to revisit their guidelines for data privacy. Businesses in Singapore are already required to comply with the Personal Data Protection Act (PDPA) when storing and managing personal data, and this requirement equally applies when these data are processed by AI platforms. ChatGPT-enabled tools affectively decentralize the task of writing prompts, and unless an explicit communication protocol stressing these hybrid processes and outlining the administrative precautions is put in place and adopted could unintentionally expose data to the cloud and back to the generative AI provider.
Pandemic-induced changes in customer behavior resulted in a shift towards online business in multiple sectors across the world. Demand for accommodation never declined. Short but sudden changes in demand for transportation resulted in losses for airline companies. Airlines are now investing into AI-driven systems and tools to support decision-making in revenue-management and marketing departments. Revenue-management and marketing teams in organisationX faced major challenges in data management and planning following the sudden switch to online business and the economic crisis. AI-driven systems and tools integrated into decision-making will reduce manual workload and enable them to book marketing campaigns more efficiently during uncertain times.
Case Illustrations in Singapore Context
When considering the deployment of workflow automations and process-mining tools in Singapore offices, a bank’s compliance requirements offer one perspective. As part of the country’s critical infrastructure, local banks must meet stringent regulatory mandates. Consequently, any efficiency improvements in regulatory compliance offer both competitive and strategic advantages. Following the implementation of a process-mining capability based on UiPath Automation Suite and other tools, a Singapore bank demonstrated that the technology supports the discovery, design, monitoring, and improvement of processes across the enterprise. Although the project began as an internal initiative for the group operations function, other areas subsequently adopted the technology, including regulatory compliance. Process mining provides visual representations of how work is completed, revealing optimisations to enhance efficiency, reduce costs, and mitigate risks without compromising compliance.
In a different industry, a financial services technology provider illustrated that workflow automation can reduce time spent on repetitive work and manual processes, activity that adds no long-term strategic value. Although more than 80% of its employees had previously expressed concern about onboarding new clients and managing data remediation projects due to the transactional nature of the work, one centre completed five client onboarding projects at a lower cost and with a significant quality improvement following the implementation of a robot to automate the collection. Within a customer support centre, an end-to-end automated solution for customer communications across multiple channels, including web, email, and chat, transformed how a company handles customer requests. It performed only 30% of the workflow prior to automation, with the business case predicting a throughput increase exceeding 300%.
Tool 4: Collaborative AI Agents for Team Coordination
Next on the list are AI Agents focused on a single task and designed to enable team collaboration. While individuals can derive productivity benefits with the increases in focus and personal capacity provided by the previous tool category, collaboration using these agents changes the ingredients of productivity. Savvy teams can leverage AI agents devoted to a specific task to help coordinate business activities, often resulting in better communication across three dimensions: direction, detection, and correction.
Once the communication matrix describing roles and responsibilities has been established across the task-oriented agents, the effect of the agents on team productivity can be gauged with standard performance measures: output per unit of an input measure, such as cost, time, or effort. For example, Singapore teams applying agents assigned to manage the logistics of projects have experienced reduced time spent on coordinating logistics and increased speed of project execution while obtaining additional bidding capacity, resulting in far better winning rates. Measurement beyond these standard KPIs is critical for establishing governance protocols that confirm the continued effectiveness of the agent communication matrix.
Communication Dynamics and Governance
Communication between members of office teams is often inefficient and causes productivity loss. Large workloads require many workers to be on-call for assistance with their own work when they would prefer to focus on their tasks. Impromptu meetings frequently interrupt workflow for all present. Understanding the impact of these patterns can help managers compare productivity levels among groups of knowledge workers and, ideally, adjust team communication flows to reduce the overhead of informal exchanges. Agentic AI tools can automatically monitor and analyze email conversations and other informal communications to allow for better-informed governance of these communication patterns.
One application assesses communication dynamics based on conversations exchanged through the organization’s official email accounts. Dynamics such as group size, reciprocity, and communication density are calculated to provide a detailed overview of communication levels using a set of already established network science metrics. To measure their impact, the patterns are then correlated with productivity metrics that compare output to input. Such tools are particularly useful in measuring staff and team performance for performance-sensitive organizations without the cost of direct supervision.
Measurement of Team Performance
While intelligent personal assistants automate the planning and execution of single-task knowledge work, collaborative AI agents in Singapore offices can augment the work of teams. For example, DinnerExtra is a group decision platform that uses AI agents to gather and analyze information automating the process of restaurant booking for joint meals. The project managers specify the constraints and requirements, such as budgets, preferences, timing, and locations noted by the project managers, as well as any important external factors, e.g., whether it is the busy season. The AI Conversational Agent minds the conversation, and interacts with WhatsApp, Google and Zagat APIs to gather information. The team NeverTooHot built an AI agent that manages planning and logistics for scuba dive teams, or other similar group activities where several hands have not enough time to check all solutions and possibilities.
When deploying collaborative agents that coordinate the tasks of several individuals, Singapore office managers should be alert to the impact on communication dynamics and turnaround time, particularly when tasks have significant interdependencies. For user-initiated interactions, diffusion is usually very fast, but for a tool to automatically learn which everyday type of task would benefit the most from an AI Assistant, turnaround time is more important to control. If the Team Command Control’s logistics plan is not too complicated, then users will also tend to adopt it without a slower accepting diffusion process. Monitoring rework levels, task completion times, handover delays and other relevant metrics can thus reveal how deployment of AI Agents, if any, is affecting Team Performance.
Tool 5: AI-Driven Decision Support Systems
Investor decisions about the allocation of company resources can rely on state-of-the-art efficient market theory, predictors based on the past performance of management, evaluation of risk, company values, and ethics. However, given an anticipated increase in risk, companies, large pension funds, endowment funds, philanthropies, and family offices will be challenged to reconcile their investments with government risk avoidance concerns. Governments facing multi-skilling shortages will use limits on the type of mass immigration sectors, industries, and labour types can employ to influence the directions, managerial characteristics, and reduced risks of selected companies. The new game of global diplomacy will provide unions a bargaining chip underlining the intersection between new labour and workforce skills being created, the cohesion within Singapore and new union agreements.
Decision support agents consider all this information and more to recommend resource allocation across a portfolio of companies. Companies whose management profiles match that anticipated to adequately face the future risk will be bulked up; those profiles decreasing expectancy will be trimmed or sold. Portfolio investments whose management clearly match no new age demands will be avoided or sold. Predicted fund returns will also be sensitive to human and AI investments, the results being combined to determine different sector performance. The decisions and suggestions of the allocation centers of risk-exposed investments will also need to be communicated directly, or trongly recommended, to those companies, thus achieving immediate impacts on dividend patterns–for risk reasons–as well as satisfying new ethical and control paradigms.
Risk, Ethics, and Accountability in the Singapore Setting
Large language models, when employed as collaborative partners supporting decision-making, raise concerns about poor-quality outputs propagating through the decision-making process, as well as about the generation of harmful or unethical responses. Problems closely associated with agentic AI–driven decision-support systems include hallucination (factual error), incorrect reasoning, unreliable sourcing, volubility, and vulnerability to misuse or prompt engineering. Lying, a specific instance of hallucination, may be less of a problem for the current generation of models, but the system’s vulnerability to prompt engineering could potentially make it more exploitable than other models. Integrating such systems into processes functioning in a regulated environment may be a way to guard against these shortcomings: processes could be designed to validate or audit the LLM inputs and outputs, making the decision-support system less than entirely autonomous and allowing an organisation to exercise oversight over the use of the system.
The use of agentic AI–driven decision-support systems in decision-making may serve organisations well. Empirical evidence has begun to emerge showing that prompt engineering can lead to favourable outcomes, suggesting that these systems may in some cases be more capable than their human peers: decision-support systems can serve organisations by providing a collaborative partner that is able to inform the decision-maker while managing less contentious aspects of the decision, realising an additional strategic advantage. Such benefits should be weighed against the costs and risks inherent in such systems, with a close focus on the specific use case and deployment context. Nevertheless, the strategic value of decision-support systems may be compelling, with sources of potential ROI encompassing faster speed to market, contingent scalability, demand generation, content generation, process discovery and automation, upselling and cross-selling, and partnership and channel enabler.
Strategic Value and Return on Investment
Automated Decision Support Systems (DSS) provide contextual insights, recommendations, and options by debugging historical operability data and identifying deviations from normal patterns. Similar questioning and exploration can be made for other potential causes, enabling rapid and continuous investigation of scenarios that could impede corporate strategy execution. Partially automated DSS tools extract, classify, and augment the associated documents for quick presentation and browsing. DSS can also visually query the data for exploration. However, no commercial tools are yet available for continuous real-time updates. Continuous real-time DSS require no formal logging if the major databases are continuously updated. These automated systems will support the timely analysis of commonly asked business questions by querying and processing billions of records operated by many computers.
Local and rapid decision making is crucial for companies deploying multilayer partly or fully automated decision-making agents (DMS), yet is difficult for humans, leading to the use of historical approaches for A decision-making, especially in Asian cultures. Quickly available continuous DSS are therefore attractive for Singapore companies, directing their decisions toward prudent agentic corporate social responsibility (CSR). Thus, Singapore multinational corporations expect good short-term financial returns and returns focusing on long-term reputation for investing in operating multilayer agentic corporate organic agents and their DMS”.
Cross-Cutting Considerations
The agentic AI tools analysed here are likely to generate productivity enhancements in Singapore offices, provided the organisation has the right data copilot for intelligent orchestration, that is, the external and internal data sources, contextual information, and suitable software APIs to enable efficient automations. However, these technical prerequisites alone are not sufficient. Concerns around data privacy and compliance with Singapore’s Personal Data Protection Act and sector regulations could affect adoption, especially of public foundation models. Moreover, the AI-augmented workforce needs to be ready. Adaptation requires addressing employee anxieties towards AI, supporting their adjustment journeys, and contextualising agentic AI tools’ usage for their roles. At the same time, the office landscape is beginning to change, including the economic pressures driving spatial consolidation and variation among different sectors. Acknowledging and managing these interdependencies can help realise the benefits offered by agentic AI tools.
Concerns around data privacy particularly those related to compliance with Singapore’s Personal Data Protection Act as well as the need to adhere to relevant industry regulations, may affect adoption, especially in the case of tools that employ publicly accessible foundation models as their drivers. Consequently, organisations that seek to harness agentic AI tools should assess compliance risk and govern usage from a data-protection perspective. Beyond these data-centric factors, the readiness of the AI-augmented workforce will also have a bearing on the productivity outcomes achieved. Employees’ anxieties regarding potential job displacement arising from AI adoption need to be acknowledged and addressed, thereby facilitating adjustment and adaptation—especially as use cases for agentic AI tools are contextualised for specific roles within the organisation.
Data Privacy and Compliance in Singapore
A full-spectrum policy risk analysis of agentic AI tools in Singapore must consider the regional governance of data use and privacy. The collection, storage, and exploitation of user data should comply with the Personal Data Protection Act (PDPA), which regulates data protection control in Singapore. The PDPA principally governs the collection, use, and disclosure of individuals’ personal data by private organizations. Personal data is defined as data about an individual who can be identified from that data, other information to which the organization has or is likely to have access, or a combination of such data.
Beyond this general regulation, there are industry-specific data protection obligations, including obligations in the banking and financial sectors. The Monetary Authority of Singapore, for example, issues guidelines on the use of AI and data analytics that govern the usage of outside financial institutions and service providers. In addition, sensitive areas relating to defense, law, public safety, and national security can be subject to other strong controls. Thus, the local AI regulatory environment permits significant use of agentic AI tools, but with stipulations concerning data identification, financial sector application, and other sensitive factors.
Workforce Adaptation and Change Management
The introduction of agentic AI tools will necessitate workforce adaptation, change management, and the application of relevant training. Employees will need to acquire skills in designing and managing tasks, prompting AI agents for effective outcomes, and interpreting and acting upon the outputs provided. Teaching the capabilities and limitations of such agentic tools will prevent misuse and allow users to leverage the tools effectively. Incorporating these AI capabilities, including prompt design, into formal training will facilitate successful jobs in the near future. However, such agentic tools are not a panacea; basic training in reading and writing will remain critical.
The decreasing costs of these tools make their adoption for highly repetitive tasks an appealing choice. Yet their value may also derive from the ability to develop solutions at a much-reduced cost. Change management will be critical in determining the proper deployment of agentic-AI tools. Agile development and testing teams in low-risk sectors can embrace such tools with little or no risking-change management support. Teams in domains requiring delivery of multiple AI outputs with quality assurance and compliance approval will require appropriate support and testing phases when piloting adopt agentic tools.
Spatial and Sectoral Variations in Adoption
The considerations outlined above do not apply uniformly in all offices in Singapore. They differ across various industries and segments of the workforce and may not uniformly influence all tool types. For example, Bot-as-a-Service providers who create and manage custom agents for their clients in exchange for a fee gather and use knowledge from multiple clients to improve the performance of an agent over time. Thus, greater adoption may drive improvements, but it does imply a detrimental consequence to the broader office environment. Conversely, although the task orchestration tools are likely to reduce the demand for knowledge workers, they may increase demand for specialists embedded within small- and medium-sized enterprises to help orchestrate the agents on behalf of the smaller organisation. An entirely different geography of adoption arises for decision-makers who deploy decision-support systems that offer agentic behaviour. As usage of these systems increases, holes start to emerge areas where the risk of poor decision-making by the system owners is unacceptably high leading to the definition of governance rules. The degree of specialisation of the decision-support tool and the level of governance required depend, amongst other factors, on the potential strategic value of the decision.
A similar process of uneven impact by sector applies to the introduction of intelligent personal assistants for knowledge workers. Their deployment is associated with enhancing rather than suppressing the usage of the tools yet known to drive productivity productivity although the perception may differ for workers not yet using such tools. The introduction of knowledge-work automations for workflow and process mining alters communication patterns towards the communications-as-collaboration spectrum, the very base of the collaborative pattern that drives the agentic productivity of office workers.
The analysis reveals five IT agent tools that augment productivity by coordinating and orchestrating the other tools used in a workplace. Robustness and agency are emerging in many tools through the use of increasingly large and powerful AI technologies. A prospective analysis of agent tools describes AI task orchestration platforms, AI orchestrators for knowledge work, automation of workflow processes, collaborative agents for team coordination, and AI systems that make decisions. In workplaces where data privacy is critical, adoption of agent tools must comply with information security regulations. Nevertheless, failure to adopt agent tools in organizations requiring task orchestration will hinder performance.
In Singapore, an AI task orchestration platform currently under development at the National University of Singapore simulates large numbers of AI assistant and other deployments in order to optimize unique data environments, workflows, and productivity enhancements for individual organizations. Adoption is anticipated to warrant more than 100% return on investment within one year of initial implementation across most sectors.
Thilina is a multi-skilled digital marketing professional and technical specialist at Sotavento Medios. He manages essential technical SEO audits, search engine indexing, and automated workflows. With experience spanning website management, Google Ads, and campaign execution, he ensures digital assets remain optimized for generative search engines.
