While still at its early stage, agentic artificial intelligence (AI) is garnering significant attention as potentially the most game-changing wave of the AI revolution.

As we will see, agentic AI promises to present accountants and other finance professionals with a suite of valuable use cases. First, let’s step back and take a context-setting look at how this technology has emerged.

Learning curve

According to Karen Ko, Managing Director APAC Financial Services Transformation at global business consultancy Protiviti, agentic AI marks the latest evolutionary stage of AI innovation that began decades ago with early, rules-based systems.

In the machine learning era, you would train an AI model with a highly specific dataset to perform a single task, Ko explains. Significant human intervention was often required to fine-tune results, and accuracy varied.

In late 2022, with the arrival of ChatGPT and similar tools, Generative AI (GenAI) emerged as a breakthrough. Unlike machine learning, GenAI uses pre-trained models – such as Large Language Models – equipped with vast, often unstructured datasets, to enable rapid deployment and versatility.

“No longer was it necessary for users to build models from scratch,” says Ko. “Instead, they simply provided prompts, and the system would perform multiple tasks. For example, semantic search and text, graphics or video generation.”

In the past year, she notes, the focus has turned towards agentic AI, which takes the form of one or more ‘AI agents’ that can carry out actions on your behalf, almost like people would.

“This means that you can automate processes by using multiple agents at once to handle a range of tasks, such as data extraction and summary, preparing reports, reviewing and making decisions. You set the goal, and the agents work autonomously to meet it.”

Taking initiative

At this early stage of agentic AI’s adoption, Ko says, the most mature and impactful use cases are in customer service, marketing and sales. Here, AI agents can support a variety of tasks at scale, such as triage of customer enquiries, sentiment analysis, campaign orchestration and personalised engagement. They can also produce sales and marketing materials.

Adoption in finance is gaining momentum, however. Leading enterprise resource planning platforms such as SAP, Oracle, Workday and Microsoft Dynamics are actively embedding AI and agentic capabilities into their workflows. Those enhancements aim to streamline processes such as forecasting, reconciliation and compliance management.

However, compared to customer-facing domains, where agentic AI applications are advancing rapidly, finance solutions remain in an earlier stage of maturity. This is largely due to the inherent complexity of financial operations and stringent regulatory requirements that demand careful implementation and testing.

That said, Ko has observed a proactive approach among some clients. Rather than waiting for major software providers to release fully developed solutions, these organisations are taking the initiative and building their own AI agents. A popular venue for this is Microsoft Copilot Studio.

“We’ve recently run a series of training sessions for our clients on how to use Copilot M365 for productivity improvement,” Ko says. “To deepen those skills beyond basic M365 training, we organised two-day AI Agent Boot Camps using Copilot Studio. The sessions enabled professionals with non-tech backgrounds – such as operations specialists, finance staff and actuaries – to learn how to design and deploy AI agents with intuitive, low-code tools.”

One interesting example that came out of those workshops was an AI agent built to resolve payment disputes: cases where a client has paid a bill, but the provider’s system has not recognised it and issues a reminder. “Usually, that scenario would turn into a long battle between customer service, operations and finance,” Ko says. “But the agent brought the resolution time down from a week to just 15 minutes.”

Agents of flexibility

Based on her close view of how AI agents are evolving, Ko sees potential for popular use cases in four finance areas:

1. Analytics One example could be an expense analytics AI agent that automates monthly analysis. It could do this by retrieving multiple reports from enterprise storage systems such as SharePoint or OneDrive, consolidating the data and performing variance checks against budgets. It could then highlight anomalies, generate visual summaries and deliver insights directly to finance managers – reducing manual effort and accelerating decision-making.

In a multi-agent framework, that could mean you have:

• a data retrieval agent to retrieve reports;
• a data processing agent to clean data;
• an analytics agent to perform variance analysis;
• a visualization agent to generate charts and dashboards;
• a communication agent to deliver reports in preferred communication channel, and
• a supervisor agent oversees all tasks and timely execution.

2. Reconciliation In a multi-agent model for bank reconciliation, different agents would fetch expense reports, pull bank statements and match transactions, while a fourth would monitor discrepancies and trigger alerts. A communication agent would then send notifications to relevant teams, while a supervisor agent coordinates tasks and validates outputs.

3. Reporting Agents could compile and validate financial data, then visualise it in charts or dashboards and draft wording for management or annual reports. In tandem, a supervisor agent reviews outputs for completeness and timeliness based on the finance calendar, while a communication agent distributes draft reports to relevant stakeholders.

4. Risk management AI agents could monitor bank statements and other cash flow data for early identification of liquidity risks. From there, they would generate a series of scenario-based recommendations – such as reallocating funds or outlining potential hedging strategies – to support proactive financial planning.

In the past, Ko points out, automation in finance was largely rules-based, limited to pre-defined workflows and static logic. By contrast, she says: “Agentic AI introduces adaptive reasoning, which enables agents to analyse data, make informed recommendations and work towards defined goals.”

However, while AI agents can be powerful allies for boosting efficiency, Ko stresses that human oversight remains critical. “Organisations should adopt a ’distrust and verify’ approach,” she notes, “with a ‘human in the loop’ protocol to validate outcomes.”