The hack (almost) everyone’s using: AI agents

By Rosalyn Page

Artificial intelligence (AI) agents have quickly emerged as a must-have tool for business. 

In Australia, 64 per cent of finance and insurance SMEs are already using AI according to National Australia Bank’s Embracing AI report. Agents have the potential to help manage payment processing, reporting and reconciliations.

Security, ethics and accuracy must be considered to minimise risk. AI agent adoption requires defined use cases and robust guardrails for organisations to enjoy the benefits.

What is an AI agent?

An agent can complete tasks using artificial intelligence. Large language models (LLMs) such as ChatGPT serve as the basis of agents, acting as the “brain” in the process. 

In an agentic workflow, an AI agent is set a task via instructions. It accesses an LLM for reasoning, planning and breaking the task into defined steps, and utilising tools to complete the task. The output is reviewed by a human and, if necessary, refined.

Unlike ChatGPT and Microsoft Copilot, which use a two-way conversation function, agents act and use tools autonomously.

“Tools can take many forms — they can query an object or a document, connect to your email or ERP system to extract information, or carry out a web search,” says Harshu Deshpande, group VP for product and engineering in ANZ at Publicis Sapient.

AI agents in finance can handle tasks such as processing invoices, reconciling accounts, generating financial reports, flagging anomalies and executing multi-step workflows.

“There are a lot of manual activities done on a repeated cycle that you can turn over to an agent and gain a lot of operational efficiency,” Deshpande says.

Agentic AI for finance operations

Many workplaces are looking to incorporate agents into their workflows. According to a McKinsey survey, 23 per cent of respondents report their organisations are actively scaling an agentic AI system in at least one business function, with an additional 39 per cent in the experimentation phase.

AI agents are moving beyond simple automation and toward more sophisticated cognitive labour. David Rajkovic, regional vice president ANZ at Rubrik has found that AI enhances investigative auditing, enabling the rapid detection of anomalies and systemic fiscal irregularities.

In analytics, where vast datasets are synthesised, agents are used to extract actionable trend insights. “This is providing a competitive edge in market volatility,” he says.

The other area in which agents can make a difference is trading. “AI agents are being deployed to execute complex, high-frequency stock trading strategies,” Rajkovic adds.

Agents can also help those professionals who manage high volumes of routine transactions, repetitive reporting cycles and large datasets.

How can I build my own AI agent?

Building personal AI agents for productivity allows finance professionals to automate routine tasks to reduce time and focus on higher value work. 

Finance professionals can start with no-code tools that require little more than logical thinking and clear instructions. 

“Today, almost anyone can build agents using low-code or no-code providers like Microsoft Copilot or n8n,” says Rajkovic, adding that the skill set varies depending on the agent’s complexity.

A dedicated few may want to progress towards more customised solutions with low-code and build-your-own agents, but this path requires more familiarity with data structuring, programming language and AI frameworks. 

 

"There are a lot of manual activities done on a repeated cycle that you can turn over to an agent and gain a lot of operational efficiency."

 

With agentic tools now available in products such as Microsoft Copilot Studio and Google Gemini Enterprise Agent Platform, finance professionals can begin experimenting with agent workflows in line with organisational policies. Pre-built templates are available to guide structured workflow design. 

“If you are working with annual reports, there is an in-built workflow that lets you ingest several reports and then use natural language to make queries about them,” Deshpande says.
Other finance-related agent templates include invoice processing, procurement approvals, budgeting and financial analysis.

A note on company culture and AI implementation

When adopting agents in finance and accounting workflows, tools and capability are only part of the equation. Creating and using agents may require a culture change to embrace experimentation and manage risk considerations with strong governance.

“Unless you put these tools in employees’ hands and give people the permission to experiment,” Deshpande says, “you do not know what use cases are in an organisation.”

Ethical considerations for AI agents in accounting

When it comes to agent deployment in finance workflows, there are important ethical elements to consider. Inaccuracy and explainability of output are two of the most reported risks of AI, although according to McKinsey’s The state of AI in 2025 report, more organisations are implementing mitigation strategies.

High AI performers are more likely to have defined processes to determine how and when model outputs need human validation to ensure accuracy, the report notes.

“A finance environment’s specific risk appetite dictates their level of acceptance. While some applications allow for a higher tolerance, critical financial functions demand absolute precision,” Rajkovic explains.

“For these high stakes environments, even marginal errors are unacceptable. Consider transaction integrity — bank balances must be exact, as losing even $0.01 during a transfer constitutes a critical failure.”

 

"You need the ability to map workflows clearly — defining triggers, actions and outcomes. Then iteratively refine through testing. Basic systems-thinking is more important than heavy technical development."

 

Organisations need to be explicit about what agents should and should not do, and build processes that ensure a human will always review their work. The rules need to reflect that LLMs are not deterministic and are, in fact, prone to fill in gaps in output. Rajkovic suggests “a rule might be: if you are missing data, do not guess”.

Regulatory reporting is another area where errors are unacceptable, he says. “Reports for regulatory bodies such as the Australian Taxation Office require total accuracy to prevent legal action or heavy fines.”

Likewise, audit trails have no margin for error. “Documentation must be 100 per cent verifiable and precise,” Rajkovic continues.

With AI, transparent governance is a business requirement rather than an optional extra, particularly in finance and accounting. Frameworks need to ensure AI remains ethical, transparent and compliant with internal policies such as data access restrictions or bans on agents modifying production code.

“As autonomous agents become central to operations,” he says, “organisations must transition from manual oversight to real-time monitoring and automated controls.”

Jasmine Liu, financial controller at Cubic Promote, has built several agentic workflows for her organisation.

The most advanced one monitors and categorises outstanding invoices and flags accounts with urgency tiers based on overdue length. It also pulls in contextual signals such as recent news or events related to a client and combines that with historical payment behaviour to assess risk. 

“This allows us to prioritise follow-ups and anticipate potential payment issues earlier,” Liu says.

Her agents are built using Claude Cowork, Zapier for workflow automation, Microsoft Outlook 365 for email integration and Xero for financial data. It is less about advanced engineering and more about process design. 

“You need the ability to map workflows clearly — defining triggers, actions and outcomes. Then iteratively refine through testing. Basic systems-thinking is more important than heavy technical development,” Liu notes.

Other agents utilised or being developed by the business include pre-checking quote margins to eliminate underpricing and processing supplier invoices against purchase orders.

Security is managed with restricted access to certain files and datasets, using only the desktop version of Claude, and human approval for pricing and payment decisions.