Following our webinar with the Institute of Singapore Chartered Accountants (ISCA), "The Agentic Finance Function: Navigating the Next Frontier of ERP and AI", we were flooded with excellent, forward-looking questions from finance leaders across industries.
The session showed us that while the shift toward Agentic AI offers great potential for ERP, navigating integration, data governance, and system customisation requires a highly practical roadmap.
To help you chart the path forward, our Senior Director, Alex Ramambason, and Director, Lau Wei Loong, have compiled their detailed answers to the top 8 questions asked by the audience. Read on to learn how to balance foundational ERP structures with next-generation AI capabilities.
Q1: Does having a heavily-customised NetSuite ERP significantly slow down the system for the end-user?
If you have customised your ERP heavily, yes, it will eventually slow down performance, especially for multi-country organisations that layer on unique customisations for every single operating market. That technical debt compounds over time.
The question is now: Where should we host this custom logic? Instead of building complex customisations directly inside your ERP system, you can layer an AI agent outside of it to handle the complexity. Over time, this allows your core ERP engine to simplify back towards a cleaner, standard version while the AI agent smoothly manages the variable business logic.
Q2: How do you ensure corporate data protection while adopting Co-pilot or any Agentic AI?
Governance is key. An AI agent should not be treated any differently from a human user. It must have a named account, an assigned system role, and strict access restrictions.
We look at this in two distinct security layers:
- Authentication: System connectors must be securely authenticated to ensure the agent is explicitly authorised to access the data pipeline.
- Granular Permissions: Once inside the ERP, the agent's permissions must be strictly defined (e.g. read-only versus read-and-write). Within modern ERPs, you have the granularity to decide whether an agent can only read General Ledger accounts or if it possesses the authority to create new purchase orders and invoices.
Q3: Is data cleansing a pre-requisite for AI adoption in an ERP system? For an ERP used for a number of years, can AI help speed up the data cleansing process?
Yes, data cleansing is absolutely a prerequisite. The cleaner your data foundation is, the more accurate your ERP will be, directly improving the quality, reliability, and predictability of your AI's output.
And yes, AI can accelerate this journey. During data migration or cleansing phases, you can deploy AI tools to automate the heavy lifting, such as identifying and inactivating duplicate vendors, fixing inconsistent naming conventions, or flagging customers who haven't been contacted in years. In a recent complex logistics project, we successfully utilised AI tools to query, structure, and enrich legacy data, significantly speeding up the readiness phase through an AI-assisted process.
Q4: How often is your AI/ERP engine updated, and do updates disrupt finance operations? Also, how do you ensure an AI agent is explainable and not a "black box"? If an agent posts an incorrect journal entry, who is accountable?
System Updates: AI features embedded within major ERPs are rolled out through standard release cycles (for instance, NetSuite pushes updates twice a year, in April and September). Even when these updates hit your account, they are subject to user enablement. They do not disrupt operations because you choose when to activate them, and software providers ensure that only broader, non-intrusive features (like generative text buttons) are rolled out automatically.
Explainability & Accountability: The more explicit instructions you give an agent, treating it with the same onboarding rigour as a new human colleague, the more predictable and accurate its outputs become.
Regarding accountability, system design should always mandate a Maker-Checker process. AI agents should never operate in a complete vacuum; they generate the transaction (the Maker), but a human supervisor reviews and approves it (the Checker). Furthermore, because the agent operates under a specific, named user account, the ERP audit trail captures exactly who created and who approved the transaction, maintaining full corporate accountability.
Q5: Have you experienced AI agents going rogue, and how do you prevent this from happening?
We have not seen agents go "rogue" because the business use cases we design and implement for clients to date have been highly specific, restricted, and controlled.
However, what could happen during development is a misalignment in assumptions. For example, when tasking an agent with generating a complex management reporting pack, we noticed some initial variance in the financial figures. Upon review, we realised the agent was making its own assumptions because we hadn't specified the exact parameters. We adapted and refined the prompt instructions to fix this. To prevent unintended actions, we establish strict guardrails during the agent design stage, ensuring for example, that an agent lacks the system deletion permissions required to wipe out historical data.
Q6: Which one should come first - ERP customisation or AI adoption?
Neither should take priority. Think of the ERP as your operational foundation and AI as your business accelerator. AI can only deliver real value if your underlying ERP and data are AI-ready.
Therefore, the true prerequisite is data readiness. Before launching any ERP customisation or AI initiative, leaders must ask: Is our data clean, governed, and easily accessible? If the answer is no, your immediate priority should be a data cleansing and governance exercise.
Once your data is stable, the choice becomes use-case driven:
- For deterministic, repeatable processes, standard ERP configuration or customisation remains the most appropriate choice.
- For non-deterministic, judgment-based processes, AI should be layered on to augment or automate decisions.
Q7: Given the speed of technological evolution, should we wait longer to see how AI matures within the ERP space?
Waiting is an operational risk. When you build AI capabilities today, you are not just buying software; you are establishing an internal capability, an agile way of working, and optimised data pipelines that remain long-term corporate assets.
The rapid evolution of technology shouldn’t pause your roadmap because a successful deployment requires significant groundwork outside of the software itself. Organisations need time to manage change, refine business workflows, and clean legacy data. Starting now ensures your infrastructure and team are mature enough to capture the value of next-generation tools as they arrive.
Q8: How should a Finance team embark on an ERP AI journey? Through self-learning, or in collaboration with the company’s IT department?
Finance should never embark on an AI journey in isolation. It must be a business-led, IT-enabled collaboration.
A purely self-led Finance approach often results in siloed tools and "pilot purgatory", while IT-only initiatives risk delivering technically sound systems that fail to solve real business problems. The ideal path is a co-creation model:
- Finance drives the "What" and "Why": They own the ultimate use cases, such as accelerating the month-end close, anomaly detection, AP automation, or predictive cash flow modelling.
- IT enables the "How" and "At Scale": They own the underlying data architecture, security, access governance, and core integration layers that make the AI outputs reliable, secure, and scalable.
Watch the full webinar recording here!
