AI Automation in Shared Services in Malaysia: What Changes Beyond RPA

The short answer: RPA automated the structured, rule-following half of shared services work. AI automation in Malaysian shared services is about the other half: scanned invoices that do not match the template, vendor documents in three languages, free-text HR queries and the exceptions RPA always handed back to a person. That is where the value is, and it is precisely where auditability and human review have to be designed into the workflow from day one rather than added after the pilot.

Why Is AI Automation in Shared Services a Malaysian Question Now?

Because the sector is large, growing and already being asked to move up the value chain. The Ministry of Digital’s press release on the GBS Malaysia Mid-Term Report 2025 (30 October 2025) puts the number of GBS companies at 749, up 66.8% since 2022, with investment rising from RM0.73 billion in 2021 to RM9.87 billion in 2024, over 36,000 high-value jobs generated, and projected GBS revenue of RM28.14 billion in 2025. The summit where that report launched was themed “GBS 5.0: Human-Centric, AI-Driven, Impact-Focused”, which tells you where the policy conversation has landed.

The incentive framework points the same way. MIDA describes its Global Services Hub scheme, introduced in Budget 2024 as the successor to the Principal Hub incentive, around higher value-added regional and global functions, and notes that modern hubs increasingly adopt AI and RPA. And the industry is saying it out loud: around a GBS summit in September 2026, an AstraZeneca finance leader described finance teams moving from talking about AI to deploying it, while a Shell transformation lead stressed keeping human judgement at the centre, as Digital News Asia reported.

The practical question is narrower. You almost certainly have RPA already. What does an LLM or an agent do that your bots do not, and which process do you trust it with first?

What Did RPA Leave Behind in the Back Office?

RPA was a good answer to a specific problem: a human copying values between systems that do not talk to each other. Given a stable screen and a structured input, a bot is fast and perfectly repeatable. Its weakness is also its definition. It follows rules over fields, so the moment an input stops being structured, it stops.

In any finance or HR shared services operation, that leaves a recognisable residue. The invoice that arrives as a phone photo. The supplier whose remittance advice is a paragraph of text in an email. The bank statement line that matches nothing because the reference was typed by hand. The onboarding pack where the company registration certificate is in Bahasa Malaysia and the bank letter in Chinese. RPA routes all of these to an exceptions queue, and the exceptions queue is where your experienced people spend their day.

That residue is not small. In an August 2025 interview with Bernama, AGOS Asia’s head of digital, citing global surveys, estimated that about 40% of roles in Malaysian GBS involve highly standardised, repetitive tasks such as invoice processing, reconciliation and employee onboarding, already partially automated depending on organisational maturity. “Partially” is the operative word. The part that is not automated is overwhelmingly the unstructured part.

How Do RPA, LLM-Based Automation and Agents Compare?

These are not generations where the newest replaces the rest. They are three tools with different input tolerances and different failure modes, and a mature operation runs all three, often inside one process.

The row that matters most is failure mode. RPA fails loudly: a screen changes, the bot stops, someone notices. An LLM fails quietly: it returns a confident, well-formatted answer that happens to be wrong. An agent can fail by compounding, because a misclassification in step one becomes the premise for steps two to six. Every control you design should follow from which of those three you are dealing with. If the agent distinction is new to you, agentic AI explained covers what separates an agent from a well-prompted model, and the multi-agent systems guide covers when splitting the work across agents helps and when it just adds surface area.

Where Is the Value, and Why Is It Also Where the Risk Is?

Here is the uncomfortable symmetry at the centre of this. The work RPA could not do is valuable precisely because it needs judgement: reading an ambiguous document, deciding which of two plausible matches is right, noticing that an invoice looks off. The same property makes it the work where a wrong answer is expensive and hard to spot.

So the review step is not a concession to nervous auditors. It is part of the product. When I built the LPA Analyzer, which takes 150 to 200 page Limited Partnership Agreements and extracts every instance of six critical clause categories, each extraction carried a confidence score and a link back to its position in the document. That was not decoration. It is what let a lawyer check the output in seconds instead of rereading the agreement, and it is the only reason the time saving was real. SureCiteAI follows the same principle: every answer is attributed to its source, so any claim can be verified against the original document.

Translated to shared services, designing the review step means deciding four things before you build anything:

• What the reviewer sees. Not just the proposed answer, but the source span it came from, the confidence and the reason it was routed to them. A reviewer shown only a conclusion will either rubber-stamp it or redo the work.
• What routes to review and what does not. Thresholds by confidence, by amount, by vendor risk and by action type. Reading a document can be automatic; changing bank details should never be.
• What gets logged. The input, the retrieved context, the prompt and model version, the output and the human decision. Your internal and external auditors will ask why a payment was released; “the model said so” is not an answer, and a reconstructable trail is.
• How corrections flow back. Every reviewer override is labelled data. Capture it and it becomes your evaluation set; ignore it and you will never know whether accuracy is improving.

This is the core of how I approach AI automation with humans in the loop: the automation and the review queue are designed as one system, with the same care, rather than a model with an approval button attached afterwards.

What Changes When Documents Arrive in Three Languages?

Regional operations run out of Malaysia routinely handle English, Bahasa Malaysia and Chinese in the same week, sometimes on the same invoice. MIDA’s own workforce overview notes that the typical Malaysian speaks at least two or three languages, which is part of why regional work lands here. It is also why template-based OCR and RPA struggle: every language and layout combination is another template to maintain.

Multimodal models read mixed-language documents far better than template systems, one of the clearest wins for intelligent document processing in Malaysia. Three cautions. First, do not assume equal accuracy across languages; build your evaluation set with real documents in each language you receive, in proportion. Second, normalise entities such as company names, which can appear in romanised, Malay and Chinese forms for the same supplier, before matching against the vendor master. Third, keep the original-language source visible to the reviewer. A translated summary is convenient; it is not evidence.

Which Shared Services Process Should Go First?

The best first process is not the one with the most hype attached. It is the one where you can measure success cleanly and where a mistake is caught before it costs anything. My shortlist, in rough order:

1. AP invoice exceptions. High volume, a clear baseline (exceptions per week, time to clear), a ground truth you already have in posted invoices, and a reviewer who already exists. The AI proposes; the AP clerk approves.
2. Vendor onboarding document checks. Multilingual documents, a checklist that defines “correct”, and a natural approval gate before the vendor record goes live. A strong fit for fraud controls too, provided bank-detail changes always need a human.
3. Reconciliation matching suggestions. Fuzzy matching on free-text references is a genuine LLM strength, and proposed matches are easy to verify. Keep journal postings behind review.
4. HR policy queries. Good value, but only with answers grounded in the current policy documents and cited, and with anything touching an individual’s pay, leave balance or contract routed to a person.
5. Contract review. High value per document, lower volume. Extraction and deviation flags work well; the judgement stays with legal.
6. Collections. Last, not because the technology cannot draft a reminder, but because it is customer-facing, tone-sensitive and a poor place to learn your controls.

Whichever you pick, baseline it before you start. Knowing today’s exception rate and handling time is what turns a pilot into a business case, and the method for doing that honestly is in AI automation ROI for enterprises. If you are weighing a platform vendor against a custom build for the chosen process, build vs buy AI sets out the trade-offs, and for the broader landscape of where generative AI is landing across Malaysian businesses, see generative AI use cases in Malaysia.

How Does PDPA Shape the Design?

Shared services data is personal data: employee records, payroll queries, vendor contact names, bank letters, identity documents in onboarding packs. The 2024 amendments to the Personal Data Protection Act, phased in across 2025, raised the stakes. As Mayer Brown’s July 2025 summary sets out, data processors are now directly subject to the security principle, both controllers and processors must appoint a data protection officer, breaches must be notified to the Commissioner as soon as practicable, the old cross-border whitelist has given way to a risk-based transfer framework, and the maximum fine for breaching the data protection principles rose from RM300,000 to RM1,000,000.

For a GBS centre processing on behalf of group entities, the processor point is the one people miss. It is no longer only the controller’s problem. In architecture terms, three questions need answers before go-live: where personal data ends up (prompts, logs, caches and vector indexes, not only the source system); where your model provider processes it, which is now a cross-border transfer assessment; and whether you can find and act on one person’s data across all of those places. (If staff are also using ChatGPT or Copilot directly, the same transfer questions apply, covered in rolling out ChatGPT Enterprise or Copilot under PDPA.) The mechanics of that last one are covered in right to erasure in vector databases.

What Happens to the Shared Services Workforce?

Honestly, roles change shape more than they disappear, and unevenly. Pure data entry shrinks. Exception handling, review and control work grow in importance, and they need people who understand the process well enough to spot a wrong answer delivered confidently. That is a skills shift, not a relabelling, and it needs investment.

Two things I would tell any head of shared services. First, your most experienced exception handlers are your most valuable asset in an AI programme, because they are the ground truth. Involve them in defining what correct looks like and they will make the system better; present the system to them as a fait accompli and they will be right to distrust it. Second, keep headcount promises out of the business case. They distort scoping and poison adoption. Measure throughput, exception rates, cycle time and accuracy, and let the operating model follow the evidence.

Frequently Asked Questions

What does AI automation change in shared services that RPA could not?

RPA follows fixed rules over structured screens and fields, so it stops at the first document or email that does not match the template. LLM-based automation can read unstructured inputs such as scanned invoices, free-text emails and contracts, and agents can work through a multi-step exception by looking things up and proposing a resolution. The gain sits in exactly the work RPA handed back to people, which is also where review and audit design matter most.

Which shared services process should we automate with AI first?

Start with a high-volume process where the input is messy but the correct outcome is checkable, such as accounts payable invoice exceptions or vendor onboarding document checks. Those give you a clear baseline, an obvious reviewer and a ground truth to measure against. Leave collections outreach and anything that changes a payment or a person’s employment record without a human decision until the review step has proven itself.

Does AI automation in a Malaysian GBS centre raise PDPA issues?

Yes, because invoices, onboarding packs and HR queries are full of personal data about employees and vendor contacts. Since the 2024 amendments, which phased in during 2025, data processors are directly subject to the PDPA security principle, controllers and processors must appoint a data protection officer, and breaches must be notified to the Commissioner. Design for that at ingestion: know where personal data lands in prompts, logs and indexes, and where your model provider processes it.

Automate the Exceptions, Design the Review

Malaysia’s GBS sector has the scale, the multilingual talent and the policy tailwind to move from cost arbitrage to something more valuable. The operations that get there will not be the ones that deploy the most agents. They will be the ones that picked a measurable process, treated the reviewer’s screen as seriously as the model, and could show an auditor exactly why every decision was made.

If you are choosing that first process, I work with shared services and operations leaders on exactly that scoping, and on building the system once it is scoped; see AI development in Malaysia for how engagements run, or book a consultation to talk through your exceptions queue.