Medical claims inflation in Malaysia ran at 15% in both 2024 and 2025, well above regional and global averages, according to RAM Ratings in July 2026. Bank Negara Malaysia’s Annual Report 2024 put the longer trend in two numbers: between 2021 and 2023 the total cost of medical and health insurance and takaful claims rose 73%, while premiums rose 21%.
Motor is no easier. The General Insurance Association of Malaysia (PIAM) reported an underwriting loss of RM289.3 million on motor business in 2025, a combined ratio of 103%, and an average cost per claim up about 20% to RM8,831.
The usual answer, raising prices, is restricted. Bank Negara’s interim measures spread medical premium increases made between 2024 and 2026 over at least three years, and the Governor has told Parliament’s Public Accounts Committee that repricing to widen profit margins is strictly prohibited. A base medical plan is due nationwide in early 2027, and RAM says it may bring thin margins at first.
That leaves the thing an insurer or takaful operator does control: what it costs to handle a claim, and how accurately it is handled.
Adoption is wide and shallow
Bank Negara’s 2024 survey of the sector, cited in its August 2025 discussion paper on AI, found that 77% of insurers and takaful operators had at least one AI application in use, up from 58% a year earlier. Most of it sits in customer analytics, internal operations and fraud detection, not in core risk functions.
A separate report commissioned by the FIDE Forum and developed with Accenture Malaysia added the uncomfortable part: no Malaysian insurer had scaled its AI strategy. One application in use is a pilot. It is not a claims operation that runs differently.
A claim is a file of documents
A hospital claim arrives as an admission form, an itemised bill, a discharge summary and a doctor’s report. A motor claim arrives as a police report, photographs, an adjuster’s report, a workshop estimate and invoices. Somebody reads each one and types what matters into the claims system. Somebody else checks it against the policy.
This is where AI is reliable today, in three jobs.
1. Reading the file into a record. Language models are good at pulling fields out of untidy documents: dates of admission and discharge, diagnosis, line items and amounts, vehicle and workshop details. We use this in production for a different kind of financial document. In our banking litigation platform, a model reads figures out of bank statements of account, including scanned ones, and a person confirms them before anything is calculated. The rule transfers directly: the model fills a strict schema, and a claims officer confirms.
2. Checking the file against itself and against references. Is the admission date inside the period of cover? Does the bill include items the plan excludes? Has this invoice been submitted before? Once the fields exist, these are software checks. The references are improving too. Price ranges for 26 common medical procedures have been published and a central medical claims database has been set up. On the motor side, one loss adjuster told the New Straits Times that disputes over inflated repair costs have dropped sharply since spare part prices and labour hours were standardised in a centralised industry database.
3. Sorting the queue. A clean, low-value claim with complete documents should not wait behind a complicated one. Triage sends simple claims down a fast path and puts the unusual ones in front of an experienced adjuster sooner. Speed is now a regulatory expectation as well. Bank Negara’s revised motor claims policy, described in August 2025, is meant to cut average processing time for own damage claims by 20 working days. RAM expects a pilot of online police reporting and end-to-end digital claims processing to improve efficiency further.
What AI should not do
Decide to reject a claim. A model can prepare the file and point at what looks wrong. A named person decides, and the reason is recorded. Malaysia’s data protection regulator points the same way. The guidelines on automated decision-making and profiling issued on 30 April 2026 say AI should not be the sole basis for a decision about a person, and expect explicit consent where the processing involves sensitive data such as health information.
Do the arithmetic. Deductibles, co-payments, annual limits and benefit schedules are rules. The standard tier of the proposed base medical plan alone carries a RM500 deductible per disability for policyholders under 61, RM1,000 for those aged 61 and above, and a 20% co-payment at out-of-network hospitals capped at RM3,000. That belongs in tested code that gives the same answer every time, not in a prompt.
Accuse anyone of fraud. In Adyen’s 2026 report on Malaysian insurance, 96% of insurers estimated that up to 24% of claims may involve some form of fraud, and 71% said fraudsters’ own use of AI has made attacks harder to prevent. The same loss adjuster was direct about it: “AI makes evidence far easier to tamper with, including manipulated scene photographs and falsified timestamps.” His firm still treats field checks as its most important tool. So a claims process that approves on photographs alone is weaker than it was two years ago. Use models to flag patterns for an investigator. Do not let them clear or condemn a claim. We cover the wider problem in deepfake scams and Malaysian businesses.
What Bank Negara expects
The discussion paper of 5 August 2025 is not a regulation, but it reads like a specification. The existing technology risk rules are technology-agnostic and are treated as broadly adequate for current uses of AI. Institutions are expected to embed fairness, accountability and transparency. The paper notes that most AI applications in the sector are designed to augment human decisions, not replace them, and it warns about the whole sector leaning on the same models and datasets.
For a claims system, that becomes four questions a reviewer will ask. Who accepted this output? What did the model see, and which version was it? Can the provider be changed without rebuilding the process? Where did the policyholder’s medical data go? We set out the detail in what the paper means if you build for banks. It applies equally to insurers, takaful operators, third-party administrators and their vendors.
A note on takaful
General takaful’s market share rose to 22% in 2025, according to RAM. The documents in a takaful claim are the same as in a conventional one. The ledger behind them is not. Contributions, funds and surplus follow their own rules, and those rules must be visible to the operator’s Shariah and actuarial teams.
That is another reason to keep calculation out of the model. Our litigation platform computes conventional and Islamic facilities with separate, deterministic logic for the same reason. A reviewer can read code. Nobody can audit a paragraph of instructions to a model.
The front desk
Agents and customers already live on WhatsApp, and insurers are meeting them there. In February 2026 Zurich Malaysia described an AI assistant for its agents that works inside WhatsApp in English, Malay, Chinese and Tamil, can take a motor renewal from eligibility check to cover note in one conversation, and passes anything needing nuance to a live support team. The handover is the part worth copying. See AI agents on WhatsApp for how to build one that knows when to stop.
What this means if you run claims
- Start with reading, not deciding. Pick one document type, such as itemised hospital bills or workshop estimates, and measure how many fields a person has to correct.
- Use AI to read. Use code to calculate. Use people to approve. Every payment and every rejection has a named owner.
- Keep the trail. Store the input, the extracted record, the model version and who confirmed it. You will be asked.
- Treat images as claims, not proof. Verify at the source where you can: the hospital, the workshop, the police report.
- Keep the model replaceable. Put it behind an interface, and know which data leaves your environment and where it goes.
Prices are constrained and claims are not. The insurers and operators that do well over the next few years will be the ones that handle each file faster and more accurately than they do today, with people still making the decisions.
Sources
- Insurance sector to hold firm in 2H (The Star)
- BNM: medicine and procedure price display key to addressing medical inflation (Malay Mail)
- Interim measures on medical and health insurance and takaful (Bank Negara Malaysia)
- Governor: Bank Negara mandated to protect insurance industry (CodeBlue)
- PIAM: motor insurance posts RM289.3 mil loss in 2025 (paultan.org)
- Discussion paper on artificial intelligence in the Malaysian financial sector (Bank Negara Malaysia)
- Malaysia’s AI financial regulation framework (Tech Wire Asia)
- Bank Negara Malaysia’s discussion paper on AI in the Malaysian financial sector (Halim Hong & Quek)
- Report on AI adoption in Malaysia’s financial institutions, FIDE Forum and Accenture (The Edge Malaysia)
- Crash investigators check 800 motor cases yearly (New Straits Times)
- Motorists can claim from own insurer without losing NCD, says Bank Negara (The Star)
- Malaysia Insurance Report 2026 (Adyen)
- PDP Commissioner launches guidelines on data protection impact assessment, automated decision-making and profiling, and data protection by design (Allen & Gledhill)
- Zurich Malaysia press release, 11 February 2026
This article is general information and is not legal advice.