AI in Indonesian Banking: What OJK’s AI Governance Guidelines Mean for Your Infrastructure

Written by
Nando Napitupulu
Publshed at
November 18, 2026
Updated at
November 18, 2026
OJK's AI governance guideline for Indonesian banks, explained for IT: where AI can run, audit trails, access control, vendor rules and what examiners ask.

Most Indonesian banks already run AI somewhere: fraud scoring, credit decisioning, customer-service chatbots, anti-money-laundering alerts. What changed in April 2025 is that the regulator wrote down what “doing it properly” looks like. OJK’s Tata Kelola Kecerdasan Artifisial Perbankan Indonesia (AI Governance for Indonesian Banking) sets out how commercial banks should develop, buy, run and monitor AI systems.

The guideline is not a regulation with its own sanctions; OJK describes it as a minimum reference for the sector. But the launch announcement positions it alongside rules banks are already examined against: POJK 11/POJK.03/2022 on IT implementation by commercial banks, SEOJK 29/SEOJK.03/2022 on cyber resilience and security, and the Personal Data Protection Law. Read together, they turn AI principles into concrete expectations about where AI systems run, who can touch them, what gets logged and how vendors are controlled.

Who this article is for

  • Compliance and risk heads: Is the AI we already use compliant, and what will an OJK examiner ask us to show?
  • Heads of IT infrastructure: Where can AI systems run, and what do the racks, network, logs and recovery site need?
  • CIOs: Can we keep using offshore AI APIs, and should AI run in our own data center, a colocation facility or a cloud region?

Short on time? Read the examiner table and the ten questions for your data center provider.

What the Guideline Asks Banks to Do

The guideline is built on three core values that OJK says must be adopted together: reliability (explainable output from secure, resilient systems), accountability (decisions traceable to a responsible person or process, with customer data protected throughout) and human oversight (people stay in the loop, and customers can get human review of AI-driven decisions).

For the technology element, the wording is direct. AI systems “must include robust security features, such as encryption, secure data storage, and continuous monitoring for cyber threats”, and they “must incorporate mechanisms to track and record decisions” (pp. 105–106). Those two sentences are where governance becomes an infrastructure requirement.

An AI Committee, Anchored in Existing IT Governance

The guideline recommends a cross-functional AI committee, accountable to the Board and chaired by the director overseeing IT or risk management, that advises on whether AI projects fit the bank’s strategy, ethics and regulatory obligations. Members include the directors and executive officers for IT, risk, compliance and legal, IT users, and AI specialists where needed.

Its functions can sit within the IT Steering Committee that POJK 11/2022 already requires, and at least one board member should have adequate AI knowledge. For infrastructure teams, AI hosting decisions (which facility, which cloud region, which vendor) now have a formal approval path, and the committee will expect evidence, not assurances.

Where the Guideline Meets Your Infrastructure

1. Where AI systems are allowed to run

The AI guideline itself is neutral on deployment and explicitly allows cloud. The location question is answered elsewhere.

POJK 11/2022, Article 35 requires banks to place their electronic systems in a data center and a disaster recovery center within Indonesia. An electronic system is defined broadly: any set of devices and procedures that collects, processes, analyses, stores or transmits electronic information. A model that scores loans or flags suspicious transactions fits that description as well as a core banking system does.

Placing an electronic system offshore requires OJK approval, and only for specific categories set out in Article 35(3), such as risk management or anti-money-laundering systems integrated with a foreign head office, or systems serving a global customer base. Article 36 adds conditions: a country-risk analysis, a statement that OJK’s supervision will not be weakened, a choice-of-law clause in the vendor contract, and a no-objection letter from the provider’s home supervisor.

The on-shore requirement also covers the recovery side: the OJK requirements for a disaster recovery center apply to AI services that the business can’t run without. Personal-data transfers raise a separate set of questions under Indonesia’s Personal Data Protection Law.

2. Audit trails, logs and the model inventory

The guideline asks for automatic logging, audit trails, human-in-the-loop controls and a model inventory recording every model in use and every change to it. That turns logs into regulated records:

  • Are inference requests, outputs and model versions logged in Indonesia, alongside the systems they describe?
  • Are logs protected from alteration, including by administrators?
  • Can you reconstruct which model version made a given decision six or twelve months later?
  • Do retention periods match what your auditors and the AI committee expect?

SEOJK 29/2022 adds time pressure: banks must notify OJK of a cyber incident, including one involving an AI system, within 24 hours of becoming aware of it and file a full report within five working days, and you can’t notify accurately without logs you can trust.

3. Security and access control

Under SEOJK 29/2022, asset protection covers network, hardware and software, user access controls to prevent unauthorized actions, and adequate protection when working with IT service providers, including in the use of cloud, backed by regular security testing including penetration tests.

AI adds assets that traditional inventories miss: trained model weights, training and fine-tuning datasets, vector databases used for retrieval, and the GPU servers themselves. In practice that means:

  • AI servers and storage in a physically controlled space with logged, individual access, not shared admin credentials
  • AI traffic separated from general corporate networks
  • private links, rather than the public internet, between your systems, cloud platforms and AI providers

A cross-connect inside the data center or a private cloud exchange connection keeps model and data traffic off the open internet, which is easier to defend in an audit than an internet-facing API key.

4. Third parties and AI vendors

Most banks will buy as much AI as they build. The guideline expects risk-based due diligence on a vendor’s information security, financial condition, regulatory compliance and operational resilience, explanations of how vendor algorithms work, testing before and after deployment, and, for generative AI, contingency plans if a third-party model fails. This third-party ecosystem may include AI model providers, cloud platforms, GPU infrastructure providers, colocation operators and network carriers, all of which can play a role in the bank’s AI supply chain. All of it sits on top of the contract terms POJK 11/2022 (Article 30(3)) already requires of any IT service provider: data confidentiality, periodic independent IT audit reports, written approval before subcontracting, critical-incident reporting, and OJK’s right to examine the provider.

5. Resilience and capacity planning

The guideline expects AI infrastructure that can scale and stays available and recoverable. AI workloads the business depends on (a fraud model in the payment path, for example) belong in the same resilience design as the systems they support, including the recovery site. Beyond rack power density, banks should also consider power redundancy, cooling capacity and the ability to scale infrastructure as AI workloads grow. Most bank AI workloads are inference rather than training, and the data center requirements between inference and training differ significantly. Size rack power to the workload you will actually run, and verify what a facility can deliver using a structured assessment of AI-ready colocation. Digital Edge supports high-density AI deployments of up to 20 kVA per rack, subject to the specific hardware configuration, power availability and cooling requirements.

What an OJK Examiner Is Likely to Ask For

This table maps each expectation to the evidence a bank should be able to produce and the team that usually owns it. It is indicative; confirm specific expectations with your OJK supervisor.

ExpectationSourceEvidence to have readyTypical owner
AI systems hosted in IndonesiaPOJK 11/2022 Art. 35–36Data center and DRC location of each AI system, including backups and logs; OJK approval for any offshore placementIT infrastructure, Compliance
Decisions traceableAI guideline (monitoring)Model inventory with version and change history; tamper-protected inference logs; retention policyCIO, AI/data team
Human oversightAI guidelineDocumented human-in-the-loop controls; staff training records; customer route to human reviewBusiness owner, Risk
AI assets protectedSEOJK 29/2022Asset register covering model weights, datasets, vector databases and GPU servers; access logs; network segmentation; penetration test reportsIT security, IT infrastructure
Incidents reported on timeSEOJK 29/2022Incident procedure that covers AI systems and meets the 24-hour initial notificationIT security, Compliance
Vendors controlledPOJK 11/2022 Art. 30(3); AI guidelineDue diligence files; contracts with confidentiality, audit, subcontracting, incident and OJK examination clauses; vendor model test resultsProcurement, Compliance
AI decisions governedAI guideline; IT Steering Committee (POJK 11/2022)Committee charter; minutes approving AI projects and hosting choicesCIO, Board
AI services recoverableAI guideline; DRC requirementsAI systems in the business continuity plan; DR test results for critical modelsIT infrastructure

Choosing Where Your AI Runs

Can we keep using offshore AI APIs?

Many generative AI pilots started by calling a model hosted outside Indonesia. Whether that can continue depends on what the API receives:

  1. No customer or confidential bank data (for example, drafting public content): placement risk is lower, but vendor due diligence and personal-data rules still apply.
  2. Customer or confidential data reaches the API: the service is likely processing data as part of the bank’s electronic system. Offshore placement then needs OJK approval, which Article 35(3) of POJK 11/POJK.03/2022 (OJK Regulation on IT Implementation by Commercial Banks) limits to specific categories that many customer-facing uses may not fit.
  3. If approval isn’t available, there are three on-shore routes: use model endpoints hosted in a Jakarta cloud region, self-host an open-weight model on your own GPUs in an Indonesian data center, or combine the two over private connections.

Before choosing, check that the specific models you need are offered in-region (availability differs by provider and changes often) and that prompts, logs and backups also stay in Indonesia. Confirm individual cases with your OJK supervisor.

Own data center, colocation or cloud region?

AWS, Google Cloud and Microsoft Azure all operate cloud regions in Jakarta, so on-shore no longer means building your own facility. The trade-offs:

Own data centerColocation in IndonesiaJakarta cloud regionOffshore AI service
Meets on-shore placementYesYesYes, if logs, backups and model endpoints stay in-regionNo, unless OJK approves
Cost profileCapital-heavy; GPU density may need power and cooling upgradesYou own the hardware; predictable monthly cost suits steady inferencePay per use; flexible for pilots, can get costly at constant high useLowest entry cost
Control of logs and physical accessFullHigh: your racks, provider access logs, OJK audit rights in the contractShared; relies on provider audit reportsLowest
Time to deploySlow if the facility needs upgradesWeeks to months, mostly hardware lead timeFastFastest
Best fitBanks with modern facilities and spare powerSteady production AI; hybrid with cloudPilots and variable workloadsNon-sensitive use, or approved categories

Many banks will end up hybrid: steady, sensitive inference on owned GPUs in colocation and experimental work in a Jakarta cloud region. Both environments can be connected through private connectivity to support secure data exchange.

Questions to Ask a Data Center Provider

Each question maps to an expectation above.

  1. Location: Are the facility and every component of my deployment, including backups and logs, physically in Indonesia?
  2. Access control: How is physical access to my racks granted, logged and reviewed? Can you provide access logs for an audit period on request?
  3. Audit rights: Will the contract include OJK’s right to examine your facility and services, as POJK 11/2022 requires?
  4. Independent assurance: Which certifications and independent audit reports can you share, and how often are they renewed?
  5. Subcontracting: Do you subcontract any part of the service, and will you seek our written approval before doing so?
  6. Incident reporting: How fast will you notify us of a security or operational incident, given our 24-hour notification duty to OJK?
  7. Private connectivity: Can we connect privately to our cloud platforms, AI providers and carriers within the facility, without traversing the public internet?
  8. Power and density: What rack power can you deliver today for our AI hardware, and what evidence supports it?
  9. Operations support: Who handles on-site tasks such as hardware swaps and cabling for our AI servers outside office hours, and under what remote hands terms?
  10. Recovery: Can our recovery site sit within the same provider’s footprint, with private replication links between sites?

Conclusion

OJK’s AI governance guideline doesn’t ban any technology or mandate a particular architecture. It makes AI accountable the way the rest of a bank’s IT already is. Combined with POJK 11/2022’s on-shore placement rule and SEOJK 29/2022’s cyber-resilience duties, the general expectation is clear: AI systems that touch Indonesian customer data should be hosted in Indonesian facilities, supported by secure connectivity, reliable audit trails and access controls that banks can demonstrate to OJK and Bank Indonesia examiners.

Planning where your bank’s AI systems will run? Talk to our team about colocation for regulated AI workloads in Jakarta, with high-density racks of up to 20 kVA, private cloud connections and on-site remote hands. Get in touch to arrange a site review.

Nando Napitupulu
FSI Account Manager

Talk to Digital Edge Indonesia Experts

Complete the form below to discuss about the modern digital infrastructure with our dedicated experts.
This site uses cookies
Select which cookies to opt-in to via the checkboxes below; our website uses cookies to examine site traffic and user activity while on our site, for marketing, and to provide social media functionality.