Back to Insights

Insight · Guide

16 January 2027: what changes for AI and data in your company.

PP 33/2026 turns Indonesia's data protection law from principles into a work list. This guide puts it in operating terms: where AI comes in, who owns each obligation, what employees may paste into ChatGPT, and a nine-week plan for a baseline you can defend.

Akses team Updated 16 min read 22 sources
Contents
  1. ·Summary
  2. 01What changes
  3. 02Where AI comes in
  4. 03The AI already in use
  5. 04Who owns what
  6. 0510-point checklist
  7. 06Employee AI policy
  8. 07A nine-week plan
  9. 08What's still open
  10. 09Common questions
  11. 10Closing
  12. RReferences
Summary
  • PP 33/2026 takes effect on 16 January 2027. Its 225 articles turn the PDP Law from principles into work that can be checked: records of processing, DPIAs, DPOs, vendor contracts, and breach notice within 3×24 hours1.
  • AI is named explicitly. The elucidation of Article 120 lists AI and machine learning as "new technology". Using AI to process personal data therefore needs a DPIA before it starts, whatever the scale1.
  • The risk is already on employees' desks. 76% of workers in Indonesia bring their own AI tools to work11. One paste of customer data into a personal AI account can touch four obligations at once.
  • The maximum fine is 2% of annual revenue, but that is a ceiling. The amount depends on impact, cooperation, and compliance history, and it can be zero2. Documented compliance work lowers the risk.
  • Nine weeks is enough for a defensible baseline. Start with an inventory and owners, then an employee AI policy, DPIAs for the riskiest AI uses, vendor contracts, and a breach drill.
16 JanPP 33/2026 takes effect on 16 January 2027, six months after promulgationArticle 225
2%Maximum fine as a share of annual revenue. It can also be set at zeroArticle 185
3×24Hours to notify data subjects and the agency after a breachArticle 114
76%Workers in Indonesia who bring their own AI tools to workMicrosoft Work Trend Index, 2024

01What changes

From principles to a work list.

Indonesia's Personal Data Protection Law (UU PDP No. 27/2022) has been fully in force since its adjustment period ended in October 20241. The trouble is that the law speaks mostly in principles: data must be protected, processing needs a basis, breaches must be reported. For most companies, principles like these are easy to agree with and hard to check.

PP 33/2026 changes that. The regulation was enacted and promulgated on 16 July 2026, has 225 articles in 12 chapters, and takes effect six months later, on 16 January 202756. It does not add new principles. It sets out how you prove you follow them: which documents must exist, who signs them, and how much time you have when something goes wrong.

For a board, three changes matter most.

  • Compliance is now evidence. Records of processing with 13 minimum fields (Article 74), a written retention policy (Article 75), documented DPIAs (Articles 120 to 122), and internal and external PDP audits (Article 138)12.
  • Sanctions need no prior warning. A written warning, suspension of processing, deletion of data, and a fine can be imposed together (Article 184)1.
  • Supervision results can be published (Article 183). For many businesses, the reputational cost is higher than the fine1.

The fine is often quoted without context. Article 185 does cap it at 2% of annual revenue or receipts. But the amount is set from a list of factors: the impact and duration of the violation, the type of data, the number of data subjects, how the violation was found, cooperation during examination, business scale, ability to pay, and compliance history. The fine can even be set at zero23. The last two factors are entirely in your hands.

PDP compliance is no longer about good intentions. It is documents you can show.

02Where AI comes in

Four doors where AI meets PP 33.

PP 33/2026 is not an AI regulation. Indonesia has no binding AI rules yet: two draft presidential regulations on an AI roadmap and AI ethics had been initialled by every relevant minister by September 2026, but still await the President's signature20. What exists is a 2023 circular from the communications ministry (SE Menkominfo No. 9/2023) on AI ethics, which is guidance only19.

But as soon as AI processes personal data, it enters PP 33 territory. There are four doors.

Figure 1
Four places where AI use meets PP 33/2026.
Articles 120 to 122DPIAThe elucidation of Article 120(2)(f) names AI and machine learning as new technology. The DPIA must be done before processing starts.
Articles 93 to 96Automated decisionsData subjects can object to fully automated decisions, including profiling, with legal or significant effect. There must be a route to human review.
Articles 14 to 15AI vendorsAn AI vendor that processes data on your behalf is a processor. A written agreement is required, and you must approve its sub-processors.
Articles 165 to 173Transfers abroadMost AI services run on servers outside Indonesia. The legal basis for transfer is tiered, and consent only works for occasional transfers.

Summarised from RSM Indonesia, PPC Land, and Veritask. Article numbers from secondary sources that agree with each other.

Door 1DPIA: AI is named

Article 120(2) lists seven kinds of high-risk processing that require a DPIA (data protection impact assessment) first: automated decisions with legal or significant effect, specific (sensitive) data, large scale, systematic evaluation or monitoring, matching or combining datasets, new technology, and processing that restricts data subject rights12. According to RSM, the elucidation of point (f) names artificial intelligence, machine learning, smart technology, and IoT as examples of new technology1.

The consequence is simple: a customer service chatbot, automatic summaries of medical check-up results, customer risk scores, or an AI assistant that reads client email all need a DPIA before use. A DPIA is done per system or use case, not per prompt. At minimum it covers a description and purpose of the processing, why it is necessary, the risks to data subjects, and mitigation, with the DPO's input recorded3.

Door 2Automated decisions: a human must be able to step in

Articles 93 to 96 give data subjects the right to object to decisions made solely by automated processing, including profiling, when the decision has legal or significant effect. Examples in the elucidation: credit refusal, job application refusal, and differential pricing2. If the objection is accepted, the company must provide human review (Article 94).

This matches SE Menkominfo 9/2023: AI should not be the decision-maker on matters concerning people19. In practice, the cheapest way to comply is to make sure every important AI-assisted decision has one human name that approves it.

Doors 3 and 4AI vendors and servers abroad

When you use AI through an API or a business subscription to process customer data, the vendor is almost always acting as a processor. Article 14 requires a written agreement with minimum terms (scope, purpose, data types, duration, audit rights), and Article 15 requires your written approval for sub-processors13.

For transfers abroad, Article 165 sets a tiered order, not a free choice: first, a destination country with equal or higher protection; failing that, adequate and binding safeguards; and only last, the data subject's consent1. The agency's list of equivalent countries and its standard contract clauses have not been issued yet. Consent is also limited to non-repetitive transfers involving a limited number of people (Article 173)2. So for routine AI use, the realistic route today is a binding contract with the vendor.

AI has no rules of its own in Indonesia. Once it touches personal data, it follows the PDP rules.

03The AI already in use

The question isn't "are we using AI".

Many boards open this discussion by asking whether it is time for the company to use AI. The data says that question has already been answered, often without them knowing.

In 2024, Microsoft found that 92% of knowledge workers in Indonesia use generative AI at work, and 76% bring their own AI tools11. The 2026 edition shows 33% of Indonesian workers are already among the most advanced AI users, twice the global average, while only 42% feel their leaders have a clear position on AI12. PwC found 69% of Indonesian workers used generative AI in the past year13.

Data from security vendors, which should be read with care because they also sell the fix, shows what goes in with it. Cyberhaven found 34.8% of company data that employees put into AI tools is sensitive, up from 10.7% two years earlier16. Netskope counts an average of 223 generative AI data policy violations per organisation per month, 32% of them involving regulated data15.

ExampleOne paste, four obligations

Picture an HR staff member asked to summarise the annual medical check-up results for a board meeting. They export a spreadsheet with the names, national ID numbers, and results of 800 employees, and paste it into their personal ChatGPT account so the summary is done in five minutes. The intent is good. The summary may well be good too.

Figure 2
One paste into a personal AI account can touch four obligations at once.
What happened

A spreadsheet of check-up results for 800 employees, pasted into a personal AI account.

  • No agreement with the vendor
  • Servers outside Indonesia
  • Recorded nowhere

5 minutes · a good summary

What it touched
  • Specific data. Health data is listed in Article 6(1)(a)
  • A processor with no contract. Article 14
  • A transfer with no basis. Article 165
  • Possibly a protection failure. The 3×24 hour clock may start, Article 114

4 obligations · nobody knows

An illustration, not an event at any particular company. Whether an incident counts as a reportable protection failure is a legal judgement made case by case.

Many companies react first with a ban. Samsung did in 2023, after engineers uploaded internal code to ChatGPT17. But BCG found 54% of workers would use AI even if it weren't authorised14. A ban moves the use to personal phones, where you can't see anything.

The more promising move is the opposite: provide an approved tool. Netskope's data shows that as use of company-managed AI accounts rose from 25% to 62%, use of personal accounts fell from 78% to 47%15. That is correlation, not proof of cause. But the direction makes sense: people take the easiest path. Your job is to make the safe path the easiest one.

A ban doesn't stop AI use. It only makes it invisible.

04Who owns what

Every obligation needs one name.

In many companies, PDP is seen as legal's job, AI as IT's job, and breaches as a "later" problem. The result is that nobody holds anything. PP 33 does define a DPO, but the DPO doesn't carry out every obligation. The DPO advises, monitors, and acts as the contact point (Articles 144 to 146)1. The work itself stays in the business units.

The split below is a starting point that usually works in mid-sized companies. Adapt it to your structure, but don't leave a single row empty.

Template 1 · Owner map
ObligationOwnerEvidence that must existArticles
DPO appointmentBoardAppointment letter, direct access to the board142 to 146
Records of processingCOO, filled in by each unitOne list, AI tools included, updated quarterly74 to 75
DPIA for AISystem owner, reviewed by the DPOA DPIA dated before go-live120 to 122
Vendor and AI contractsProcurement and legalA processing agreement per vendor, sub-processor list14 to 15
Transfers abroadIT, reviewed by legalA data flow map and transfer basis per vendor161 to 173
Employee AI policyCOO and HRA one-page policy, signed acknowledgementsGood practice
Breach responseIT, with the DPORunbook, notice templates, drill records114 to 118

The column most often forgotten: Evidence. If you can't show it within one working day, assume it doesn't exist.

A note on the DPO. Article 142 requires one for public service processing, large-scale systematic monitoring, and large-scale processing of specific data1. Since a Constitutional Court ruling in July 2025, meeting any one condition is enough, not all of them9. Clinics, labs, medical check-up providers, and companies that hold large volumes of employee health data should assume they need a DPO.

05Checklist

A 10-point checklist for the board.

Answer honestly. "Yes" means the evidence can be shown today, not "in progress".

0 of 10 answered yes

#PointQuestionYes / Not yetIf not yet, start here
1Owners Does every PDP obligation have one named owner, and has a DPO been appointed if the conditions apply?
Yes or not yet, point 1
Fill in the owner map above in one board meeting, then decide whether a DPO is required and who it is
2Records Is there one list of all personal data processing: what data, for what, stored where, and shared with whom?
Yes or not yet, point 2
Ask each unit head to fill in one row per process, starting with the five processes that hold the most customer data
3AI inventory Do you know every AI tool employees use, including AI features inside software and automatic meeting notes?
Yes or not yet, point 3
Send an anonymous three-question survey: which tools, for what work, and what data goes in
4DPIA for AI Does every AI use that touches personal data have a DPIA completed before it went live?
Yes or not yet, point 4
Pick the riskiest AI use from the inventory and write the first DPIA as your template
5Automated decisions Where AI helps decide something about people (credit, hiring, triage, pricing), is there a human who approves and a way to object?
Yes or not yet, point 5
Name an approver for each AI-assisted decision, and one address where people can object
6Vendor contracts Does every vendor that processes personal data for you, including AI and cloud vendors, have a written processing agreement?
Yes or not yet, point 6
Rank vendors by volume of personal data, then request a DPA from the top ten
7Transfers abroad Do you know which data leaves Indonesia, to which country, and on what basis?
Yes or not yet, point 7
Add "server location" and "transfer basis" columns to the vendor list and fill them in with IT
8AI policy Do employees know which AI tools are approved and which data must never go in, and have they signed it?
Yes or not yet, point 8
Publish a one-page policy with the traffic light rule in section 06, and provide one approved AI tool
93×24 hour breach If a breach is found on Friday afternoon, does everyone know who decides and who sends the notice before Monday afternoon?
Yes or not yet, point 9
Write a one-page runbook and notice template, then drill it with one "wrong paste into AI" scenario
10Retention and rights Is there a policy on how long data is kept, and a way for people to ask for their data to be accessed or deleted?
Yes or not yet, point 10
Set retention periods for your three largest data types, and one email address for data subject requests

The order is deliberate. Points 1 to 3 are the foundation: without owners and a list, the other points have nothing to stand on. If time is short, work on the first "not yet" from the top.

06Employee AI policy

A traffic light, not a ban.

AI policies that work are usually short. Employees don't read a 20-page document, but they can remember three colours. The core is one question before pasting anything: what colour is this data?

Figure 3
One question before pasting: what colour is this data?
Green · goPublic informationGeneral email drafts, published material, public research, polishing your own non-confidential writing. Fine in approved AI tools.
Amber · company account onlyInternal, no personal dataInternal SOPs, anonymised figures, notes with names removed. Only in a company AI account where data isn't used to train models.
Red · don'tPersonal and confidential dataNames with ID numbers, phone numbers, addresses, health data, personal financial data, passwords and API keys, source code, contracts, unreleased financials. Only through a system that has had a DPIA.

Built from common AI use policy practice. The red category follows specific data in Article 4 of the PDP Law and the data types most often leaked according to Netskope and Cyberhaven.

TemplateA one-page AI policy

  1. Approved tools

    Name them, with company accounts and company sign-in. Personal accounts aren't used for work data. One route to request a new tool.

  2. Three colours

    The traffic light rule above, with examples from everyday work in your own company.

  3. Settings

    No model training on your data, retention as short as possible. Note: "temporary chat" on a personal ChatGPT account can still be kept for up to 30 days18, so it is no substitute for a business account.

  4. Output

    A human reviews before use. AI doesn't decide important things about people. Numbers, quotes, and legal citations are checked.

  5. Wrong paste

    Report within one hour, no penalty for prompt reports. The DPO assesses whether the 3×24 hour clock has started.

  6. Sign-off

    One short training session and a signature, repeated every year. Policy owner: the COO, reviewed every 6 months.

Employees don't read a 20-page policy. They can remember three colours.

07Plan

Nine weeks, five steps.

This plan works whenever you start, before or after 16 January 2027. Nine weeks isn't enough for a perfect compliance programme. It is enough for a baseline you can defend if someone asks one day. If the rules are already in force, don't wait: every week without documents is a week without compliance history.

  1. Weeks 1–2

    Owners and inventory. Fill in the owner map, decide on the DPO, send the AI tools survey, and ask each unit to fill in records of processing for its five biggest processes.

  2. Weeks 3–4

    AI policy and an approved tool. Publish the one-page policy, switch on one AI tool with company accounts, and announce the "wrong paste" channel.

  3. Weeks 5–6

    DPIAs and contracts. DPIAs for the two or three riskiest AI uses. Request processing agreements from the top ten vendors and record server locations and transfer bases.

  4. Weeks 7–8

    Breach. A one-page runbook, notice templates, and a one-hour tabletop drill with a realistic scenario.

  5. Week 9

    Board review. Rerun the 10-point checklist and record what is still "not yet" with a target date. That list is itself evidence that you are doing the work.

Notice what isn't on the list: buying compliance software, certification, or a large consultancy. Those may help later. For these nine weeks, what you need most is decisions and documents.

08What's still open

What's still moving.

An honest guide has to say what isn't known yet. As of 11 November 2026, five things are still open.

  • The PDP supervisory agency hasn't been formed. Its draft presidential regulation has gone to the President and is described as being at the final stage. Until the agency exists, Komdigi carries out its functions through a transition task force7.
  • About 30 matters are delegated to agency regulations, including the list of countries with equivalent protection and standard contract clauses for transfers abroad2.
  • "Large scale" has no numeric definition yet. It decides who needs a DPO and which DPIAs are triggered by scale.
  • The AI presidential regulations aren't signed. Once issued, extra duties on AI ethics and labelling AI content may follow20.
  • We haven't accessed the official text of PP 33/2026. Article numbers in this guide come from secondary sources that agree with each other. We will update it once the official copy is available.

None of these five is a reason to wait. The obligations still apply from 16 January 2027, and data subjects can still sue for damages through the civil courts8. What changes later is the detail, not the direction.

09Questions

Common questions.

Do small companies have to comply too?

The PDP Law applies to every personal data controller, with no exemption by size. What depends on scale are certain obligations, such as the DPO and some DPIA triggers. Business scale is also one of the factors in setting a fine.

Can employees use ChatGPT for work?

Yes, on three conditions: a company account, a written agreement with the vendor, and no red data (personal data of customers, employees, or patients) unless it goes through a system that has had a DPIA. The same applies to Gemini, Copilot, Claude, and other tools.

Do we need a DPIA for every prompt?

No. A DPIA is done per system or use case, such as "customer service chatbot" or "AI summaries of check-up results". It is reviewed when the risk changes, for example a new data type or a new vendor (Article 121).

The agency doesn't exist yet. Why not wait?

Because the obligations apply from 16 January 2027, Komdigi supervises during the transition, and data subjects can still sue. Compliance history and cooperation are also factors that lower a fine. The documents you make today are your compliance history later.

We process employee check-up results. What's different?

Health data is specific data (Article 6(1)(a)). That alone triggers a DPIA, and at large scale it triggers the DPO requirement. If AI is used to read or summarise the results, two DPIA triggers apply at once. HR should receive group-level figures, not individual results, unless there is a clear basis.

Our servers are in Singapore. Is that a problem?

Not a problem, but it is a transfer abroad that needs a basis. Until the list of equivalent countries is issued, the realistic basis is a binding agreement with your cloud provider, recorded in your records of processing and reviewed regularly.

10Closing

Closing.

PP 33/2026 is often read as a fine threat. We think the more useful reading is as a work list that should have existed anyway: knowing what data you hold, who is responsible, where the data goes, and what happens when something goes wrong. Companies that can answer those four aren't only more compliant. They are also readier to use AI with confidence.

If it has to come down to one step for this week: write a list of every AI tool used in your company, and one owner for that list. The rest gets easier after that.

The AI in your company is already running.
What's missing is the list and its owner.

Akses is a technology partner that builds systems, apps, and data reports. We don't give legal advice. But if the technical work needs sorting out, such as a data inventory, role-based access, activity logs, or moving processes out of spreadsheets and personal accounts into systems that can be audited, that is our work. See what we can do.

Useful? Share it with a director or COO who hasn't had time to read PP 33.

Method note and disclaimer

This guide draws on law firm and consultancy summaries (mainly RSM Indonesia's Client Alert of 18 September 2026), regulation databases, and news coverage as of 11 November 2026. At the time of writing we could not access the official copy of PP 33/2026 (State Gazette 2026 No. 88) on JDIH, so article numbers follow secondary sources that agree with each other. The content of the elucidation of Article 120(2)(f) naming AI rests on RSM alone. Third-party survey and telemetry figures are given with their year; security vendor data is marked as vendor data. The example in section 03 is an illustration, not an event at any particular company.

This is an operational guide, not legal advice, and it does not replace the opinion of legal counsel. This page will be updated when the official text, agency regulations, or the AI presidential regulations are issued. Akses · aksesgroup.tech

References

  1. RSM Indonesia. Client Alert: PP 33/2026, Six Implementation Priorities before 16 January 2027. 18 September 2026. rsm.global/indonesia/…/RSM Indonesia Client Alert - PP 332026.pdf
  2. PPC Land. Data controllers face 2% revenue fines under Indonesia's new data rules. 2026. ppc.land/data-controllers-face-2-revenue-fines-under-indonesias-new-data-rules
  3. Veritask. Technical regulation on personal data protection and the obligations of controllers and processors finally issued through Government Regulation No. 33 of 2026. 2026. veritask.ai/en/artikel/technical-regulation-on-personal-data-protection-…
  4. Justisio. PP 33 Tahun 2026, text of Article 6 (partial). justisio.com/regulation/indonesia/dbwgcly3epltypka7sebq61yx
  5. Meridian Hukum. PP No. 33 Tahun 2026 (LN 2026 No. 88, TLN No. 7190). meridianhukum.com/peraturan/pp-no-33-tahun-2026
  6. Katadata. PP Pelindungan Data Pribadi berlaku Januari 2027, ini aturan baru data anak. 3 September 2026. katadata.co.id/finansial/makro/6a9979b7c5745/…
  7. Katadata. Komdigi: Perpres lembaga pengawas PDP masuk tahap final. 5 November 2026. katadata.co.id/digital/teknologi/6aec368e80c97/…
  8. Bisnis.com. Masyarakat dapat gugat dan terima ganti rugi atas pelanggaran data pribadi. 2 September 2026. teknologi.bisnis.com/read/20260902/84/2000857/…
  9. Makarim & Taira S. Constitutional Court ruling means more firms must appoint a data protection officer. September 2025. makarim.com/news/constitutional-court-ruling-means-more-firms-must-appoint-a-data-protection-officer
  10. Republic of Indonesia. Law No. 27 of 2022 on Personal Data Protection. pasal.id/akn/id/act/uu/2022/27
  11. Microsoft & LinkedIn. 2024 Work Trend Index, Indonesia. 11 June 2024. news.microsoft.com/id-id/2024/06/11/…
  12. Microsoft. Work Trend Index 2026: 33% of Indonesian workers are at the forefront of AI adoption. 30 June 2026. news.microsoft.com/source/asia/2026/06/30/…
  13. PwC Indonesia. Global Workforce Hopes and Fears Survey 2025, Indonesia. February 2026. pwc.com/id/en/publications/consulting/hopes-and-fears-2025-indonesia.pdf
  14. BCG. AI at Work 2025: Momentum Builds, but Gaps Remain. 26 June 2025. 10,600+ workers in 11 countries. bcg.com/press/26june2025-beyond-ai-adoption-full-potential
  15. Netskope Threat Labs. Cloud and Threat Report 2026. January 2026. Security vendor data. threatlabs.netskope.com/reports/2026/cloud-and-threat-report-2026
  16. Cyberhaven. 2025 AI Adoption & Risk Report. 23 April 2025. Telemetry from 7 million workers, security vendor data. cyberhaven.com/press-releases/…
  17. CSO Online. Samsung bans staff AI use over data leak concerns. May 2023. csoonline.com/article/3695170/samsung-bans-staff-ai-use-over-data-leak-concerns.html
  18. OpenAI Help Center. Temporary Chat FAQ. help.openai.com/en/articles/8914046
  19. detikInet. Kominfo issues a circular on the ethical use of AI in Indonesia (SE Menkominfo No. 9 of 2023). December 2023. inet.detik.com/law-and-policy/d-7103666/…
  20. Kompas. Komdigi: AI presidential regulation awaits the President's signature, targeted this year. 3 September 2026. tekno.kompas.com/read/2026/09/03/20290047/…
  21. Hukumonline. Pemerintah akhirnya terbitkan kerangka implementasi pelindungan data pribadi, berlaku mulai 16 Januari 2027. 2026. pro.hukumonline.com/a/lt6a97c3fe4527c/…
  22. OJK. Tata Kelola Kecerdasan Artifisial Perbankan Indonesia (AI governance for Indonesian banking). April 2025. ojk.go.id/id/Publikasi/Roadmap-dan-Pedoman/Perbankan/…

All links accessed 11 November 2026.

Let's talk.

Tell us which point your checklist stops at. We help with the technical side: data inventory, role-based access, activity logs, and systems that can be audited.