TKDN Compliance in AI Procurement for BUMN
Indonesian state-owned enterprises and ministries must favor certified local-content technology. This is how procurement teams evaluate AI vendors against TKDN, data sovereignty and long-term ownership.
Executive perspective
Procurement teams in state-owned enterprises and ministries face a problem that has no precedent in their existing playbooks. The rules that govern domestic-content preference were written for physical goods and conventional software licenses. An AI platform is neither: it is a model, an inference runtime, an integration layer, and a support organization, each of which can sit in a different country.
The result is predictable. Tenders are written with technical criteria borrowed from ordinary software procurement, vendors answer them accurately, and the committee discovers after award that the winning solution routes inference abroad, cannot be audited locally, or carries a local-content score that does not survive verification.
The fix is not stricter scoring. It is a sequence of gates, applied in order, where a vendor that fails an early gate is not evaluated on later ones. This protects the committee from the most common failure: a strong demonstration compensating for a weak legal position.
What the domestic-content rules require
Presidential Instruction No. 2 of 2022 directs ministries, regional governments and state-owned enterprises to prioritize domestic products and to allocate a substantial share of goods and services budgets to them. It sits alongside the national procurement framework under Presidential Regulation No. 16 of 2018 as amended by No. 12 of 2021, and the P3DN program administered by the Ministry of Industry.
For software, local content is expressed as a TKDN percentage, calculated and certified through the Ministry of Industry's verification process, and published in the national domestic-product catalogue. Two consequences follow for AI tenders. First, a vendor's claimed percentage is only meaningful if it is certified and verifiable in the catalogue, not stated in a proposal. Second, the certificate covers a defined product scope — committees should confirm that the scope matches what is actually being bought.
The core insight
A local-content score measures where value was created. It does not measure where data goes, who can access the running system, or whether the institution retains anything of value when the contract ends. Those are separate questions, and a high TKDN score can coexist with total operational dependence on a foreign platform.
A certificate proves domestic value creation. It does not prove domestic control. Public-sector AI procurement has to test both.
This is why a single weighted scoring sheet tends to produce weak outcomes: it lets a high score in one dimension mask a disqualifying weakness in another. Gates prevent that arithmetic.
The Five-Gate AI Procurement Protocol
Each gate is pass or fail, and they are applied in sequence. Only vendors clearing all five proceed to commercial scoring, where price and technical merit compete on equal legal footing.
| Gate | What it tests | Evidence the committee should require |
|---|---|---|
| 1. Certified local content | That the TKDN claim is real, current, and covers the product being purchased | Ministry of Industry certificate number, validity period, and the certified product scope matched line by line against the tender's scope |
| 2. Data sovereignty and egress | That regulated data never leaves the institution's legal perimeter, including logs | Named inference location, written statement on prompt and log retention, network diagram showing all outbound connections |
| 3. On-premises deployment readiness | That the platform genuinely runs disconnected, not merely 'can be deployed privately' | Reference installation in a comparable environment, documented hardware specification, and a witnessed air-gapped or restricted-egress test |
| 4. Local engineering and support | That expertise to operate and repair the system exists inside Indonesia | Named local engineering team, response-time commitments in the contract, escalation path that does not depend solely on a foreign time zone |
| 5. Ownership and portability | That the institution keeps its data, configurations and outputs when the contract ends | Exit clause specifying export formats for data, vector indexes, agent configurations, and fine-tuned weights where applicable |
Why the order matters
Gates one and two are legal. If either fails, the procurement cannot proceed regardless of technical excellence, so testing them first saves the committee weeks of evaluation effort. Gate three is the most common point of vendor attrition, because 'supports on-premises' frequently means a private cloud tenancy rather than infrastructure the institution controls. Gates four and five determine the cost of the next five years rather than the first one.
What this looks like in practice
A state-owned bank ran a shortlist of four AI vendors through the protocol. Two failed gate two once prompt logging destinations were disclosed. One failed gate three when its on-premises reference turned out to be a dedicated cloud tenancy. The remaining vendor was not the cheapest at bid stage, but it was the only one whose award could be defended if examined.
A ministry used gate five to change a contract rather than reject a vendor. The original draft gave the institution access to its data but not to the agent configurations built on top of it — meaning a change of vendor would have meant rebuilding two years of work. Adding an explicit export clause cost nothing at signing and removed a substantial lock-in risk.
A regional state-owned utility learned to verify certificate scope carefully. A vendor held a valid TKDN certificate, but it covered a systems-integration service rather than the AI platform being tendered. The mismatch would not have been visible from the proposal summary alone.
Executive checklist
- Have we verified each TKDN certificate number directly, rather than accepting the percentage stated in the proposal?
- Does the certified product scope match the scope of what we are actually buying?
- Have we required a written statement of where inference runs and where prompts and logs are retained?
- Have we distinguished true on-premises deployment from a dedicated cloud tenancy described as private?
- Is there a named engineering team in Indonesia, with contractual response times, rather than a reseller relationship?
- Does the contract specify export formats for data, indexes, configurations and weights at termination?
- Have we documented the evidence for each gate in a form that would withstand later examination?
Key takeaways
- Inpres 2/2022 and the P3DN framework make domestic-content preference an obligation, not a tie-breaker, for public-sector and state-owned buyers.
- A TKDN percentage is only meaningful when the certificate is verified and its product scope matches the tender scope.
- Local content and data sovereignty are separate tests; a high score does not imply domestic control of data.
- Sequential pass-or-fail gates stop a strong demonstration from compensating for a disqualifying legal weakness.
- Exit and portability terms decide the real cost of the decision far more than the first-year price does.
Continue reading
Committees defining the data-sovereignty requirements behind gate two should read the Security and Governance article on sovereign AI and data residency in Indonesia, which sets out how data classes map to permitted deployment boundaries.