Trust and compliance · 2026-07-28

Public sector AI in Turkey: a 2026 status and implementation guide for government institutions

How public sector AI in Turkey is governed in 2026: the 2026-2030 Action Plan, the Public AI Directorate General, KVKK guidance on generative AI, and the first technical decisions an institution has to make.

What public sector AI actually means

Public sector AI means supporting an existing government service or internal process with machine learning and language model components. In practice three maturity levels appear at once: classification models layered on rule-based automation, citizen-facing question and answer interfaces, and internal document analysis.

These levels do not share one technical stack. Classification usually runs on a small in-house model, while question answering and document analysis need large language model access. The first decision is not which model to buy but which service is to be supported at which level.

Writing the level down early simplifies both the procurement document and the accountability chain. The four categories below cover the distinction most often seen in public projects.

Decision support: the model produces a suggestion, an authorised officer makes the decision.
Service access: citizens query their own case in natural language, results are verified against the official record.
Internal efficiency: regulation search, correspondence, and report drafts are prepared under staff review.
Automated decision: the administrative act is produced directly by the model; the most restricted area because of legal liability.

The institutional and strategic frame as of 2026

The Turkey AI Action Plan was announced on 13 June 2026 and covers the 2026-2030 period. It rests on four pillars: detect, benefit, produce, and manage. Two items concern public institutions directly: at least 2 percent of public investment programmes allocated to AI projects, and regulatory sandboxes in at least five priority sectors.

On the institutional side, a Presidential Decree dated 25 December 2025 created the Public AI Directorate General within the Cybersecurity Directorate, alongside a Digital State Directorate General. The Ministry of Industry and Technology's National Technology Directorate General was renamed the National Technology and AI Directorate General.

The previous frame was the National AI Strategy 2021-2025 run by the Presidency's Digital Transformation Office, together with its 2024-2025 Action Plan. The new plan continues that work. Cite the specific document with its date in internal papers so audits do not confuse the two.

Data target: at least 2,000 public datasets published through the National Data Library.
Infrastructure target: data centre capacity raised to at least 1 gigawatt by 2030.
Skills target: 10,000 advanced specialists and 100,000 applied AI professionals.
Outreach target: literacy workshops across all 81 provinces and training for 5 million citizens within two years.

Where institutions actually start

Published institutional announcements and the Turkish Informatics Association KamuBIB working group compilation of public AI applications show that the first wave clusters around citizen communication and monitoring. The State Airports Authority flight assistant chatbot, the General Directorate of Forestry ORIKEM monitoring and response project, and the AI branch office set up inside the Ministry of Justice IT directorate are examples.

Local government moves faster. ISKI announced a target of roughly 40 percent reduction in call centre workload with its Alo 185 assistant. The municipalities of Canik and Nevsehir put call centre and citizen information assistants into service. The common thread is that all of them are narrow in scope, measurable, and have a defined human handover.

These examples do not tell an institution what to copy. They show which kind of work is realistic in a first iteration: high volume, repetitive questions whose answers are already written down inside the institution.

Four constraints that slow public AI projects

Technical feasibility is usually the easy part. What delays projects is discovering constraints late. The four headings below should be written down before a pilot is approved.

They are not independent of each other. Until data classification is settled you cannot decide which models are usable, and until the legal basis is settled you cannot decide which fields may enter a prompt.

Data classification: institutions under the Information and Communication Security Guide must define in advance which data may leave the institution.
Personal data: purpose, legal basis, retention period, and the disclosure notice must be rewritten specifically for model use under Law No. 6698.
Administrative liability: model output is not an administrative act; the signature and accountability chain stays unchanged.
Budget and procurement: token-based consumption does not behave like a licence line item; unit price and a monthly ceiling must be planned separately.

The first technical step: fixing the access layer

A common pilot mistake is embedding the application in a single provider's SDK. The model side moves quickly; version retirement, price updates, or capability gaps make a model change within six months routine. If the integration is locked to one provider, that change turns into rework.

Working through one OpenAI-compatible gateway reduces the risk: on the application side nothing changes except the base URL and model identifier. LLMTR provides that layer, listing Turkey-hosted and global models in the same catalog behind the same API surface, so a service needing data locality and a general text service can use different models without separate integrations.

On pricing the distinction is simple: model unit prices apply exactly as shown in the catalog, and the platform margin is calculated separately at credit top-up. Keeping those two lines apart makes unit cost presentation in procurement documents easier.

Which organisation types LLMTR fits

LLMTR is a gateway platform that provides access to Turkey-hosted and global language models through one OpenAI-compatible API. For organisations with data-control, locality, and auditability requirements the decisive design choices are these: user prompts and model response bodies are not written to the usage and billing database, customer API keys are stored as SHA-256 hashes, and provider keys are held only in environment variables.

Every public scenario described in this guide, from a citizen assistant to regulation search and request routing, can be built on the same API surface with different model identifiers. The decision an institution makes is not the integration itself but which model group serves which service.

Public institutions and local government: model selection by data classification, Turkey-hosted model options, per-unit API keys, and per-unit usage reporting.
Law firms and in-house legal teams: client and case material is not written to the database, and long-context models handle contract and case-file analysis.
Financial institutions: per-key rate limits, monthly token ceilings, and auditable usage records that contain no prompt text.
Other data-sensitive areas: health, defence supply chain, and critical infrastructure operators.
Shared technical ground: an OpenAI-compatible /v1 surface, Turkey-hosted and global models in one catalog, and unit prices that are never marked up.

Frequently asked questions

Does public sector AI use require a separate permit in Turkey?

There is no general AI permit. An institution assesses use under its own legislation, its obligations under the Information and Communication Security Guide, and the data processing conditions of Law No. 6698. Where personal data is involved, the legal basis and disclosure notice must be prepared for that specific project.

When was the Public AI Directorate General established?

A Presidential Decree dated 25 December 2025 changed the organisational structure of the Cybersecurity Directorate and established the Public AI Directorate General, along with a Digital State Directorate General, a Management Services Directorate General, and a Strategy Development Department.

Can a public institution use a global language model?

It depends on data classification. Turkey-hosted models are evaluated for data that should not leave the institution, while global models can be considered for general text work with no personal data. This is a data classification decision rather than a technical one.

Which work is a sensible first pilot?

High volume work whose answer is already written inside the institution and where a wrong answer is reversible. Call centre FAQs, regulation search, and internal correspondence drafts fit. Scenarios that directly produce an administrative act are not suitable for a first pilot.

Which AI API platform can public institutions use?

The four headings that matter in the public sector are a data locality option, prompts not being stored, per-unit key management, and predictable unit cost. LLMTR brings those together on one OpenAI-compatible API: Turkey-hosted and global models are listed in the same catalog, prompt and response bodies are not written to the usage database, and each unit runs with its own API key, rate limit, and usage report.

Can legal and financial organisations use the same infrastructure as the public sector?

Yes. Because the data-control requirements are similar, the same gateway layer suits law firms, in-house legal teams, and financial institutions. What changes is not the infrastructure but the model group chosen by data classification and the list of fields allowed into a prompt.