Etken Yapay Zekâ (Agentic AI) Guidance — KVKK's March 2026 Framework
On 12 March 2026, the Kurul published Etken Yapay Zekâ Sistemlerine İlişkin Rehber — its guidance for agentic AI. "Etken yapay zekâ" is KVKK's own Turkish term for what the market calls agentic AI; use it when searching kvkk.gov.tr directly, the English term won't surface it.
Turkey was not late to this. Market commentary places the publication alongside three other major authorities doing the same thing in the same window: the UK ICO (8 Jan 2026), the EDPS (Nov 2025), and Spain's AEPD (18 Feb 2026). For a KOBİ or enterprise deciding whether Turkish regulatory readiness lags the EU/UK on this specific question — it doesn't. Full comparison, emphasis by emphasis: KVKK Among Four — How Agentic AI Guidance Compares Internationally.
How KVKK defines the thing it's regulating
The guidance characterizes agentic AI by three properties, more pronounced than in "traditional" AI: goal-orientation, autonomy, and environmental interaction. Operationally, the system runs through AI agents, and multi-agent systems may activate as task scope expands — language that maps directly onto how a corporate-brain deployment actually grows (single agent → agents that read and write shared knowledge → agents that hand off to each other).
The four things the guidance asks for
1. Human oversight (insan gözetimi). Human-centered design and meaningful oversight — not a rubber-stamp human in the loop. Critically, the guidance asks for this to be decided in advance: at which stages, under which conditions, does a human actually intervene? "We'll have a human review it" is not itself a control unless the review points are specified.
2. Traceability and transparency (izlenebilirlik ve şeffaflık). Interactions between system components — not just system-to-user — must be traceable, with control mechanisms that catch unwanted outcomes early. For a multi-agent setup, this means the handoffs between agents need an audit trail, not just the final output.
3. Accountability structure. Clear division of task, authority, and responsibility between developers, deployers ("yerleştiriciler"), and other actors in the chain, with coordination between them treated as a requirement, not a nicety.
4. Privacy by design / by default, applied across the full system lifecycle — the same standard Art. 4 principles already require, restated for autonomous systems specifically.
DPIA: what changes
Existing Data Protection Impact Assessments do not automatically cover an agentic upgrade. The guidance says DPIA processes should be updated to address:
- Increased autonomy levels — a DPIA written for a single-turn assistant does not cover an agent that now takes multi-step actions unsupervised.
- Multi-agent architectures — risk compounds at the handoffs, not just within one agent's reasoning.
- Special-category data inference — an agent can infer special-category data (health, religion, etc.) from ordinary inputs even without being fed it directly; the guidance calls for technical measures to prevent this incidental processing, and for verification mechanisms at critical decision points.
Why this page reads like a description of the corporate-brain method
Re-read "traceability," "verification mechanisms at critical decision points," and "clear division of responsibility" above. That is, almost verbatim, the argument for page contracts, lint gates, and audited write access — except here it's a regulator requiring it, not an architecture opinion. The practical takeaway for deployment patterns is in Pattern 5: the decision trail this guidance asks for and the decision trail good agent architecture already wants are the same artifact.
---
Educational reference maintained by Avalanche AI — not legal advice. This is recent guidance and among the fastest-moving material in this wiki — verify against the current text at kvkk.gov.tr before relying on it.