Secure AI agents for public institutions

On June 17, 2026, Siesta AI presented its approach to secure AI agents at the Czech Chamber of Deputies during an internal event focused on e-Stat. The discussion centered on a practical question public institutions face every day: how to speed up research, drafting, and administrative work without weakening security, auditability, access control, or existing IT architecture.

As public institutions explore AI, the key question is no longer whether models can draft text or summarize documents. The harder question is whether they can operate in regulated environments where every source, permission, action, and output must be explainable and reviewable.

eStat and the challenges of legislative workflows

eStat was discussed as a platform intended to support legislative workflows and contribute to the wider modernization of public administration. In that context, Siesta AI shared how secure AI agents can support document-heavy work without replacing core systems or changing established decision-making processes.

Public institutions typically have access to a lot of information. The operational challenge is managing it: documents spread across multiple repositories, parallel versions circulating by email, permissions that differ by department, and decisions that must be justified with references that hold up under review. This is exactly where a governed “assistance layer” can reduce manual work while keeping responsibility where it belongs.

Where secure AI agents can help, without replacing core systems

In legislative and administrative workflows, value often comes from removing “hidden labor” that sits around the official decision: searching for sources, checking whether a reference is still valid, comparing versions, or turning meeting outcomes into structured next steps. When those steps are done manually, the process becomes slower and more error-prone, especially under time pressure.

A concrete reference point is the Siesta AI demo environment at https://zakony.siesta.ai/. It shows how source-grounded retrieval and structured outputs can support legal and legislative work, for example by:

surfacing potentially relevant provisions of Czech law for reviewer assessment
highlighting areas that may warrant constitutional-law review
linking to relevant EU law and obligations where applicable
flagging potential duplications, overlaps, or conflicts with existing regulation
supporting legislative-technical review, including missing definitions, broken references, and internal inconsistencies
preparing structured drafts, summaries, and amendment-style text based on provided inputs and attachments

Beyond drafting, governed agents can also assist with administrative work such as summarizing long documents, comparing document versions, preparing structured meeting follow-ups, and extracting action items for human confirmation. In regulated environments, these capabilities are most useful when citations, permissions, and approval steps are visible, not implicit.

What public institutions should measure

To keep evaluation grounded, focus on outcomes you can verify in real work: the time it takes to find relevant sources, prior drafts, and related amendments; the end-to-end drafting cycle time for briefs, summaries, amendments, and committee notes without lowering review quality; the error rate caused by outdated versions or missed dependencies; and the citation acceptance rate, meaning how often reviewers can approve outputs without sending them back for rework.