Key facts
| Workflow fit | Guidance, summarisation, translation and drafting; decisions stay with officials |
| Deployment tiers | Plugsky cloud, VPC, on-prem and air-gapped for classified or restricted networks |
| Residency | Region-locked planes including GCC (UAE), EU (Frankfurt), APAC (Singapore) and US (Virginia) |
| Identity | Scoped API keys for services; RBAC and SSO/SCIM for staff |
| Audit | Admin, key and inference events exportable to SIEM for records and oversight |
| Endpoints | Chat completions, streaming, JSON mode and function calling are live; assistants and responses endpoints are coming soon |
| Models | 30+ models including open-weight options suitable for offline operation |
| Procurement | Enterprise terms, SLA and deployment review documented for public sector buying |
TL;DR
- Automate preparation and explanation, not entitlement decisions.
- Design every agent around an accountable official and a reviewable record.
- Use region-locked or air-gapped deployment where data cannot leave the perimeter.
- Keep citizen data minimised and retention configured per service.
- Build on live chat completions; assistants-style endpoints remain roadmap items.
How it works, step by step
- Choose a citizen-facing task where the agent prepares or explains and a named official decides.
- Classify the data involved and select the deployment tier: cloud, VPC, on-prem or air-gapped.
- Create workspaces pinned to the required region with scoped keys per service and SSO/SCIM for staff access.
- Implement the tool loop with read-only access to the case management system, and route all write actions to human queues.
- Write the escalation rules first: when confidence is low, when the query is sensitive, or when the citizen asks for a person.
- Stream audit events to your SIEM and records system, with prompt retention set to the minimum your policy allows.
- Publish a plain-language notice explaining what the agent does, what it records and how to reach a human.
Try it yourself
Open the sovereign AI readiness score →
Where agents help in public services
The highest-value public sector agents reduce administrative load without inheriting legal authority. Common patterns: form completion guidance across long eligibility rules, case summarisation for caseworkers, correspondence drafting in a house style, translation for multilingual services, and internal knowledge search across policy manuals.
In Plugsky terms each is a tool loop on /v1/chat/completions. The model reads authorised context and returns either an answer or a tool_call your service executes against the case system. Entitlement calculations, sanctions and enforcement remain deterministic code or human decisions — never model judgement.
Sovereignty, records and oversight
Public sector deployments answer to records law and oversight bodies, so three properties matter: where processing happens, what is retained, and who can prove it. Plugsky supports region-locked planes, VPC and on-prem deployments, and fully air-gapped operation where networks are restricted or classified, using open-weight models that run inside your perimeter.
- Records: export administrative, key and inference events to your SIEM and records platform so decisions can be reconstructed.
- Minimisation: strip national identifiers and unnecessary personal detail before prompting; configure short retention.
- Access: separate platform administration from case data access so no single role holds both.
Designing for trust and escalation
Trust is a design property, not a disclaimer. Show the citizen when they are talking to an agent, give a one-step path to a human, and make the escalation reason visible to caseworkers. Log the agent's suggested answer and the official's final decision separately, so oversight can compare them.
Evaluate before scaling: build a test set of real query types, measure whether escalations land in the right queue, and review low-confidence handling with domain experts. Because 30+ models sit behind one compatible endpoint, you can adjust the quality and cost balance per service without changing the surrounding platform.
Honest comparison
| Requirement | Plugsky agents | Typical hosted agent platform | Building in-house |
|---|---|---|---|
| Restricted-network operation | On-prem and air-gapped tiers with open-weight models | Rarely supported | Possible but slow |
| Residency | Region-locked planes plus customer-controlled deployment | Limited region options | Wherever you host |
| Audit for records | Key, admin and inference events exportable to SIEM | Vendor dashboards | You build it |
| Model choice | 30+ models behind one endpoint | Vendor catalogue | You serve each model |
| Escalation design | Implemented in your application and case queues | Basic handoff features | Entirely your build |
| Time to pilot | Days on a free workspace | Procurement dependent | Months |
Frequently asked questions
Can AI agents make eligibility decisions?
No. Agents should prepare, explain and route; the decision belongs to an accountable official working from a reviewable record. Keep entitlement logic in deterministic code or human process.
Does Plugsky support air-gapped deployment?
Yes. On-prem and air-gapped tiers run open-weight models fully inside your perimeter, with updates delivered through your controlled process and the same OpenAI-compatible API.
How are audit requirements met?
Administrative changes, key lifecycle events and inference metadata can be exported to your SIEM and records systems, so reviewers can reconstruct what the agent produced and who decided.
What about multilingual services?
Agents can translate and draft in multiple languages through the same endpoint. Validate quality with native speakers before citizen-facing use.
How is citizen data minimised?
Redact identifiers before prompting, grant tools read-only access, and configure retention per service. Prompt retention is configurable in the platform.
Can we start without a full procurement cycle?
A free workspace with plugsky-micro and plugsky-lite, no card, supports an internal pilot on synthetic or public data while procurement runs in parallel. Plan details are on the live pricing page.
Who supports restricted-environment deployments?
Enterprise engagements cover deployment review, support expectations and SLA terms for restricted networks. Confirm specifics with the enterprise team during scoping.