Key facts
| Safe scope | Administrative and support tasks; clinical decisions remain with licensed staff |
| API | OpenAI-compatible /v1/chat/completions with streaming and function calling (live) |
| Residency | Region-locked planes: EU (Frankfurt), GCC (UAE), APAC (Singapore), US (Virginia) |
| Deployment | Cloud, VPC private endpoint, on-prem and air-gapped for PHI-sensitive workloads |
| Data controls | Configurable prompt retention and PII minimisation before inference |
| Audit | Key, admin and inference events exportable to SIEM for investigation and review |
| Access | Scoped API keys, RBAC and SSO/SCIM for clinical and administrative users |
| Endpoint status | Chat completions is live; assistants and responses endpoints are coming soon |
TL;DR
- Automate paperwork and summarisation; keep diagnosis and prescribing human.
- Minimise PHI before prompting and set retention to the shortest defensible window.
- Pin inference, embeddings and logs to one region-locked plane.
- Give agents read-only tools and route all writes into clinical review queues.
- Export audit events so any output can be traced to model, prompt and reviewer.
How it works, step by step
- Select an administrative workflow — referral summarisation, prior-auth drafting or patient-message routing — and define what the agent may never do.
- Classify the data it touches and choose the deployment tier, keeping PHI inside the region or perimeter your policy requires.
- Create a workspace with scoped keys per service and SSO/SCIM for staff, separated by role.
- Build the tool loop with read-only access to the record system; return drafts to a clinician review queue rather than writing directly.
- Redact direct identifiers before prompting and configure prompt retention to the minimum your policy allows.
- Add escalation rules for urgent, ambiguous or emotionally sensitive messages, with a visible path to a human.
- Run a clinical review of a sample of outputs, then stream audit events to your SIEM and monitor override rates.
Try it yourself
Open the agent workflow designer →
Safe first use cases
Healthcare value concentrates in documentation and coordination. Discharge summary drafting, referral letter preparation, prior-authorisation packets, coding suggestions with citations, patient-message triage and policy search all save clinician time while leaving judgement intact.
Each is a tool loop on the OpenAI-compatible endpoint: the agent reads authorised context, drafts an output, and hands it to a clinician. Nothing the agent produces reaches the record or the patient without human review, and every draft carries its provenance — model, prompt version and timestamp.
PHI handling by design
Treat PHI as radioactive: minimise what leaves your systems, constrain where it lands, and keep evidence of every access. Plugsky supports configurable prompt retention, region-locked data planes and deployment in your VPC, on-prem or air-gapped, so sensitive workloads can run inside your perimeter.
- Minimise: strip names, record numbers and dates not needed for the task before prompting.
- Constrain: pin workspaces and log streams to one region; keep keys in the same jurisdiction.
- Prove: export access, key and inference events to your SIEM for investigation and audit.
Human-in-the-loop and evaluation
Design the review step before the automation. Drafts go to queues routed by urgency and specialty; the clinician sees the model's suggestion, its sources and a one-click way to reject it. Track acceptance, edit distance and override reasons — those metrics tell you whether the agent is genuinely helping or quietly adding work.
Evaluate on representative cases with clinical input, and re-run the suite whenever a prompt, tool or model changes. The 30+ model catalogue lets you test a smaller model against a frontier one on your own data before deciding what each workflow needs.
Honest comparison
| Requirement | Plugsky agents | Typical hosted agent platform | Building in-house |
|---|---|---|---|
| PHI residency | Region-locked planes plus VPC, on-prem, air-gapped | Limited region control | Wherever you deploy |
| Retention control | Configurable prompt retention per workload | Often fixed | You implement it |
| Clinical safety | Human review enforced in your application | Basic approval features | Entirely your build |
| Audit | Key, admin and inference events exportable to SIEM | Vendor dashboards | Your pipelines |
| Model choice | 30+ models behind one endpoint | Vendor catalogue | You serve each model |
| Time to pilot | Days on a free workspace | Varies | Months |
Frequently asked questions
Can an agent diagnose or prescribe?
No. Agents should summarise, draft and route. Diagnosis, prescribing and triage decisions remain with licensed clinicians who review every output.
How is PHI protected during inference?
Minimise identifiers before prompting, pin processing to a region or your own perimeter, set short retention, and export access events to your SIEM. Enterprise plans add customer-managed keys.
Does Plugsky sign a DPA?
Enterprise agreements include data-processing terms; review them with counsel before processing PHI. The published DPA guidance article lists the clauses to check.
Which workflows give fastest value?
Documentation and coordination: referral summarisation, prior-authorisation drafting, patient-message routing and policy search. They reduce administrative load without touching clinical judgement.
How do we measure whether it works?
Track draft acceptance rate, edit distance, override reasons and time saved per case. Review a clinical sample monthly and before any scope expansion.
Can we run entirely on-prem?
Yes. On-prem and air-gapped tiers use open-weight models with the same OpenAI-compatible API, so application code does not change between tiers.
How do we start cheaply?
Use the free plan with plugsky-micro and plugsky-lite on de-identified data, then scale to a paid plan or the 14-day full-access trial. See the live pricing page for details.