Key facts
| Access control | Scoped API keys with rotation; enterprise SSO and RBAC options |
| Tenancy | Separate keys and retrieval indexes per client environment |
| Deployment | Cloud, VPC, on-prem or air-gapped, per client if required |
| Auditability | Request, model and response logging for service delivery review |
| Data grounding | Embeddings and RAG are live for runbooks and client documentation |
| Models | 30+ models behind one OpenAI-compatible API |
| Pricing model | Flat monthly self-serve plans; no per-token billing on self-serve |
| Endpoint roadmap | Files, batch, assistants and fine-tuning are coming soon |
TL;DR
- Classify each client's data before any shared workflow touches it.
- Separate keys and retrieval indexes per client, not per team.
- Document subprocessors and data flows for client assurance packs.
- Log requests, sources and reviewers for service-delivery disputes.
- Keep a named owner for each client's AI workflow and exceptions.
How it works, step by step
- Inventory AI use cases across service desk, delivery and managed operations.
- Classify data per client and record contractual handling restrictions.
- Choose deployment per client: shared cloud, dedicated VPC, on-prem or air-gapped.
- Issue per-client and per-application keys with rotation and central inventory.
- Define log fields and retention: request ID, model, sources, output, client, reviewer.
- Approve a model allow-list and document it in client assurance materials.
- Review client consent and subprocessor disclosure before extending to new data.
Try it yourself
Open the AI API key security checklist →
Client environments and data boundaries
An IT services provider holds data from many client environments, each with its own contractual rules. Classification should record the client, the data class and the permitted processing location. Public vendor documentation and internal runbooks are low risk; client source code, tickets, logs and production configuration are not.
Write the boundary down before automating anything, because a single shared retrieval index across clients is the most common and most damaging design mistake.
Keys, tenants and least privilege
Issue a distinct API key per client environment and per application, store them in a secrets manager, and rotate on a schedule. Keep retrieval indexes and conversation history separated per client so one tenant cannot retrieve another's content. Enterprise SSO and RBAC options keep console access aligned with delivery-team changes, and offshore or subcontractor access should be scoped and time-bound.
Never place client identifiers or credentials into prompts; retrieve only the fields a task requires.
Residency, retention and service audit trails
Decide where processing happens and how long records live, per client. Region selection covers many residency needs; VPC, on-prem and air-gapped deployment covers clients who require data to stay in their own environment. Retention is your policy applied to prompts, outputs, logs and retrieval indexes, and it may differ per contract.
Log enough to reconstruct a service action: request ID, model and version, retrieved source identifiers, output, client and reviewer. See AI audit logs for a schema.
Model governance and delivery accountability
Keep an approved model list with evaluation evidence and re-test when versions change. Ground answers in current client documentation with citations so engineers can verify before acting. A named delivery owner should remain accountable for each client's workflow, and the checklist should document how subprocessors are disclosed and how client consent is captured before extending AI into a new service line.
Honest comparison
| Control area | Plugsky capability | Common gap | Owner |
|---|---|---|---|
| Tenancy | Separate keys and indexes per client environment | One shared index for all clients | Architecture |
| Identity | Scoped keys, rotation, enterprise SSO and RBAC options | Shared engineer credentials | Security engineering |
| Data boundary | Shared cloud, dedicated VPC, on-prem or air-gapped | Client rules undocumented | Service delivery |
| Retention | Configurable logging per client policy | Single blanket retention | Compliance |
| Audit trail | Request, client and reviewer logging | Actions not traceable | Internal audit |
| Disclosure | Model and endpoint documentation for assurance packs | Subprocessors undisclosed | Account management |
Frequently asked questions
Does Plugsky make us compliant?
No. Compliance is your program. Plugsky provides deployable controls - scoped keys, separated indexes, private deployment, logging - that you document in client assurance materials.
How do we keep client data separated?
Use a distinct key and retrieval index per client environment, and never let a shared workflow query across tenants. Private deployment is available per client where contracts require it.
What should we log?
Request IDs, model names and versions, retrieved sources, outputs, the client and the reviewer, retained under the applicable contract so service actions can be reconstructed.
Can we run AI in a client's own environment?
Yes. VPC, on-prem and air-gapped deployment keeps prompts, documents and embeddings inside the client estate, and the API stays OpenAI-compatible.
What do clients ask about subprocessors?
They ask which vendors process their data and where. Keep the model list and deployment architecture current so you can answer without a bespoke review each time.
Can we white-label this for clients?
Yes. The API is OpenAI-compatible and keys can be scoped per client, so you can operate AI services under your own brand and contract terms.
Where should a pilot start?
Pilot on public vendor documentation and internal runbooks with the free plan, then extend to client-specific content with per-client keys and private deployment.