Key facts
| Access control | Scoped API keys with rotation; enterprise SSO and RBAC options |
| Deployment | Cloud, VPC, on-prem or air-gapped to match client requirements |
| Auditability | Request, model and response logging for review trails |
| Data grounding | Embeddings and RAG are live for policy and workpaper search |
| Structured output | JSON mode returns review notes and extraction records in fixed schemas |
| Retention | Log and data retention configured by your firm under its own policy |
| Pricing model | Flat monthly self-serve plans; no per-token billing on self-serve |
| Endpoint roadmap | Files, batch and fine-tuning are coming soon |
TL;DR
- Classify client data before connecting any AI workflow to it.
- Give each application and engagement-bounded system its own scoped key.
- Decide residency and retention explicitly; do not inherit defaults.
- Log requests, retrieved sources and model versions for review.
- Keep a human reviewer accountable for every accounting judgment.
How it works, step by step
- Inventory AI use cases and the client data each one touches, then label data classes.
- For each class, choose the deployment mode: cloud, VPC, on-prem or air-gapped.
- Issue per-application keys, store them in a secrets manager, and set a rotation schedule.
- Define log fields and retention: request ID, model, retrieved sources, output, reviewer.
- Approve a model allow-list tied to evaluation results, and re-test on model changes.
- Require citations and a qualified reviewer before any output informs client work.
- Run a tabletop exercise for a wrong or leaked answer, including client notification paths.
Try it yourself
Open the AI API key security checklist →
Scope and data classification
Security starts with knowing what the system can see. Split content into classes: public guidance, firm methodology, client financial data, and identifiers such as tax numbers or bank details. Each class gets its own rules for whether it may leave the firm's boundary and what logging it requires.
Most firms start with methodology and public guidance, then extend to client data only on private deployments where prompts, documents and embeddings stay inside the environment.
Identity, keys and least privilege
Model credentials deserve the same discipline as database credentials. Issue a distinct API key per application and environment, store it in a secrets manager, rotate it on a schedule, and revoke on role changes. Where the platform supports enterprise identity, connect SSO and role-based access so console permissions follow your joiner-mover-leaver process.
Never share a production key across engagements, and never place keys or client identifiers in prompts.
Residency, retention and audit trails
Decide explicitly where processing happens and how long records live. Region selection covers many residency needs; VPC, on-prem and air-gapped deployment covers mandates that require data to stay inside the firm. Retention is your policy applied to prompts, outputs, logs and retrieval indexes - not a platform default.
Log enough to reconstruct a decision: request ID, model name and version, retrieved source identifiers, output and the reviewing user. See AI audit logs for a schema.
Model governance and human accountability
Keep an approved model list with evaluation evidence for each model, and re-run evaluations when versions change. Restrict high-risk workflows to models and prompts that passed review, and require citations from approved sources. Regardless of automation, a qualified professional remains accountable for the work product, so the checklist should name that reviewer in the workflow design.
Honest comparison
| Control area | Plugsky capability | Common gap | Owner |
|---|---|---|---|
| Identity | Scoped keys, rotation, enterprise SSO and RBAC options | Shared production keys | Security engineering |
| Data boundary | Cloud, VPC, on-prem, air-gapped deployment | Data path undocumented | Risk and compliance |
| Retention | Configurable logging of requests and responses | No defined retention | Privacy officer |
| Audit trail | Request, model and source logging | Outputs not traceable | Quality review |
| Model governance | Model choice behind one API with evaluation | Silent model changes | Technical committee |
| Human sign-off | Citations and structured output for reviewers | Unsigned AI drafts | Engagement lead |
Frequently asked questions
Does using Plugsky make our firm compliant?
No. Compliance is your program. Plugsky provides deployable controls - scoped keys, private deployment, logging - that you document and audit against your own professional obligations.
Can client data stay inside the firm?
Yes. VPC, on-prem and air-gapped deployments keep prompts, documents and embeddings inside your environment, and the API stays OpenAI-compatible.
What should we log?
Request IDs, model names and versions, retrieved source identifiers, outputs and the reviewing user, retained under your policy. That is enough to reconstruct how an answer was produced.
How do we handle tax identifiers?
Keep them out of prompts where possible, retrieve only the fields required, and use private deployment for workflows that must process them.
Is fine-tuning available for firm templates?
Fine-tuning and batch endpoints are coming soon. Today, use retrieval over approved templates with JSON mode for consistent output structure.
How often should keys and models be reviewed?
Rotate keys on a fixed schedule and after staff changes; re-evaluate models whenever a version changes or a new model is added to the allow-list.
Where do we start?
Pilot on internal methodology documents using the free plan, prove accuracy and logging, then extend to client data on a private deployment.