Running a pilot in your own cloud: exactly what IT has to do

Your cloud, your controls

NIST SP 800-190 describes risks and defences for containerised applications, including image integrity, registry access, and runtime isolation — relevant when agent workers run as services in your cluster. NIST SP 800-53 provides a control catalogue teams often map to cloud reviews. These publications do not endorse any vendor; they give shared language for the checklist your security team already uses.[1][2]

Stellitron’s public trials run in multi-tenant workspaces on dummy data. An enterprise pilot can instead deploy connectors and agent runtimes in your AWS, Azure, or GCP account (or on-prem Kubernetes) so documents and API responses do not leave the region you designate. That model trades convenience for control; IT tasks below are the usual minimum.[1][2]

Network and identity

Provision a private subnet or serverless environment that can reach your ERP, P2P, or bank APIs — and nothing else. Document inbound and outbound allow-lists: Coupa or SAP endpoints, your SFTP drop, optional email intake, and the Stellitron support bastion if you permit break-glass access. Deny open internet egress by default.

Create an integration identity per connector with least privilege. Prefer short-lived tokens or workload identity over long-lived passwords in configuration files. Map that identity to your IdP for operator sign-on to the agent UI — SSO through SAML or OIDC so offboarding an employee removes console access promptly.

Separate secrets: API keys for Coupa, SAP, or bank feeds live in your secret manager; the runtime reads them at boot. Rotation should not require redeploying application code — only updating the secret reference and rolling pods.

Logging, residency, and roll-back

Forward application logs and agent traces to your SIEM. Include correlation IDs that tie a proposed hold to the document versions retrieved. Retention should match your records policy; pilots often start with ninety days and extend after legal review.

Data residency is where compute and storage run, not where the vendor is incorporated. Pin regions to match your ERP and backup strategy. If models are invoked, document which subprocessors process text and whether prompts contain production identifiers — many pilots redact or hash vendor names in logs while keeping full detail in encrypted object storage inside your account.

Roll-back is disabling the connector identity and scaling workers to zero — Coupa and SAP see no further pulls, and no write-back jobs queue. Keep a runbook: who can revoke credentials, how to drain in-flight batches, and how AP reverts to the pre-pilot workflow without losing audit history already exported to your archive.

Security sign-off before production data

Complete a threat sketch: stolen integration credential, misconfigured egress, prompt injection on uploaded PDFs, and insider misuse of break-glass support access. For each, note detection and response — not just prevention. Red-team optional PDFs with hidden instructions to confirm the verifier and rules layer do not execute arbitrary actions.

Run a tabletop with AP or finance: connector disabled mid-batch, duplicate write attempt, and stale snapshot used for a hold recommendation. Measure whether operators can explain state without vendor intervention.

This checklist complements your own cloud standards; it does not replace a formal risk assessment. Timelines depend on your change advisory board, not on Stellitron alone.