Key facts
| Deployment models | In-region cloud, private endpoint in your VPC, on-prem, air-gapped |
| Residency | Region-locked planes: EU (Frankfurt), GCC (UAE), APAC (Singapore), US (Virginia) |
| Resilience | Automatic upstream failover; status page for component health; SLA at /legal/sla |
| Identity and access | SSO/SCIM, RBAC at workspace/role/resource level on Enterprise |
| Key custody | BYOK via AWS KMS, Azure Key Vault, HashiCorp Vault or on-prem HSM |
| Audit | SIEM audit log export; region-locked log storage |
| Models | 30+ models including open-weight options suitable for on-prem |
| Compliance posture | SOC 2 Type II and ISO 27001 readiness in progress (not yet certified) |
TL;DR
- Choose the tier from consequence of failure, not from cost alone.
- Keep AI out of safety-critical control loops unless the full assurance case is complete.
- Segment AI infrastructure from OT networks; treat the AI endpoint as an untrusted peer.
- Design for degraded operation: what the plant does when AI is unavailable.
- Have a governance path for changes, approvals and incident response that works offline.
How it works, step by step
- Map use cases to assets and classify by safety, availability and regulatory consequence.
- Select a deployment tier per use case: cloud, VPC, on-prem, air-gapped or edge.
- Segment AI infrastructure from operational technology networks with explicit boundary controls.
- Define degraded-mode behavior and manual fallbacks for every AI-assisted process.
- Implement identity, key custody and audit controls that operate within your environment.
- Validate models on representative operational data before production use.
- Exercise incident response and recovery with the AI systems in scope.
Try it yourself
Open the private LLM cost estimator →
Deployment tier selection
In-region cloud works for corporate functions: document drafting, knowledge search, support assistance. Private endpoint in your VPC removes shared-tenancy exposure for operational data. On-prem suits sites where network dependency is unacceptable or data cannot leave the facility. Air-gapped applies only where no external connectivity is permitted at all. Plugsky supports the same OpenAI-compatible API across these tiers, so a use case can move as its assurance requirements grow.
Availability and degraded modes
Operators need to answer one question before deployment: what happens when the AI is unavailable? Define a degraded mode for every assisted process — manual procedures, cached responses, or simply pausing the workflow. Model failover should be automatic and observable, and the status page plus audit logs should give you evidence during an incident. Treat the AI endpoint as one dependency among many, with the same availability discipline you apply elsewhere.
Security and segmentation
Industrial environments increasingly reference frameworks such as IEC 62443 for control-system security. The architectural implication is segmentation: AI infrastructure should sit in its own zone, communicate through controlled interfaces, and never be treated as implicitly trusted by OT systems. Pair this with strong identity (SSO/SCIM), least privilege, BYOK key custody and audit export so every request is attributable.
Regulation and governance
Sector regulators and national cybersecurity authorities set expectations that vary by country and asset class. Keep a register of AI use cases mapped to assets, approvals and controls, and review it with your security and safety teams. Where vendor certifications are still in progress — as with Plugsky's SOC 2 and ISO 27001 readiness — document compensating controls and the timeline in your risk register.
Common pitfalls
- Connecting AI directly to control networks without segmentation.
- No manual fallback when the model is unavailable.
- Using shared endpoints for operational data without isolation.
- Ignoring model version changes in change management.
- Treating a pilot as proof of plant-wide readiness.
Honest comparison
| Capability | Plugsky | Hyperscaler AI platform | Building in-house |
|---|---|---|---|
| Deployment tiers | Cloud, VPC, on-prem, air-gapped | Shared cloud, limited hybrid | You own the stack |
| Offline operation | On-prem and air-gapped with open-weight models | Limited offline support | Native if you build it |
| Failover | Automatic upstream failover | Service-dependent redundancy | You build routing |
| Identity and audit | SSO/SCIM and SIEM export | Mature enterprise controls | You build and staff |
| Residency | Region-locked planes | Regional footprint per service | Your facilities |
| Certifications | SOC 2 / ISO 27001 readiness in progress | Completed audits in many regions | Your own programme |
Frequently asked questions
Should critical infrastructure AI run in the cloud?
For administrative workloads, in-region cloud is usually acceptable. For operational data, prefer a private endpoint; where connectivity or jurisdiction demands it, use on-prem or air-gapped deployment.
How do we keep AI separate from OT networks?
Use zone segmentation with controlled interfaces, least-privilege access and strict egress rules. Treat the AI service as an untrusted peer, even when it runs on-prem.
What happens when the model fails?
Plan a degraded mode with manual procedures or cached results, and monitor failover through the status page and audit logs. Autonomy should never remove the manual path.
Can models run offline?
Yes — on-prem and air-gapped tiers run open-weight models without external network dependency, with updates delivered through your controlled process.
How do we govern model changes?
Treat model version changes like any operational change: assess quality and safety impact, test on representative data, and record approval in the use-case register.
What certifications should we require?
Use sector frameworks where mandated. Where a vendor's certifications are in progress, document compensating controls and monitor milestones rather than assuming coverage.
How do we get support in restricted environments?
Define the support model up front — remote assistance where permitted, on-site procedures where not — and capture response targets in the enterprise agreement.