Keep documents and prompts internal
Process confidential inputs locally without sending them to external AI APIs.
Security & enterprise control
Security for
local AI deployments.
Financial statements, valuation materials, transaction documents, client records, and proprietary research can contain confidential information, including material nonpublic information (MNPI). Local processing keeps this information within your enterprise environment.
Kairon designs and deploys workflows that reduce reliance on external processing and storage, subject to the client’s security policies.
Discuss your workflowData flow
Processing within
client infrastructure
This architecture describes fully local client deployments. It is separate from this public website and Kairon’s internal administration tools, which may use external services.
- 01 / INPUT
Internal documents
Files and context within your infrastructure
- 02 / PROCESS
Local workflow and AI processing
Models run on your enterprise compute
- 03 / OUTPUT
Internal outputs
Results and workflow data remain inside
No cloud AI inference required. No external storage required for workflow data.
Local processing
Models, documents, and outputs stay on client infrastructure. A local installation may still be connected to an internal network.
Offline operation
Core processing can run without internet connectivity. Setup, updates, and support arrangements are agreed for each deployment.
Network isolation
An air-gapped environment is isolated from external networks. That requires deliberate deployment configuration; it is not an automatic property of local software.
External exposure
Reduced dependence
on external services
Fully local workflows remove the need for external processing and storage in the core document workflow.
Keep workflow storage in your environment
Avoid maintaining workflow data in a separate provider’s cloud.
Run core processing independently
Use local models without depending on an external inference service or internet connectivity.
Provide access within the enterprise
Employees can use an internal deployment without a public-facing application endpoint. Deployment and access remain subject to your security policies.
Deployment comparison
Local inference costs
Fully local inference eliminates provider API token charges for that processing. Additional work uses available local capacity, which can reduce marginal inference costs for recurring, document-heavy workflows.
Scroll horizontally to compare both deployments.
| Dimension | Cloud API-based workflow | Kairon local deployment |
|---|---|---|
| Processing location | External provider infrastructure | Client enterprise infrastructure |
| External model API | Required for cloud inference | Not required for local inference |
| Workflow data storage | Depends on provider and configuration | Within client infrastructure |
| Internet dependency | Required to reach the external inference API | Not required for core local processing |
| Inference cost structure | Provider usage charges; product fees vary | Hardware and operating costs; no cloud-model API token charges |
Total costs include hardware, electricity, maintenance, and support. Savings depend on workload, utilization, hardware, and model requirements. Local models must be evaluated for the quality and speed your workflow needs; they do not universally match every cloud model.
Kairon provides workflow design, integration, deployment, and ongoing support, with commercial terms agreed for each engagement. Cloud providers’ storage, retention, and training policies vary by service and configuration; this comparison does not assume they all retain prompts or train on customer data.
Reference material
Security context
and further reading
Selected primary sources on sensitive information, service dependencies, and AI system design. These are different kinds of risk—not evidence that every breach is a cloud or AI failure, or that Kairon would have prevented an incident.
Reviewed
Industry report
2026 Data Breach Investigations Report
Verizon’s annual analysis examines breach patterns across industries, including vulnerability exploitation, third-party involvement, and unapproved AI use. The 2026 edition draws on 2025 data.
Why it mattersA broader basis for reviewing dependencies and security fundamentals, beyond any single incident.
AI security guidance
Prompt injection is not SQL injection (it may be worse)
The NCSC explains why malicious instructions within AI inputs create a persistent design risk, and recommends limiting the actions a model can cause a system to take.
Why it mattersKeeping inference local does not remove prompt-injection risk. Workflow boundaries and human review still matter.
Organization disclosure
Recent cybersecurity incident information and FAQ
Harvard disclosed unauthorized access discovered on November 18, 2025, following a phone-based phishing attack. The affected alumni and development systems held contact, donation, and related personal information; the notice does not establish exactly which records were accessed.
Why it mattersSensitive relationship records need protection, and identity verification remains essential regardless of hosting location.
Vendor disclosure
MySonicWall Cloud Backup File Incident
SonicWall’s October 8 investigation update confirmed unauthorized access to firewall configuration backups for all customers who had used its cloud backup service. The notice was first published September 17, 2025; these are disclosure dates, not the date of initial access.
Why it mattersCopies held by a separate service create another data boundary to assess. This incident concerned configuration backups, not AI inference.
Project inquiries
Discuss your requirements
Share an overview of your current process, data sources, and required outputs. We can assess the scope, integration requirements, and deployment options.