Skip to content
Contact

SECURITY AND PRIVACY

Data stays within the organization's own boundaries

If you work in a regulated industry, your question is not "what can AI do." Your question is: where does the data sit, where does the model send calls, and who can see what. What follows is how ARCH deployments answer.

01 /

The model runs on the organization's hardware

ARCH deployments run on the organization's own infrastructure (on-premises). Retrieval, inference, and generation all take place within this boundary; organizational data is never sent as a prompt to an outside model.

Security perimeter — every step inside the organization's boundaryA single organizational boundary enclosed by a perimeter wall: retrieval, inference, and generation run in a single processing pipeline inside, and answers are generated only from the organization's own document set. The line to the outside model is cut and no call passes over it; every question-document-approval chain is written to the audit trail, and records whose retention period has expired are destroyed automatically.EXTERNAL MODELno call crosses the boundaryDOCUMENT SETclosed knowledge boundaryORGANIZATION BOUNDARYon-premises perimeter wallPROCESSING PIPELINEretrieval · inference · generationACCESS LINESsame question, two document setsRETENTION AND DESTRUCTIONautomatic at end of periodAUDIT TRAILquestion → document → approvalDIAGRAM · SECURITY ZONE
A single processing pipeline inside the organization's boundary: a question takes its answer only from the organization's own document set; documents the user cannot see do not enter the answer, and no call goes out to an external model. Every question–document–approval chain is written to the audit trail, and records whose retention period has expired are destroyed automatically.

A single processing pipeline runs inside the wall around the organization's boundary; answers are generated only from the document set inside, and no model calls go out. Elements:

  • Organization boundary: on-premises perimeter wall
  • Processing pipeline: retrieval · inference · generation
  • Document set: closed knowledge boundary
  • Access lines: same question, two document sets
  • Audit trail: question → document → approval
  • Retention and destruction: automatic at end of period
  • External model: no call crosses the boundary

02 /

  • 01

    On-premises operation

    Deployment takes place on the organization's own hardware. Data does not leave the organization's network to be processed; no model calls go outside.

  • 02

    Closed knowledge boundary

    An answer to a question is generated only from the organization's own document set. The model does not use general knowledge outside that set as a source — and it can be shown that it does not.

  • 03

    Who can see what

    Access is defined at the document and record level. A document a user cannot see does not appear in the answer to that user's question either.

  • 04

    Audit trail

    Who asked what, which document the answer was generated from, who approved what — all traceable after the fact. The audit trail is not a deletable side record; it is the system itself.

  • 05

    Retention and destruction

    The retention period for each data type is defined at deployment, and destruction runs automatically at the end of the period. The period is as long as legislation requires; no longer.

  • 06

    Certification

    Our information security and quality management systems are certified. Certificate details are shared on request through our contact channels.

03 /

Not a wrapper application

Many "AI" products on the market are wrappers that pass the user's data to a third-party model and display the returned answer in their own interface. In a regulated organization, this creates two problems: data leaves the organization's boundary, and it cannot be shown which source the answer came from.

ARCH does not use this model. Deployment is on the organization's hardware, the knowledge boundary is the organization's document set, and the source of every answer is recorded.

Certification

  • ISO 27001:2022 certificate

    Scanned image of the ISO 27001:2022 certificate certificate

    CERTIFICATE

    ISO 27001:2022

    Information security management system

    Certificate no.
    2026/ISMS/002587
    Founded
    IQR International (IAS accredited, MSCB-135)
    Valid until
    26.07.2027
  • ISO 9001:2015 certificate

    Scanned image of the ISO 9001:2015 certificate certificate

    CERTIFICATE

    ISO 9001:2015

    Quality management system

    Certificate no.
    QMS-017210
    Founded
    IQR International (IAS accredited, MSCB-135)
    Valid until
    26.07.2027

Let's discuss deployment boundaries

Let's write down together which data sits where, who holds which permission, and how the audit trail is kept.