Skip to content

Enterprise AI architecture

Enterprise AI data-flow atlas

Compare managed, customer-cloud, self-hosted and local-model deployments by tracing where each kind of data goes and who is left operating it.

Definition

The questions to ask an enterprise AI vendor are the same at every node of the data flow: which data class sits here, which processor holds it, in what location, for what purpose, for how long, who can reach it and who owns it. This atlas maps those questions across managed, customer-cloud, self-hosted and local-model patterns rather than promising how Marella will be deployed, so replace every generic node with the named service, operator and evidence for the proposed configuration.

01

Use one inventory

For each pattern, follow every copy your content makes, from the original file through to the backups and the deletion event at the end.

  • Source files and parsed content
  • Metadata and embeddings
  • Prompts, inputs and outputs
  • Logs, backups and deletion events
02

Managed pattern

  • Vendor-operated service boundary
  • Data and inference regions
  • Subprocessors and support access
  • Retention and recovery
03

Customer-cloud pattern

  • Account and resource ownership
  • Vendor identity and support routes
  • Network paths and upgrade control
  • Backup and cross-account processing
04

Self-hosted or on-premise pattern

  • Package dependencies and infrastructure
  • Secrets, identity and egress
  • Model services and monitoring
  • Backup, recovery and updates
05

Local-model variant

Verify that inference and every supporting dependency actually remain inside the intended boundary, and test whether the supported local model meets the workflow requirements.

  • Inference stays in boundary
  • Dependencies stay in boundary
  • Local model meets requirements
06

Record evidence for every node

  • Data class and processor
  • Location, purpose and retention
  • Deletion, access and encryption
  • Contract evidence and named owner

Architecture plates

Four patterns; every node still needs an owner

Pattern 01

Managed

Vendor-operated service boundary

  1. 1Approved source
  2. 2Managed parsing and retrieval
  3. 3Configured model path
  4. 4Answer, logs and support

Confirm location, providers, retention, support access and service responsibilities.

Pattern 02

Customer cloud

Customer account with shared operation

  1. 1Customer source
  2. 2Customer-cloud data services
  3. 3Approved inference path
  4. 4Shared monitoring and support

Name account ownership, vendor access, updates, backup and cross-account paths.

Pattern 03

Self-hosted / on-premise

Customer-operated infrastructure

  1. 1Customer source
  2. 2Customer runtime and storage
  3. 3Selected model path
  4. 4Customer logs, backup and recovery

Verify prerequisites, egress, support route, upgrade, vulnerability and recovery ownership.

Pattern 04

Local-model variant

Depends how the local model is set up

  1. 1Customer source
  2. 2Local retrieval context
  3. 3Locally operated model
  4. 4Local output and local logs

Verify that every dependency is local, supported and adequate for the evaluated workflow.

What this page does not prove

  1. B1A diagram does not prove the deployed system follows it.
  2. B2Self-hosted and local-model are separate properties.
  3. B3Support, telemetry and backup paths can cross an assumed boundary.
  4. B4Legal roles and transfer mechanisms require qualified review.

Related pages

Plan an evaluation