Skip to content

ai-stack-layers.md

Latest commit

 

History

History
40 lines (29 loc) · 3.04 KB

ai-stack-layers.md

File metadata and controls

40 lines (29 loc) · 3.04 KB
Raw
Error in user YAML: (<unknown>): mapping values are not allowed in this context at line 2 column 35
---
title: AI Stack Layers — From Data to Oversight
source-id: (multi-source synthesis: stanford-ai-index-2026, mit-cisr-2024, isg-genai-2025, landing-ai-playbook)
wiki-page: (02_References/enterprise-ai/wiki/frameworks/ — multi-source synthesis)
last-synced: 2026-06-07
---

AI Stack Layers

What this reference contains

A six-layer framework for diagnosing an enterprise AI technology stack. Each layer is a potential binding constraint — weakness at any layer prevents value at the layers above.

The six layers

Layer What it is Binding-constraint signal
1. Data Data pipelines, quality, access, governance, warehousing Pilots fail at scale because data "should exist" but isn't pipeline-ready
2. Model Foundation model selection, fine-tuning, model lifecycle management Model is over-invested relative to orchestration; moat-building potential is low
3. Orchestration Workflow routing, agent coordination, model abstraction layer, tool integration Agents work in isolation; no cross-function value; cannot scale to Stage 3
4. Tools and applications APIs, SaaS integrations, end-user-facing tools Tool fragmentation; shadow AI; no shared consumption layer
5. Oversight and governance Human-in-loop design, audit trails, alert thresholds, model performance monitoring No visibility into AI behavior post-deployment; accountability gaps
6. Operations and lifecycle Model versioning, retraining triggers, incident response, decommissioning Models drift silently; no update process; deployed systems outlive their validity

Key insights

  • Data is the most common binding constraint. Data engineers and software engineers are tied as the most in-demand AI roles (Stanford AI Index 2026). The limiting factor is rarely the model.
  • For every $1 of visible tech investment, up to $10 is invisible — mostly data and change management (Stanford AI Index 2026).
  • The orchestration layer is where the moat lives (MIT NANDA 2025; Accenture). A European energy company using Accenture's AI core: 5 months to ship AI apps vs. 18 months pre-AI-core.
  • Model selection is the most over-optimized decision. Foundation models commoditize. The strategic question is orchestration, workflow integration, and data.
  • Oversight and operations are the Stage 2 → Stage 3 governance prerequisite. MIT CISR Stage 3 requires dashboards, transparent AI outcomes, and test-and-learn culture — all Layer 5–6 capabilities.

When to consult

  • For skill tech-stack-diagnostic: this is the primary assessment framework — walk each layer, score against the signal indicators, identify the weakest link.
  • For skill tech-buy-vs-build: the orchestration layer is where BUILD investment is justified; models and tools are where BUY almost always wins.

Source

Multi-source synthesis: Stanford AI Index 2026 (data as binding constraint, invisible costs), MIT CISR 2024 (Stage 3 platform requirements), MIT NANDA 2025 (orchestration as moat), Accenture (AI core case study, 5 vs. 18 months).