Briefs 7-9:Enterprise AI Ecosystems, Governance, and Human Oversight Study Notes

Vocabulary flashcards covering Enterprise AI Ecosystems, AI Ethics, Bias, Governance, Generative vs. Agentic AI, and Human-in-the-Loop design from Briefs 07, 08, and 09.

Enterprise AI Ecosystem

The full set of technologies, vendors, partners, data assets, and people through which AI capability reaches business processes.

Enterprise AI Stack

The layered assembly of AI capabilities comprising infrastructure, foundation models, data, enterprise platforms, agents/applications, and people/partners.

Build-Buy-Partner Decision

The strategic sourcing logic to build what differentiates, buy commodity capabilities, and partner where speed and expertise matter more than ownership.

Ten-Twenty-Seventy (10-20-70) Rule

BCG finding that roughly 10 percent of AI value comes from algorithms, 20 percent from technology, and 70 percent from people, process, and operating-model change.

Agent Sprawl

The uncoordinated proliferation of AI agents across different vendors, tools, and departments.

Agentic AI Mesh

An architecture for coordinating custom-built and off-the-shelf agents within a unified framework.

Digital Core

The modernized platforms, processes, and data foundation a company runs on before AI can scale.

Vendor Lock-in

The switching costs created when an organization's business processes and data become deeply entangled with one supplier's platform.

Shadow AI

Unsanctioned employee use of external or personal AI tools outside official organizational channels and data controls.

Lilli

McKinsey's internal generative AI tool built on more than 100,000 internal documents to synthesize proprietary firm knowledge.

AI Ethics

The values and responsibilities that guide appropriate AI system design and use, asking what should be done rather than what can be automated.

AI Governance

The policies, roles, controls, metrics, and monitoring practices used to manage AI deployment responsibly.

Responsible AI Governance Lifecycle

The structured process following AI system design through use case definition, data assessment, model evaluation, oversight design, controlled deployment, and monitoring.

Historical Bias

A form of bias where an AI system learns from past decisions or outcomes that already reflect unfair treatment or systemic inequities.

Selection, Sampling, and Representation Bias

Bias occurring when dataset samples do not adequately represent the full population or subgroups affected by the AI system.

Measurement Bias

Systematic error created by how data are collected, labeled, or measured, such as utilizing convenient proxies that misrepresent actual needs.

Algorithmic Bias

Unfair outcomes generated when a model learns correlations or proxies for protected traits, even if sensitive variables are excluded.

Interaction Bias

Corruption of AI system behavior over time resulting from exposure to abusive or malicious user interactions.

Automation Bias

The human tendency to over-rely on and rubber-stamp AI outputs because they appear objective or authoritative.

Contextual Bias

Bias that arises when patterns learned in one business context are applied to a different domain where they do not fit.

Data Minimization

The governance practice of collecting and using only the specific data required for a defined purpose to limit privacy and security risks.

Explainability

The capability to provide a meaningful, understandable rationale for an AI-supported output or recommendation.

Responsible AI Risk Map

A framework mapping AI business processes across key risk dimensions including bias, privacy, transparency, accountability, hallucination, human autonomy, IP/labor, and agentic control.

Generative AI

Software built on large foundation models that creates new content—text, images, code, or audio—upon request.

Agentic AI

AI systems that sense context, reason, plan, and execute multi-step tasks autonomously toward a goal across tools and software applications.

Foundation Model

A very large neural network trained on vast unstructured data that can be adapted to perform many diverse downstream tasks.

Hallucination

A primary failure mode of generative AI where the model produces confident, fluent, but factually wrong or unsupported output.

AI Adoption Funnel

The industry dynamic where 88 percent of organizations use AI tools, but only about 6 percent qualify as high performers capturing over 5 percent EBIT impact.

AI Fluency

The learned ability to collaborate with AI effectively, efficiently, ethically, and safely across four core competencies: Delegation, Description, Discernment, and Diligence.

Delegation

The first D of AI fluency; deciding deliberately which work steps to hand to AI and which to keep human.

Description

The second D of AI fluency; communicating a task to AI with sufficient context, constraints, and format requirements.

Discernment

The third D of AI fluency; critically evaluating AI output for factual accuracy, sound reasoning, fit, and bias.

Diligence

The fourth D of AI fluency; practicing responsible AI use through verification, disclosure, and maintaining human accountability.

Human-in-the-Loop (HITL)

An oversight pattern in which a human explicitly reviews and approves every AI output before it takes effect.

Human-on-the-Loop

An oversight pattern where the AI system acts autonomously while a person monitors aggregate operation and intervenes on exceptions.

Human-out-of-the-Loop

Full AI autonomy reserved exclusively for narrow, low-risk, and highly instrumented workflow tasks.

Project Vend

An Anthropic research experiment featuring an autonomous shopkeeper agent ('Claudius') that highlighted agent risks like runaway discounts before strict procedural guardrails were introduced.

Walmart Super Agents

Walmart's architecture of four unified agent entry points—Sparky, associate agent, Marty, and developer agent—built to prevent agent sprawl.