Putting It Into Practice
The 90-Day Executive AI Learning Plan
This plan sequences the topic catalog into a working study schedule. It assumes 3-5 hours per week: enough to read each week's topic entries closely (they are already written at executive length, roughly 10-15 minutes each), go deeper on one or two primary sources, run a small experiment, and sit with one hard question. Topics are referenced by their catalog number (Domain.Topic); go back to the full entry in the catalog for definitions, sources, and the complete list of questions to ask your team.
Treat the week numbers as a sequence, not a calendar. If Week 9 (Security) needs eight days instead of seven, take them: the plan is a priority order, not a countdown.
PHASE 1: FOUNDATIONS (Days 1-30 / Weeks 1-4)
Goal: build the shared vocabulary and mental models everything else depends on. By the end of this phase you should be able to sit in a vendor pitch or an architecture review and know which questions are the load-bearing ones.
Week 1: Strategy and the Operating Model
Topics: 1.1 AI Operating Models (Centralized/Federated/Hub-and-Spoke) · 1.2 AI Portfolio Management & Build vs Buy vs Configure · 1.3 Multi-Model Strategy & Vendor Dependency Recommended Reading: DORA AI Capabilities Model report (cited under 1.1): the single best evidence-based framework for what actually makes AI investment pay off. Experiment: List your organization's current AI initiatives on one page. For each, mark who owns the outcome and whether it has a kill date. If you can't fill in both columns for most rows, you've found your first governance gap. Executive Question: If I shut down every current AI pilot tomorrow, which ones would a business unit leader personally fight to keep, and why don't I already know the answer to that?
Week 2: What an Agent Actually Is
Topics: 2.1 What Is an AI Agent · 2.2 Agent Architectures · 2.6 Human Approval & Autonomy Levels · 2.7 Where Agents Beat vs Lose to Deterministic Software Recommended Reading: Anthropic's "Building Effective Agents" engineering post (cited under 2.2): the clearest primary-source explanation of when orchestration adds value versus overhead. Experiment: Pick one workflow your team automated with an agent (or wants to). Walk through it step by step and ask at each step: could this have been deterministic code instead? If the honest answer is yes for most steps, you have an agent-washing problem, not an AI strategy. Executive Question: For every agent we're running or planning, who has explicitly signed off on what level of autonomy it has, and can they show me that decision in writing?
Week 3: Interoperability and the Model Layer
Topics: 3.1 Model Context Protocol (MCP): Architecture & Enterprise Adoption · 3.2 MCP Security & Governance · 9.1 Frontier, Reasoning & Multimodal Models · 9.2 Open-Weight vs Proprietary Models Recommended Reading: The MCP first-anniversary and 2026 roadmap posts (cited under 3.1) for where the protocol is headed under Linux Foundation governance; skim one frontier-model announcement (cited under 9.1) to calibrate current capability. Experiment: Ask your platform team for a list of every MCP server currently connected to any internal system: internal or third-party. If nobody can produce that list in a day, you have a shadow-integration problem before you have an AI problem. Executive Question: Are we building a dependency on one model vendor's ecosystem, and if that vendor's pricing or terms changed tomorrow, what would it cost us to move?
Week 4: How Work Changes: Software and Knowledge Work
Topics: 4.1 The Software Engineering Continuum · 6.1 The Maturity Ladder (knowledge work) · 10.1 Prompt Engineering Fundamentals · 10.2 Context Engineering & System Architecture Around the Model Recommended Reading: DORA's 2025 State of AI-assisted Software Development report (cited under 4.1): read the velocity/stability tension section closely; it's the most important single finding in this document for a large engineering organization. Experiment: Ask five engineers and five knowledge workers, independently, where they'd place their own daily AI use on the two continuums in Topics 4.1 and 6.1. The spread in answers tells you more about your real adoption maturity than any dashboard. Executive Question: Is our organization optimizing for developers who feel faster, or for measurable delivery stability and business outcomes, and do we actually know if those are currently the same thing?
PHASE 2: ARCHITECTURE & TRANSFORMATION (Days 31-60 / Weeks 5-9)
Goal: go deep enough on the technical architecture, security posture, and workflow redesign questions that you can evaluate your own team's and vendors' proposals rather than rubber-stamping them.
Week 5: Agent Architecture, Deep Dive
Topics: 2.3 Long-Running, Computer-Use & Browser Agents · 2.4 Agent Memory & Persistent State · 2.5 Agent Sandboxes, Observability, Rollback & Idempotency · 4.2 Coding Agents & Autonomous Coding Agents · 4.3 Specification-Driven Development & Context Engineering for Code Recommended Reading: The OSWorld 2.0 benchmark paper (cited under 2.3) for a realistic read on current computer-use agent reliability: useful ammunition against vendor demos that only show the happy path. Experiment: Ask your engineering leadership what percentage of coding-agent output currently merges without human review versus with review. If nobody has that number, that's the finding. Executive Question: If one of our production agents took an irreversible action based on a bad decision, could we reconstruct exactly what it did and why, today, in under an hour?
Week 6: Software Engineering Operations and Workflow Redesign
Topics: 4.4 AI Code Review, Testing & Debugging · 4.6 Parallel Coding Agents, Supervision Ratios & Software Factories · 4.7 CI/CD & Quality Gates in an Agentic Development Model · 4.8 AI-Generated Code Risk & Secure Development Practices · 7.1 RPA vs BPM vs API Automation vs Agentic Workflows Recommended Reading: GitClear's code-quality research and the arXiv paper on security degradation in iterative AI code generation (both cited under 4.8): read these back to back; they're the sobering counterweight to velocity claims. Experiment: Pick one business process currently run as RPA or manual work. Have your team sketch what an agentic redesign would look like and where a deterministic step should stay deterministic on purpose. Executive Question: Are our quality gates (code review depth, test coverage requirements, architecture review) the same as they were two years ago, even though the volume and origin of code changed completely?
Week 7: Enterprise Knowledge and Data Architecture
Topics: 8.1 RAG, Embeddings & Vector Databases · 8.2 Hybrid Retrieval, Knowledge Graphs & GraphRAG (Monitor) · 8.3 Permissions-Aware Retrieval & Access Control Propagation · 8.4 Data Classification, Residency, Freshness & Provenance · 8.5 Enterprise, Personal & Agent Memory Architecture Recommended Reading: Anthropic's Contextual Retrieval post and the OWASP RAG Security Cheat Sheet (both cited under 8.1/8.3): the second one is the one to hand your CISO. Experiment: Ask your data team to demonstrate, live, that a search assistant correctly refuses to surface a document a test user isn't permissioned to see. If they can't demo it on request, don't trust that it works in production. Executive Question: When our RAG system is wrong, can we trace the answer back to the specific document it came from, and would that document have been visible to the user through normal permissions anyway?
Week 8: AI Platform Architecture and Evaluation
Topics: 11.1 AI Gateways & Model Gateways · 11.2 Centralized Policy Enforcement · 11.3 Gateway Architecture Choices (Monitor) · 12.1 AI Evals & Golden Datasets · 12.3 Hallucination, Groundedness & LLM-as-Judge · 12.4 Red Teaming & Continuous Production Evaluation Recommended Reading: OWASP Top 10 for LLM Applications, 2025 edition (cited under 12.4 and 13.1: read the evaluation-relevant items now, the security ones in Week 9). Experiment: Ask what percentage of your production AI features have a golden dataset and an automated eval that runs before every deploy. Most organizations at your scale will answer close to zero: that's the baseline you're trying to move. Executive Question: If a model provider silently updated the model behind one of our production features tomorrow, would we detect a quality regression before a customer or employee did?
Week 9: Security and Identity (this week runs long, budget extra time)
Topics: 13.1 OWASP Top 10 for LLM Applications · 13.2 Prompt Injection · 13.3 Excessive Agency & Goal Hijacking · 13.4 Memory/RAG Poisoning · 13.5 Insecure Output Handling · 13.6 MCP & Tool/Plugin Supply-Chain Risk · 13.7 Zero-Trust Architecture for Agents · 14.1 Non-Human Identity & Workload Identity · 14.2 Delegated Authorization & JIT Access · 14.3 Least Privilege, Transaction Limits & Audit Attribution Recommended Reading: Simon Willison's "lethal trifecta" post (cited under 13.2) and the Invariant Labs GitHub MCP exploit writeup (cited under 13.6): both are short, concrete, and worth reading in full rather than summarized. Experiment: Ask your security team the single framing question that ties this whole domain together: "who authorized this AI to take this action, with this data, using these systems, at this time?": for your three highest-privilege agents. Silence or hand-waving is the finding. Executive Question: Do we have a non-human identity for every agent with write access to a production system, or are agents currently sharing service-account credentials with humans?
PHASE 3: SCALING THE ENTERPRISE (Days 61-90 / Weeks 10-13)
Goal: shift from "can we build this safely" to "are we running this as a managed capability": governance, workforce, measurement, and where to place your remaining strategic bets.
Week 10: Workforce Transformation and Knowledge Work at Scale
Topics: 5.1 Which Engineering Skills Appreciate vs Commoditize · 5.2 The Junior Engineer Pipeline Problem · 5.3 Changing Roles, Team Topology & Hiring Frameworks · 6.2 Enterprise AI Assistants & Enterprise Search · 6.3 Deep Research Agents & Document Generation (Monitor) · 6.4 Meeting Intelligence & Departmental Digital Workers Recommended Reading: Whatever your own engineering leadership can show you on junior-hire acceptance rates and time-to-productivity over the last two years (cited under 5.2): this is the topic where your own data matters more than any external source. Experiment: Ask how many entry-level engineering or analyst roles you've opened in the last 12 months versus three years ago. Then ask who decided the trend line, and whether anyone modeled what it does to your leadership pipeline in five years. Executive Question: If AI tools make experienced people more productive but reduce the number of junior roles that used to train the next generation of experienced people, who is accountable for closing that gap ten years out?
Week 11: Governance, Legal, and Compliance
Topics: 15.1 NIST AI Risk Management Framework & Generative AI Profile · 15.2 ISO/IEC 42001 · 15.3 EU AI Act · 15.4 AI, Intellectual Property & Copyright Risk · 15.5 AI System Inventory & Risk Classification Recommended Reading: NIST's AI RMF Playbook and the EU AI Act Service Desk implementation timeline (both cited under 15.1/15.3): the second one matters even if you have no EU operations today, because vendor compliance postures are converging on it. Executive Question: Do we have a single inventory of every AI system in production, who owns it, and what risk tier it's classified at, or would building that list today take weeks because it doesn't exist?
Week 12: Economics, Observability, and AI-Native Product Design
Topics: 16.1 Token Economics & Cost-Per-Outcome · 16.2 Cost Controls (routing, caching, substitution) · 16.3 Chargeback, Showback & Runaway Cost Risk (Monitor) · 17.1 LLMOps/AgentOps & Tracing Agent Trajectories · 17.2 Incident Response & Audit Replay · 18.1 Beyond the Chatbot · 18.2 Progressive Autonomy & Approval UX Recommended Reading: The FinOps Foundation's "Token Economics" material (cited under 16.1) and Nielsen Norman Group's "AI Agents as Users" research (cited under 18.1): the second one will change how you brief your product teams. Experiment: Pull last month's aggregate model spend and ask your finance and platform teams to reconcile it against actual business outcomes, not usage volume. Expect this reconciliation to be harder than it should be: that gap is the finding. Executive Question: If one team's agent started running away (retrying, looping, or fanning out unexpectedly) would our cost controls catch it before finance did, three weeks later, in a monthly bill?
Week 13: Measurement, Adoption, and the Technology Radar
Topics: 19.1 Executive & Role-Based AI Literacy · 19.2 AI Champions & Communities of Practice · 19.3 Employee Displacement Fear & Change Management · 20.1 Engineering Metrics in the AI Era · 20.2 Knowledge-Worker Productivity Metrics · 20.3 Enterprise Value Metrics · 20.4 Why Prompt/Seat Counts Are Vanity Metrics · Domain 21, full read-through (Emerging Technology Radar) Recommended Reading: The NBER working paper on generative AI productivity gains by skill segment (cited under 20.2) (the single best corrective to flat "AI made us X% more productive" claims) and a full pass through the Executive Technology Radar table in this document. Experiment: Take your current AI dashboard, whatever it is, and for every metric on it ask: is this a proxy for an outcome, or is it a vanity metric dressed as one (seats provisioned, prompts sent, "AI-assisted" tickets closed)? Cut anything that fails that test. Executive Question: If I had to justify our entire AI investment to the board using only outcome metrics (revenue enabled, cost reduced, capacity released) and none of the adoption or activity metrics, could I do it convincingly today?
After Day 90
This plan gets you to fluency, not to done. Three habits carry it forward: revisit the Executive Technology Radar table quarterly and move items between Monitor, Experiment, and Adopt as evidence changes; re-score your organization against the Enterprise AI Capability Model every two quarters: maturity here moves in months, not years; and treat any Monitor-tier topic in the catalog (3.3, 4.5, 6.3, 9.5, 11.3, 12.2, 13.4, 13.5, 14.2, 15.2, 16.3, 18.3) as a standing item to re-read once it starts showing up in vendor pitches or team escalations, rather than something to schedule proactively.