Session 2.3: Where AI Sits in the Pipeline (And Where It Doesn't)

Course → Module 2: AI as Infrastructure, Not Magic

Session 3 of 5

AI is excellent at some tasks and terrible at others. The mistake most people make is treating AI as a general-purpose tool that can handle any part of the content pipeline. It cannot. Mapping AI capabilities to specific pipeline stages gives you a clear picture of where it adds value and where it destroys it.

What AI Does Well

AI excels at tasks that involve transforming structured inputs into prose, processing in bulk, and operating within well-defined constraints. These tasks share a common trait: the human has already made the important decisions.

Task	Why AI Excels	Pipeline Stage
Expanding bullet points into paragraphs	Structured input, clear transformation target	Generation
Generating variations (5 headlines from 1)	Combinatorial task, no judgment needed	Generation
Reformatting content (article to email)	Mechanical transformation with clear rules	Formatting
Summarizing long documents	Compression with known source material	Research
Translating between languages	Pattern matching with high training data	Formatting
Drafting from detailed outlines	Structure provided, only prose generation needed	Generation
Processing bulk content (100 product descriptions)	Repetitive task with consistent format	Generation

What AI Does Poorly

AI fails at tasks that require judgment, experience, originality, or accountability. These tasks require inputs that do not exist in training data: your expertise, your audience's needs, your editorial standards.

Task	Why AI Fails	Pipeline Stage
Deciding what is worth writing about	Requires market knowledge, audience insight, strategy	Planning
Knowing what your audience actually needs	Requires relationship with audience, domain expertise	Research
Having opinions	Opinions require experience and values AI does not have	All stages
Verifying facts	Cannot access real-time data, hallucinates citations	Review
Judging when content is done	No quality standard, no sense of "enough"	Review
Understanding context and timing	No awareness of current events, industry shifts	Planning
Accountability	Cannot stand behind claims or accept consequences	Publishing

AI is a text-generation engine. It generates. It does not decide, verify, judge, or take responsibility. Those are your jobs.

The Pipeline Map

The diagram below shows a complete content pipeline with AI involvement marked at each stage. Green stages are where AI adds value. Red stages are where AI must not operate unsupervised.

graph TD A["Planning
HUMAN: What to write, why, for whom"] --> B["Research
AI assists: summarize sources
HUMAN: select sources, verify"] B --> C["Specification
HUMAN: Define structure, voice,
requirements"] C --> D["Generation
AI: Produce text within spec"] D --> E["Review
HUMAN: Check against spec,
verify facts"] E -->|"Fail"| D E -->|"Pass"| F["Editing
HUMAN primary, AI assists
with compression"] F --> G["Formatting
AI: Convert between formats"] G --> H["Publishing
HUMAN: Final approval,
upload, verify live"] style A fill:#2a2a28,stroke:#c47a5a style C fill:#2a2a28,stroke:#c47a5a style E fill:#2a2a28,stroke:#c47a5a style H fill:#2a2a28,stroke:#c47a5a style D fill:#2a2a28,stroke:#6b8f71 style G fill:#2a2a28,stroke:#6b8f71

The Danger Zones

Three pipeline stages are particularly dangerous to hand to AI without oversight.

Planning. When AI decides what to write, it produces content that matches the statistical average of what already exists on the internet. It does not identify gaps, opportunities, or audience needs. It does not have a content strategy. Letting AI plan your content calendar is equivalent to publishing whatever everyone else has already published.

Fact verification. AI cannot verify facts. It generates text that looks like it contains verified information, but the verification is cosmetic. "According to a 2024 study published in the Journal of Marketing..." may refer to a study that does not exist. AI does not know the difference between a real citation and a plausible-sounding fabrication.

Final approval. Publishing is an act of accountability. When you publish something, you are telling your audience: "I stand behind this." AI cannot stand behind anything. The final decision to publish must be a human decision, made by someone who has reviewed the content and is willing to attach their name to it.

The Two-Column Test

For your specific content type, create a two-column document. Column one: "AI Does This." Column two: "I Do This." List every task in your production process and assign it to the correct column. Be honest. If you have been letting AI do something it should not be doing, that is the first thing to fix.

AI Does This	I Do This
Generate draft text from my outline	Create the outline based on my expertise
Produce headline variations	Select the headline that fits my audience
Summarize research sources	Choose which sources are credible and relevant
Reformat content for different platforms	Decide which platforms and how to adapt the message
Generate first-pass metadata (descriptions, tags)	Review and approve all metadata before publishing

The pattern is consistent: AI handles execution. You handle judgment. When this division is maintained, AI is a force multiplier. When it breaks down, when AI starts making judgment calls, quality collapses.

Assignment

Create a two-column document: "AI Does This" and "I Do This."
List every task in your content production process and assign it to the correct column. Be specific. Not "writing" but "generating first draft from outline" vs. "creating the outline."
Review your column assignments honestly. Are there tasks in the "AI Does This" column that should be in the "I Do This" column? Mark them in red.
Write a one-paragraph action plan: what will you move from AI's column to yours, and why?

Where AI Sits in the Pipeline (And Where It Doesn't)