Insert AI as a transformation step between workflow stages — provide output + instruction, never delegate decisions

Why This Is a Rule

AI excels at transformation: converting data between formats, summarizing long documents, extracting structured information from unstructured text, generating drafts from outlines. AI struggles with decisions: choosing between options based on values, determining whether output meets your quality standards, judging whether something is "good enough." When AI is inserted into composed workflows, the integration point must be structured to leverage the transformation strength while maintaining human sovereignty over decisions.

The integration pattern is: provide the previous stage's concrete output (the intermediate artifact from Preserve intermediate outputs as concrete artifacts at every workflow boundary — enables resumption, failure recovery, and reuse) plus a specific transformation instruction ("Summarize these research notes into 5 key findings," "Convert this outline into a first draft," "Extract all dates and deadlines from this email thread"). The AI transforms input A into output B according to your specification. You then review output B before it enters the next stage — the human checkpoint that prevents automation from becoming abdication (Sovereignty check before automation — if the automation produces wrong output, will you notice? Automate only when the answer is yes).

This pattern treats AI as a sophisticated macro, not an autonomous agent. The macro receives inputs and instructions, produces outputs, and has no authority to decide what to do next, whether the output is acceptable, or whether the workflow should proceed. Those decisions remain with the human operator. This preserves the composability of the workflow (Test composability by replacing one sub-workflow without breaking others — forced changes to adjacent workflows mean interfaces are leaking) because the AI step has clean interfaces: defined input, defined transformation, defined output.

When This Fires

• When designing workflows that include AI-assisted steps
• When deciding where in a workflow to introduce AI tools
• When AI steps are producing unpredictable or uncontrollable outputs (likely treating AI as a decision-maker)
• Complements Sovereignty check before automation — if the automation produces wrong output, will you notice? Automate only when the answer is yes (sovereignty check) and Preserve intermediate outputs as concrete artifacts at every workflow boundary — enables resumption, failure recovery, and reuse (boundary artifacts) with AI-specific integration guidance

Common Failure Mode

Delegating decisions to AI: "AI, review this draft and decide if it's ready to send." The AI will always produce an answer, but the answer requires judgment the AI doesn't have — your relationship context, your quality standards, your strategic intent. The draft gets sent based on the AI's assessment, and the AI missed something only you would catch.

The Protocol

(1) Identify which workflow steps are transformations (format conversion, summarization, extraction, drafting) vs. decisions (quality assessment, approval, strategic choices). (2) AI is eligible for transformation steps only. (3) For each AI step, structure the prompt as: "[Previous stage output] + [Specific transformation instruction] + [Output format specification]." (4) The AI's output becomes an intermediate artifact (Preserve intermediate outputs as concrete artifacts at every workflow boundary — enables resumption, failure recovery, and reuse) that you review before it enters the next stage. (5) Never chain AI steps without human review between them — each AI output must be verified before becoming the next AI step's input, to prevent error compounding.