Directional, Not Absolute AI Rule
Let AI set direction, then apply human judgment to the final 20 percent
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 5
- Confidence
- 96%
The Directional, Not Absolute AI Rule separates dependable analytical assistance from less dependable recommendations. Verify calculations, joins, rankings, and charts against the supplied data, then use AI-generated themes, categories, and ideas as directional inputs rather than finished decisions. Human judgment supplies the final 20 percent: selecting the strongest option, adding context, sharpening originality, and rejecting generic suggestions. This avoids two symmetrical errors—blindly executing AI output and discarding it because it is imperfect. The mechanism is iterative: accept the useful direction, build on it, and apply domain expertise where differentiation matters most. Confidence should therefore vary by output type rather than being assigned uniformly to everything the system produces.
Origin
Extracted from Marketing Against The Grain, where the hosts distinguished Code Interpreter's data handling from its more generic content and guest recommendations.
Core principles
- 01Trust verified computation more than generated taste
- 02Use AI recommendations to establish direction
- 03Reserve final differentiation for human judgment
- 04Build on AI output instead of accepting it unchanged
- 05Match confidence to the type of output
How to run it
- 1
Classify the output
Separate verifiable computation from interpretive recommendations, creative ideas, and judgments.
Pro tip Mark each claim as calculated, inferred, or generated.
Watch out A confident tone does not make a recommendation factual.
- 2
Verify the analytical core
Check the input data, transformations, and calculated results before using them as the basis for decisions.
Pro tip Request intermediate tables or code when available.
Watch out Recommendations built on faulty data remain faulty even when they sound plausible.
- 3
Extract the direction
Identify useful themes, categories, commonalities, or opportunity areas in the AI's recommendations.
Pro tip Look for signals that narrow the search space rather than expecting a final answer.
Watch out Do not mistake averaged best practices for distinctive strategy.
- 4
Apply the final 20 percent
Use expertise, taste, context, and originality to select and improve the strongest possibilities.
Pro tip Ask what would make the recommendation uniquely appropriate for this audience or business.
Watch out Copying the draft unchanged usually preserves its generic qualities.
- 5
Test and iterate
Run a bounded experiment, observe the result, and feed the evidence into the next decision cycle.
Pro tip Define a measurable success criterion before executing.
Watch out Directional guidance should not be treated as certainty after only one test.
In the wild
Code Interpreter used performance data to suggest broad themes such as AI experts, entrepreneurs, and marketing strategists. The hosts treated those categories as useful direction but recognized that they still needed to choose compelling guests and create sharper titles.
→ AI narrowed the opportunity space while the hosts retained responsibility for editorial quality.
An AI identifies three campaign territories supported by customer data. A marketing lead verifies the analysis, rejects the most generic territory, and reshapes another around a customer tension the model overlooked.
→ The team gains analytical speed without surrendering differentiation or judgment.
Common mistakes
Treating suggestions as commands
Generated recommendations may be plausible but generic. They should inform judgment rather than replace it.
Distrusting every output equally
Verified calculations and speculative creative suggestions carry different levels of reliability and should be evaluated separately.
Skipping human refinement
The final layer of context, taste, and specificity is often what turns an adequate recommendation into a valuable one.
Is it for you?
Best for
It is best for analytical and creative tasks where AI can accelerate preparation but human taste determines the final quality.
Not ideal for
It is not ideal for fully deterministic tasks that can be validated automatically or high-stakes decisions requiring formal expert review.
From the transcript
“it is directional, not absolute.”
“But you're gonna have to take your brain and apply the last 20% to actually make them awesome.”
“Okay, well how can I build on what the AI has given me," that is gonna make speed of thought, speed of iteration, speed of…”
From the episode
ChatGPT’s Code Interpreter: Your $20 Personal Data Analyst (#141)