MMarketing Against The Grain
← All frameworks
Strategy

Model Council Decision Review

Compare multiple models against first-party data before making a decision

Difficulty
Easy
Time to result
~days to results
Steps
5
Confidence
98%

Model Council Decision Review sends the same business question and relevant evidence to multiple AI models, then structures their responses into areas of agreement, disagreement, and model-specific insight. The method becomes materially stronger when the models can inspect first-party systems such as a CRM rather than relying only on generic third-party information. A human reviews both the recommendation and its supporting evidence, paying particular attention to concentrated risks, inconsistent assumptions, and distinctive analytical lenses. The output is not a majority vote: consensus offers confidence, disagreement reveals uncertainty, and unique findings expose considerations that another model may have missed. This produces a compact decision brief that is easier to challenge than a single opaque answer.

Origin

Nate Folin described using Perplexity Computer's model council to choose an executive-dinner city from open CRM opportunities on Marketing Against The Grain.

Core principles

  • 01Do not trust a single model for consequential work
  • 02Ground recommendations in company-specific data
  • 03Separate consensus from disagreement and unique findings
  • 04Make the reasoning easy for a human to review

How to run it

  1. 1

    Frame the decision

    State the specific choice to be made and the practical outcome required. Define any constraints that the models must respect.

    Pro tip Ask for a simple review format before analysis begins.

    Watch out A vague question will produce conclusions that are difficult to compare.

  2. 2

    Supply first-party evidence

    Connect the systems containing the organization's actual customers, opportunities, performance, or operational constraints.

    Pro tip Prioritize current CRM and internal data over generic market assumptions.

    Watch out Third-party data alone may produce a plausible but irrelevant recommendation.

  3. 3

    Convene multiple models

    Have several current models independently analyze the same question and evidence.

    Pro tip Use models with different strengths to increase analytical diversity.

    Watch out Do not treat repeated wording as independent corroboration if the models share the same unsupported assumption.

  4. 4

    Map agreement and disagreement

    Organize the output into consensus findings, disputed findings, and insights unique to individual models.

    Pro tip Request the evidence supporting every important conclusion.

    Watch out Consensus can still be wrong when all models inherit poor data.

  5. 5

    Review and decide

    Challenge the logic, correct weak assumptions, and select or revise the recommendation.

    Pro tip Record why the final choice differs from the council when overriding it.

    Watch out Do not automate execution until the decision logic has been inspected.

In the wild

Choose an executive-dinner city

A RevOps lead asks several models to inspect open CRM opportunities and recommend a city for an executive dinner. The council identifies common target criteria, flags that Chicago's pipeline is concentrated under one representative, and presents evidence for each candidate city.

The team chooses a city using account quality and concentration risk rather than generic market size.

Common mistakes

Using generic data only

A model may recommend a theoretically attractive market that has little relationship to the company's actual pipeline or customers.

Treating consensus as proof

Agreement is a confidence signal, not a substitute for checking evidence and assumptions.

Ignoring unique findings

Model-specific observations can reveal risks or opportunities concealed by the consensus summary.

Is it for you?

Best for

Teams making thoughtful, data-dependent decisions such as choosing markets, audiences, campaigns, or investments.

Not ideal for

It is unnecessary for low-risk mechanical tasks with a single objectively verifiable answer.

From the transcript

But I also uh like to not trust one model.

Nate Folin · 04:30

You can have this in any format you'd like, broken down by where the models agree, disagree, and unique findings for each of those.

Nate Folin · 06:00

I think that's super important because if you just go third-party data or what what is the data say, it's it's less relevant to your…

Nate Folin · 06:00

From the episode

How 1 Human + AI Replaced a 15-Person RevOps Team