Marketing determinism scorecard

Score each marketing job on five yes-or-no questions. High score means a rule. Middle means rules around one narrow model step. Low means a model can draft and a person approves.

One row per job. Score it out loud with the person who owns the job. Write the date and the score so you can see what changed after an incident.

The score decides the shape of the workflow. It does not decide whether the job is worth automating. Use the process scorecard for that.

flowchart TD
    J["Job"] --> Q{"Score 4 to 5?"}
    Q -->|Yes| R["Rule or script"]
    Q -->|No| M{"Score 2 to 3?"}
    M -->|Yes| H["Rules + one model step + human approval"]
    M -->|No| D["Model drafts, sampled review"]

The five questions (one point per yes)

Question	Yes means	Point
Is there a written rule a new hire could follow?	The logic exists. Encode it.	1
Would two different outputs on the same input be a bug?	You need repeatability.	1
Does a wrong output touch money, consent, legal exposure, or a customer record?	A quiet error is expensive.	1
Must you explain the decision to a client, auditor, or colleague?	You need a reason you can show.	1
Is the input structured (fields, enums, numbers)?	A model adds risk and no skill.	1

What the score means

Score	Shape	Gate
4 to 5	Deterministic. Rule, table, or script.	Owner and review date for the rule
2 to 3	Hybrid. Rules in front and behind, one narrow model step.	Schema on the model output, human approval on anything that reaches a customer
0 to 1	Model-friendly. Drafting, summarizing, rewording.	Golden set, claims check, weekly sample review

Score your jobs

Job	Q1	Q2	Q3	Q4	Q5	Total	Shape	Date
Write the job name
Write the job name
Write the job name
Write the job name
Write the job name

Jobs that usually score high

Check these first. If one is already run by a model, review it this week.

Lead routing by region, size, or product
Consent and unsubscribe checks before any send
Campaign tags and analytics event names
Send time and frequency caps
Discounts, commissions, and invoice totals
Field mapping between the customer relationship management (CRM) system and email tool

If the score is 2 or 3

Deterministic validation runs before the model step
The model has one narrow job (extract, label from a closed list, or draft)
Model output is validated against a schema. Failures go to a person.
A table or rule makes the decision. The model never picks the owner, the price, or the send.
A consent or approval gate sits between the model and any message
Prompt and context pack version logged with every run

Re-score when

After any incident involving the job
When weekly volume changes by a large factor
When the model, prompt, or platform changes
Once a quarter, even if nothing broke