ai-ledger

“Why This Model?” Approval Checklist

A companion template for Does This Decision Actually Need a Large Language Model?

When an AI use case comes forward for approval, the usual questions are about security, privacy, accuracy, and the vendor. They matter. This checklist adds the one that comes first: why is this the model being proposed?

It complements the model-selection checklist in Model Selection Is an Accounting Control and the onboarding steps in Don’t Just Prompt AI. ONBOARD It.. Use it per step, not per process.

---

Use case / step: [ ] Proposed method: [ ] Rule [ ] Decision model [ ] LLM [ ] Person [ ] Combination

Proposed by: [ ] Reviewed by: [ ] Date: [ ]

1. Does the step need a model at all?

• Is the rule already known? If it can be written as an IF statement, a lookup, or a match, it belongs in deterministic logic, not a model of any kind.
• If a model is proposed for a step with a known rule, the reason is documented: [ ]

2. What shape is the answer?

• Can the possible outcomes be listed in advance? List them: [ ]
• Is there an “other” or “cannot tell” option, so the model is never forced into a wrong box?
• Does the step need text to be written (a summary, a draft, an explanation), or only chosen from the list?
• Does it need extended reasoning across a long document, or a single bounded judgment?

3. Is the proposed model matched to that shape?

• LLM proposed for a bounded choice (e.g. “classify notes into six categories”): why is generation or extended reasoning needed? [ ]
• Decision model proposed for open-ended analysis (e.g. a contract-accounting conclusion): why is a bounded answer enough? [ ]
• Is the model the smallest one that performed acceptably on our work, our data, and our acceptance criteria, rather than the winner of a public benchmark?

4. Evidence and data

• What is the minimum evidence the model needs? Only those fields are sent.
• Where does the model run, what does the provider retain, and does that fit policy? (See If the Data Can’t Leave, Bring the Model In.)
• Do any customer contracts restrict AI, machine learning, automated decision-making, or external processing for this data? (See Your AI Workflow Was Approved. Did Anyone Read the Customer Contract?.) Changing the model type does not change the answer.
• Could any input text be written to steer the answer (customer notes, employee explanations, vendor emails)? How is that contained?

5. Errors and uncertainty

• How will errors be measured? Pilot reference: [ ]
• What does a wrong answer cost? Minor / costly / control failure.
• What happens when confidence is low? (Attach the Confidence Band Routing Policy.)
• If the model returns no rationale, what evidence will the reviewer and the auditor rely on instead? (See From AI Answers to Audit Trails.)
• Has cost been assessed as cost per acceptable decision, including review time and errors, not just model price? (See the cost workbook.)

6. Authority

• Which decisions stay with a person? [ ]
• The model’s output is a classification or a recommendation, not an approval, a posting, or a payment.
• No decision about an individual (credit, employment, a customer’s standing) rests on the model’s score alone.

Decision

[ ] Approved as proposed [ ] Approved with a different method: [ ] [ ] Not approved

Reason: [ ]

---

*PythonMuse LLC

www.pythonmuse.com. Educational purposes only; not professional advice. Content licensed under CC BY-NC-SA 4.0.*

This site is open source. Improve this page.