AI decision record: evaluate and document a trial
Choose a tool using evidence from your task. This record connects a specific output, observed results, human cost and the ability to switch solutions. It does not assign performance ratings to providers.

How to build a useful record
- Before testing, state a format and observable threshold: for example, extract ten amounts without errors and retain page references. An example is not a universal standard.
- Give candidates the same permitted inputs. Record version, plan, settings and date. Keep rejected outputs too: they reveal failure conditions.
- A blocking error requires correction or rejection even if other checks pass. Minor successes cannot compensate for an incorrect essential fact.
- Then decide: adopt, retest or stop. State retest triggers such as model, data or integration changes.
Seven questions to settle
- Expected output and audience
- What observable result will be delivered, to whom and in which format?
- Tested tool, model and plan
- Which tool, version and plan actually match the trial?
- Acceptance criteria and blocking errors
- Which defects require rejection even if other results are good?
- Permitted data and privacy constraints
- Which information is permitted, and which must be removed?
- Evidence: input and output references, errors and corrections
- Can each conclusion be traced to an input, output and check?
- Tested export and replacement option
- Can another solution continue the work using the available formats?
- Decision, owner and retest trigger
- Who approves the decision, and which change triggers a new evaluation?







Fictional example: extract invoices
- Criterion
- Amount, currency and invoice number match each document; missing tax stays unknown.
- Blocking error
- An invented total or currency makes the result unusable, even with correct formatting.
- Evidence to retain
- Fictional input, raw output, human correction, time and page reference for each amount.
- Reversibility
- Open the export in another tool and check columns, units and document links.




Complete the decision
Example: a summary produced from an incomplete document
In this fictional case, two of three pages are readable and the remaining page was not converted. The summary looks correct for the recovered pages. The decision depends on coverage checks, not only writing quality.
- Retest
- The system flags the missing page and avoids conclusions about the whole document. Recover that page, then repeat the same check.
- Reject the output
- The system invents missing content or presents a subtotal as the full document total. That error blocks use of the output.
- Accept with conditions
- All pages have been recovered, critical values reviewed and remaining limitations fit the need established before testing.


