Automate the task, review the claim

Video transcript

1. 0:00–0:06

   Observe the whole cycle • Separate extraction, calculation and commentary. • Retrieve • Expected records? • Calculate • Reconciled measure? • Distribute • Permitted recipient? • Commentary is a separate optional task

   Three reviewed workstations separate retrieval, defined calculation and distribution. Each has its own evidence question: expected records, reconciled measure and permitted recipient. Commentary is a separate optional task; recurring reporting need not use generative AI.

2. 0:06–0:12

   Validate the foundation • Fluent text can repeat a wrong total. • One source order: O1 = 100 • O1 = 100 + O1 = 100 • Draft repeats 200 • Reconcile by order: 100 • Invented duplicate join · fluent wording cannot fix the sum

   In an invented teaching example, one source order O1 has an amount of 100. A join repeats O1 twice, and a draft repeats the naive total 200. The independent reconciliation gate returns to the one source order and amount 100. Fluent language cannot validate an inflated sum.

3. 0:12–0:18

   Give AI a bounded task • Provide approved definitions and test questions. • Definition · period · source status • Draft + review • Missing source? Disclose the gap. • Approved context does not guarantee every answer

   A bounded context packet contains the approved definition, reporting period and source update status. It enters the reviewed drafting-and-human-review branch. Approved context is not a guarantee of every answer. A missing-source question must disclose the gap under the reporting rule.

4. 0:18–0:24

   Review descriptive and causal claims • The example shows−5%, not its cause. • Prior: SAR 1,000 • Current: SAR 950 • (950 − 1000) / 1000 = −5% • Cause unproven • Invented complete weeks · order amounts, not revenue

   Two invented complete weekly cohorts use the same order definition and currency: prior SAR 1,000 and current SAR 950. The native left-to-right calculation is (950 − 1000) / 1000 = −5%. These are order amounts, not recognised revenue or collected cash. A separate dashed delivery illustration says Cause unproven: the amounts alone do not establish that late deliveries caused the change.

5. 0:24–0:30

   Evaluate the final reviewed task • Include corrections, access and operating effort. • Validated report remains usable • Optional draft fails review • Review + corrections + access + cost • Evaluate the whole task; keep the report path open

   The validated permitted report remains usable while an optional draft fails review. Review, corrections, access and operating cost enter evaluation of the full task. The report route stays available; no time saving, predictive accuracy or successful AI acceptance is claimed.