How to use this order
Work from the top and stop at the step that explains your problem. Each step names a check you can ask for and the outcome that covers it. None of them replaces a person who knows your data, and this is an authored order, not a vendor procedure.
Steps one and two: get data in, then match it
Step 1, reading documents: invoices, receipts, paperwork dates and order emails become rows or an exceptions list, tested on items you labelled. Supplier contract dates have their own job in the finance area. Step 2, matching and filling: link company or product records with software rules first and a model for the unclear pairs, and send look-alikes to an unsure list. Never let a tool invent an identifier.
- Step 1: invoices, receipts, contract dates (finance job), paperwork dates, order emails.
- Step 2: company names, old and new codes, missing attributes, tags.
Steps three and four: define the numbers, then watch the data
Step 3, definitions: write each metric once as code, and prove it against control totals. A semantic layer answers only what it models and says so otherwise. Step 4, watching: data checks shown to fire on planted failures, and alerts that name where a number moved, not why.
- Step 3: metrics, data dictionary, lineage for one report, saved questions, a question tool, workbook logic, warehouse AI cost.
- Step 4: data checks, metric-change alerts, supplier file changes, failed jobs.
Steps five and six: report and forecast, then keep AI steps honest
Step 5, reporting: commentary where every figure is traced to a table, and forecasts that must beat a seasonal baseline or say they do not. Step 6, AI steps in production: a golden set that fails on planted regressions, scored each month with fresh rows.
- Step 5: report commentary (one-off and monthly), forecasts.
- Step 6: golden-set test, monthly drift check.
Keeping it going
If you want these checks to continue without buying each one, the data-operations lane bundles up to two monthly services and up to two bounded jobs a month. Personal data needs extra care at every step: ask for the personal-data finder and masking jobs first, and remember neither can promise to find everything.
Sources and limits
- dbt: semantic layer vs text-to-SQL, 2026 Checked 2026-10-12.
- The semantic layer answers only questions that fall within what has been modelled, and returns an error rather than a wrong number when it cannot answer.
- Splink Checked 2026-10-12.
- Splink is a Python package for probabilistic record linkage that needs no training data, and its input should be standardised with several columns that are not highly correlated.
- Forecasting: Principles and Practice, time series cross-validation Checked 2026-10-12.
- Forecast methods are compared with mean, naive, seasonal naive and drift benchmarks, and a method that does not improve on them is not worth considering.
- Microsoft Presidio Checked 2026-10-12.
- Presidio's page says there is no guarantee that it will find all sensitive information.