Unit conversion: enforce dimensions before aggregation
A unit conversion is valid only when the quantities have compatible dimensions and scope. Before AI aggregates a table, distinguish scale changes from currency conversion, duration normalization and economic redefinition. Turning thousands into millions is not the same operation as turning a quarterly flow into an annual figure.
The W3C tabular-data model distinguishes a cell’s string representation from its typed value and associated annotations. That structure supports preserving meaning during parsing. It does not supply financial conversion factors or certify that two reported metrics are economically interchangeable.
Attach a dimension record
Store metric, unit, scale, currency, time basis, population and sign convention alongside each value. Preserve the original text and normalized number. A percentage, a currency amount and currency per share can all be numeric while requiring different operations.
Classify the proposed transformation. A scale conversion uses a fixed factor within the same dimension. Currency conversion needs an exchange-rate convention and relevant dates. A time conversion needs a defensible aggregation rule. A scope change cannot be solved by a numeric factor if it changes what is being measured.
Require the output dimension to be explicit before calculating. If the target is revenue in millions of one currency for a fiscal year, each input must be compatible with that currency, duration and reporting perimeter. A matching column label is only a starting clue.
Hypothetical mixed-scale aggregation
Assume fictional Company Birch reports annual revenue of 2,400 in thousands of currency Q. Company Fern reports 3.1 in millions of the same Q under a compatible annual definition. Converting Birch to 2.4 million allows a combined total of 5.5 million Q. Adding 2,400 and 3.1 directly would mix scales.
Now suppose Fern’s 3.1 is currency R, not Q. The 5.5 total is no longer justified without a selected exchange-rate method. If it is a quarterly flow rather than an annual flow, multiplying by four is an annualization assumption, not observed annual revenue.
Finally, assume Birch reports currency Q per share rather than total Q. A scale factor cannot convert it into total revenue without the appropriate share denominator and metric meaning. Each variation changes the permissible operation. All values and company names are hypothetical.
Conversion ledger
| Field | Example treatment |
|---|---|
| Raw representation | 2,400, source label retained |
| Input dimension | Annual revenue, thousand Q |
| Target dimension | Annual revenue, million Q |
| Factor and direction | Multiply by 0.001 |
| Result | 2.4 million Q |
| Compatibility conditions | Same metric, period and reporting scope |
For a chain of conversions, show every intermediate unit. Units should cancel in the intended direction. If a rate is stated as Q per R, multiplying R produces Q; using its inverse without changing the operation produces the wrong dimension. Verify a small illustrative amount manually before processing the full table.
Validate aggregates after conversion
Reconcile subtotals where the source provides compatible checks. Inspect factor outliers such as thousandfold jumps and values that changed sign unexpectedly. A plausible grand total does not prove every row was converted correctly, because opposite errors can offset.
Give AI the approved normalized dataset together with the conversion ledger. Ask it to explain exceptions rather than silently repairing units. If a transformation needs an unsupported assumption, keep that row out of the comparable aggregate and report the exclusion. The final result should name its output unit and scope every time, so a clean number never travels without its meaning.
FAQ
Can quarterly revenue simply be multiplied by four?
Only as a labeled annualization assumption. It does not create observed full-year revenue or resolve seasonality.
Does converting all values to millions make them comparable?
No. Currency, metric definition, period, population and reporting perimeter can still differ.