Utility bills to reviewed structured data
Parallel extraction paths with discrepancy review for varied utility bills.
My contribution
Developed a utility-bill extraction and review workflow for energy assessments.
Handle format variation before analysis
The ITAC-LSU Bill Analyzer detects provider formats, extracts bill fields through text/OCR and image paths, compares the outputs, and routes discrepancies to human review. Structured output supports later energy analysis. Agreement between two extraction paths is a review signal, not ground truth by itself.
What the dataset establishes
The manuscript describes 440 bills: 300 digital and 140 scanned. Provider 1 contributes 230 documents, while three other providers contribute 70 each. The digital/scanned composition is a reportable dataset characteristic.
The manuscript reports 100% extraction accuracy for digital documents. Its exact field-level evaluation definition and denominator still require confirmation before this becomes a public metric chart. Table 2 agreement rates are illustrative, and Table 3 scanned-performance values are estimated or illustrative. Neither supplies a measured scanned-document benchmark.
Review before downstream calculations
Provider detection, date ranges, consumption, demand, costs, and repeated line items need inspectable output. A missing or inconsistent field should reach a reviewer before a downstream energy calculation uses it. Public examples need an approved synthetic or redacted bill. Source bills and private client reports stay outside the public site.
System architecture
Select a step, or a block in the drawing, to see its input, computation, output and source.
Two extraction paths, compared before use
Bills are matched to a provider format, then extracted through a text path and an image path. Disagreements go to a person before the data feeds any energy calculation.
- Extraction
- Comparison
- Human resolution
Bill intakeInput
- Input
- Digital or scanned utility bills.
- Computation
- None.
- Output
- Bill documents.
- Source
- Utility-bill manuscript, Figure 1
Provider detectionProcessing step
- Input
- Bill document.
- Computation
- Detects the provider format.
- Output
- Provider-specific handling.
- Source
- Utility-bill manuscript, Figure 1
Text or OCR extractionProcessing step
- Input
- Bill document.
- Computation
- Extracts fields from text, using OCR where needed.
- Output
- Field values from the text path.
- Source
- Utility-bill manuscript, Figure 1
Image-based extractionProcessing step
- Input
- Bill page images.
- Computation
- Extracts fields from the page image.
- Output
- Field values from the image path.
- Source
- Utility-bill manuscript, Figure 1
Discrepancy comparisonDeterministic
- Input
- Field values from both paths.
- Computation
- Compares the two outputs field by field. Agreement is a review signal, not ground truth.
- Output
- Agreed fields and discrepancies.
- Source
- Utility-bill manuscript, Figure 1
Human resolutionHuman review
- Input
- Discrepant fields.
- Computation
- A reviewer resolves each discrepancy.
- Output
- Resolved field values.
- Source
- Utility-bill manuscript, Figure 1
Structured outputOutput
- Input
- Agreed and resolved fields.
- Computation
- Structured record for later energy analysis.
- Output
- Structured bill data.
- Source
- Utility-bill manuscript, Figure 1
Adapted from Automated Utility Bill Data Extraction and Analysis for Energy Management, Table 1; Table 2 qualification; Table 3 qualification; evaluation text.
Which branches run for digital versus scanned bills still needs confirmation, so both are drawn. Scanned-document values in the manuscript are illustrative and are not shown.
Animated pulses show conceptual signal flow only. They do not show measured execution time or a live inference run.
Findings
Reported results, grouped by sample. Each figure keeps its own scale and links to its source and method.
Utility-bill dataset
440 bills. 300 digital, 140 scanned; provider counts 230, 70, 70, 70.
Document format composition
Reported in sourceLargest reported count: Digital, 300.
Data table
| Condition | Value |
|---|---|
| Digital | 300 |
| Scanned | 140 |
Denominator: 440 bills in Table 1.
Adapted from Automated Utility Bill Data Extraction and Analysis for Energy Management, Table 1; Table 2 qualification; Table 3 qualification; evaluation text.
Provider composition
Reported in sourceLargest reported count: Provider 1, 230.
Data table
| Condition | Value |
|---|---|
| Provider 1 | 230 |
| Provider 2 | 70 |
| Provider 3 | 70 |
| Provider 4 | 70 |
Denominator: 440 bills in Table 1; provider labels anonymized by the manuscript.
Adapted from Automated Utility Bill Data Extraction and Analysis for Energy Management, Table 1; Table 2 qualification; Table 3 qualification; evaluation text.
Evaluation Scope & Methodology
Samples, denominators, uncertainty, sources and limits
Approach
Publish dataset composition separately from extraction claims; exclude illustrative agreement and scanned accuracy values from empirical charts.
Cohorts and samples
| Cohort | Sample | Definition |
|---|---|---|
| Utility-bill dataset | 440 bills | 300 digital, 140 scanned; provider counts 230, 70, 70, 70. |
Reported measures
| Measure | Unit | Cohort | Denominator | Uncertainty | Evidence |
|---|---|---|---|---|---|
| Document format composition | Count | Utility-bill dataset | 440 bills in Table 1 | None reported | Reported in source |
| Provider composition | Count | Utility-bill dataset | 440 bills in Table 1; provider labels anonymized by the manuscript | None reported | Reported in source |
Limitations
- Scanned-document performance in Table 3 is estimated or illustrative, not a measured result.
- Pre-resolution agreement rates in Table 2 are explicitly illustrative.
- Measured scanned-field ground truth and extraction logs are not supplied.
- The digital result needs a verified field-level denominator before a public accuracy chart.
- No measured time-savings result is established.
Sources
- Automated Utility Bill Data Extraction and Analysis for Energy Management, Table 1; Table 2 qualification; Table 3 qualification; evaluation text (Supplied manuscript reviewed November 2026). Public release pending