Map the queue first
| Observed constraint | Possible tool | What to measure |
|---|---|---|
| Manual intake and retyping | OCR and classification | Correction time and missed fields |
| Large documentary records | Extraction and issue maps | Omissions, contradictions and review time |
| Routine draft preparation | Template and draft assistance | Unsupported statements and time to approved draft |
| Hearing scarcity | Scheduling support | Adjournments, utilisation and wait time |
| Vacant judicial posts | No direct AI substitute | Recruitment, allocation and capacity |
The rebound risk
Making one stage faster may increase pressure downstream. Faster intake can create a larger hearing queue. More drafts can create more review work if their error rate is high. A pilot therefore needs end-to-end measures, including correction time and the destination of displaced work.
Quality is a throughput variable
An inaccurate summary does not save time when a judge must reconstruct the record. Recognition quality, omission rate, factual support and authority verification are operational metrics, not abstract ethics. A system should permit partial progress: accepted pages can be processed while an unclear page is reported for later replacement.
A credible pilot
- Choose one case type and one stage.
- Measure the current time, error and rework distribution.
- Use a representative historical test set with protected handling.
- Set failure thresholds before seeing results.
- Run human review and record corrections.
- Check whether time moved to another stage before claiming a backlog effect.