AI / Automation
Process Automation
Automation with ROI measured on the process, not in a sandbox. If the metric doesn’t move, the project didn’t work, however good the demo looked.
Measured on the process
Before anything is built, the process gets a baseline: cost per case, cycle time, error rate, human hours. The automation is then accountable to those numbers, and to nothing else. This is the entire method; everything below is execution discipline.
Where LLM automation actually pays
The sweet spot is the semi-structured middle: document intake, classification and routing, first-draft generation, reconciliation, triage. Work with judgement, but bounded judgement: too irregular for old-school RPA, too voluminous for people. Fully deterministic flows don’t need a model; fully open-ended judgement shouldn’t get one.
Execution discipline
- Human-in-the-loop by design, with confidence thresholds deciding what routes to review. The automation earns autonomy case-type by case-type, on evidence.
- Failure paths are first-class. What happens on model error is designed, not discovered.
- Unit economics tracked from day one: cost per processed case, including inference, so scaling the automation never becomes its own budget surprise.
- Boring reliability: versioned prompts, evaluation sets, monitored drift. The same operational discipline we apply to any production system, because this is one.
