Labs
Where narrow and wide models earn their place in a workflow.
Claims, pharmacy and freight. Each lab tests what the models could change for the business and its customers.
- For the business
- Which routine work can leave the desk, and what errors cost
- For customers
- Whether answers can come faster without more mistakes
- Across industries
- Which lessons carry over, and which don't
Labs
-
Tested ·
A hundred dots, every set of fresh claims pooled. Violet: paid with no person. Violet ring: asked for documents, not resolved yet. Ink: to a person, not certain enough or over the payment limit. Small grey: must reach a person by rule.
Claims
About half of 3,000 fresh claims paid with no person; one wrong decision in the 1,613 decided.
Declared in advance. An earlier version failed. Synthetic claims, not real use.
Read the evidence: Claims -
Demos ·
Twenty dots per demo, over all requests. Violet: settled by code. Struck ring: approved with no person, and wrong. Ring: sent back for an item. Navy: to a person.
Pharmacy
Triage caught most serious reports. Prior authorisation approved about four in ten complete requests, none wrong.
Refills, as built, were not safe to approve alone. Three small demos, not validated.
Read the evidence: Pharmacy -
In progress
The narrow model on the status step: measured. The ladder: results pending.
Freight
Being built: loads followed from tender to invoice through rules, a narrow model reading carrier messages and a wide model reading documents.
People keep the exceptions. Results pending.
See the design: Freight
Predictions written before results. Failures published. Synthetic data unless stated.