Read and write access to the systems that fix a ticket, scoped exactly like a tier-one human, with escalation above that.
Attributes, descriptions and categorisation generated and validated at scale, reviewed by merchandising before publish.
Retrieval that understands how customers actually phrase things, measured on conversion rather than relevance alone.
Load testing, caching and cloud structure so a campaign is a planned event rather than an incident.
Short cycles, visible progress, and a scope you can change. You see working software every week rather than a status report.
The repository, the infrastructure code, the evaluation set and the documentation. In your accounts, under your licence, from the first commit.
Read real ticket and catalogue samples, agree the actions an agent may take, and write the evaluation set from past cases.
A working slice on sandbox systems, measured against how your team handled the same cases.
Live on a percentage of volume, with a kill switch and daily review, expanded as the numbers hold.
Seasonal readiness reviews, plus evaluation runs before any model or prompt change ships.
Generated copy follows your tone and claims rules, and merchandising signs off before anything goes live.
Anything with a money value has a ceiling the agent cannot exceed without a person.
Order and profile data is scoped per request and never used for training.
Success is resolution rate and conversion, not chat volume.