Three weeks of deep operator interviews. We sit with the support engineers fielding escalations, the controllers reviewing accruals, the SDRs answering objections. Every recurring pattern, every unwritten rule, every "we just know to do X when Y happens" becomes a candidate for the skill file.
Output is a structured map of how the business actually runs — not how the org chart says it does. The gap between those two is where most agent deployments fail.
The extracted know-how is encoded into a living, versioned skills file inside CAEL. Each rule has provenance — which thread, which controller, which policy document. When the source changes, the skill flags drift and proposes an update.
Audit-ready by construction: every action a downstream agent takes can be traced back to the operator instruction that authored the rule.
Kassper, Gogins or Dabloom is wired up to the relevant CAEL skill. The first four weeks run in shadow mode: the agent proposes actions, operators review, nothing touches a customer record.
After enough operator-confirmed cycles to trust the routing, the agent moves to supervised live mode — still gated by a confirmation queue for anything material.
The agent learns the company, not the other way round. Each engagement self-funds within the first two quarters. ROI is modelled upfront and paced against measurable, traceable return.
Quarterly skill reviews keep policy and execution aligned — when leadership changes the playbook, the skill updates within the cycle, not the next planning quarter.