AI in business7 min
First the task and KPI
The model is chosen after metrics: response time, auto-routing share, input error rate. Otherwise you end up with a demo chat that isn’t embedded in the workflow.
Pilot on a narrow scenario
One or two streams, baseline vs after. This shows whether a ready-made API is sufficient or a custom pipeline is needed.
Quality: benchmarks and escalation
Evaluation set, confidence threshold, human-in-the-loop. If the model is unsure, we create a ticket instead of fabricating an answer.
In the production system, not on a separate tab
The assistant lives within tickets, CRM, or the portal. A separate “talk to AI” window rarely changes operations.
Risks
PII, drift, inference cost, leaks to external APIs. This is part of design, not an add‑on to a presentation.