The problem
Most support volume is the same twenty errands — refunds, address changes, booking changes, account questions. Human teams burn out on them; deflection bots that can’t actually do anything make customers angrier. The market has now proven full resolution works when the agent is integrated deeply enough to act.
The system
An agent connected to the systems of record — billing, orders, bookings, CRM — that resolves whitelisted errand types end to end: verify identity, take the action, confirm, log. Every action type has explicit permission boundaries and spend limits; sentiment, vulnerability, and anything off-list escalate to a human with full context attached. Resolution rate, containment, and CSAT are measured against your human baseline from day one.
How it's built
- Errand-type whitelist with per-action permissions, limits, and rollback paths
- Deep integration: the agent acts through your APIs, never screen-scrapes prod
- Guardrails: identity verification, sentiment/vulnerability routing, human handoff with context
- Eval harness replaying historical tickets; containment and accuracy reported weekly
Delivery
Sprint replays history to size the automatable share; Build ships one errand type live behind limits, then expands by evidence.
What to expect
- A measured containment rate on covered errands, not a vendor promise
- Resolution time on automated errands measured in minutes
- CSAT tracked against the human baseline before autonomy expands
Documented results in the wild
Independent, published deployments of this class of system — cited as market evidence that it works at scale. These are not our clients.
- Klarna AI assistant handled two-thirds of support chats in month one — the work of ~700 agents — cutting resolution from 11 minutes to under 2, with an estimated $40M profit improvement. Klarna, 2024 ↗
- Air India GenAI assistant handles ~40,000 queries daily with 97% full automation; 13M+ conversations resolved. Microsoft Customer Stories, 2025 ↗