The pattern is depressingly common: a CIO reads about Now Assist, sponsors a pilot, sees impressive demos, and then watches adoption stall at 15-20% six months in. The technology works. The deployment pattern is what fails.
Here are three patterns that have worked across our production deployments — at a tier-1 bank, a major telco, and a multi-state health system.
Pattern 1 — Grounded summarization on case worklog
Where it works: contact-center, claims, IT support — anywhere agents work with long case histories.
The play: Now Assist summarizes the case timeline at the top of the case form. Agent reads the summary instead of scrolling through 47 work-notes. We measure time-to-context, not draft acceptance.
Outcomes we’ve seen: 30-40% reduction in average handle time, sustained.
Pattern 2 — Drafted-response, agent-edits
Where it works: written customer responses (email, chat) where tone and accuracy matter.
The play: Now Assist drafts the response based on case context + the agent’s knowledge base. Agent edits before sending. Track edit-distance per draft as the quality signal.
Critical setup: ground the model on YOUR knowledge base, not a generic LLM. Production deployments need your content shaping responses.
Outcomes we’ve seen: 73% draft-acceptance rate (low edit distance) in the bank deployment; 30% reduction in handle time at the telco.
Pattern 3 — Self-service triage with explicit hand-off
Where it works: customer-facing portals where users want fast answers but also need a human path.
The play: virtual agent triages requests, resolves what it can with full identity verification + transactional capability, and hands off cleanly to a human agent with the conversation context preserved.
The non-obvious thing: invest more in the hand-off than the conversation. A clean hand-off with full context preserved is what makes users tolerate a less-perfect virtual-agent experience.
The pattern that keeps failing
Generic “ask me anything” deployments. Without scope, grounding, and measurement, they’re demo-ware.