Agent Assist on Amazon Connect + Bedrock

The streaming transcript path

The transcript path is the load-bearing part of the architecture. Contact Lens streams partial transcripts into Kinesis. A lightweight Lambda enriches, chunks and redacts PII before the payload ever reaches the model.

Total round-trip from utterance to on-screen assist should stay under 2.5 seconds at p95. Above that, the assist surface stops feeling real-time and agent adoption falls.

Retrieval layer — the three-way choice

OpenSearch gives the most control and the best cost at scale. Bedrock Knowledge Bases is the fastest to production. Kendra excels on unstructured content with strong relevance out of the box.

Pick with the evaluation harness against your actual knowledge corpus, not with a vendor slide.

Prompt orchestration and guardrails

System, task, tool and guardrail prompts are separated and versioned independently. Bedrock Guardrails run as independent policies with veto authority — they never share weights or context with the assist model.

Provider failover to Anthropic on Bedrock is wired from day one. Prompt-prefix and retrieval-result caching cuts LLM cost 30% to 45% on estates with high intent repetition.

CRM write-back that actually closes the loop

Assist that ends when the call ends is a demo. Post-call summary, disposition and next-best-action write back to Salesforce Service Cloud or Dynamics 365 Customer Service using the platform-native record schemas.

The agent never manually copies the summary. It is present when they open wrap-up.

Terraform-first, no click-ops

The reference ships as a Terraform module layout with environment separation, cost tagging by workload and business unit, secret handling via Secrets Manager, and IAM least-privilege by default.

The six SLOs to run in production

Transcript latency p95. Assist latency p95. Retrieval precision. Guardrail veto rate. Agent adoption. Cost per resolved contact. If you cannot see all six on one dashboard, you cannot operate the workload.