Pair your devices with a code and playback position follows you: pause on this device, hit resume on the other. Position is saved to the site every minute and on pause.
Open this panel on your other device and enter the same code.
Starts this lesson and continues through 10 more to the end of the course.
The last mile of being the Quick expert isn't a feature — it's being able to answer "what did the AI do, who told it to, and what did it cost?" This lesson covers Quick Research (the most expensive button in the product), the logging architecture that answers auditors, and the cost traps across everything this module introduced.
What it is (using-amazon-quick-research, verified 2026-08-26): "conduct comprehensive research by analyzing multiple data sources and generating detailed reports… gather, analyze, and synthesize information from various sources including web search, uploaded files, connected data spaces, knowledge bases, actions, and third-party data providers… AI-generated research reports with proper citations and source tracing."
The operational facts (view-research-report, verified 2026-08-26):
Research also appears as a step type inside Flows — the page "Quick Research steps in Amazon Quick Flows" exists, but its content wouldn't render on fetch (2026-08-26), so treat the details as unverified and check the page before designing around it. Similarly, we found no documented hard limits for Research (report length, source count, runs/month beyond agent hours) — say "not documented" rather than inventing a number.
The suite's audit story is split across services, and the split is the exam question. Incident response, logging and monitoring (verified 2026-08-26):
| You want to know | Look in |
|---|---|
| "Which API calls occurred for AI features (flows, agents, automations, action connectors)" | CloudTrail (management and data events) |
| Chat conversations and feedback content | CloudWatch vended logs — CHAT_LOGS, FEEDBACK_LOGS |
| Index consumption, knowledge-base syncs | INDEX_USAGE_LOGS, KB_FILE_SYNC_LOGS |
| Agent-hours consumption | agent-hours usage logs (CloudWatch) |
| Non-API user events like dashboard views | CloudTrail's "documented set of non-API events" |
| Routing/alerting | EventBridge; CloudWatch metrics |
| The built-in usage analytics dashboard | gated by the IAM action quicksight:QuickSuiteUsageMetrics |
The sentence to internalize: CloudTrail has the actions; CloudWatch vended logs have the conversations. An auditor asking "show me what users asked the AI" gets CHAT_LOGS, not CloudTrail — and if vended logs were never enabled, that history may simply not exist. Enabling them is a day-one governance task, not a post-incident one.
For Automate specifically (lesson 5's pages, verified 2026-08-26): per-run "logs, metrics" with inputs/outputs "available as structured artifacts in the logs panel on the Runs page", plus per-agent unit testing with metrics ("Total execution time, Number of tools used, Number of tasks created"). Flows and automations are also governable pre-deployment: flow sharing "may require approval review", automations deploy only from committed versions, and the quota table carries approval policies (500/account, 100/asset type — both adjustable; verified 2026-08-26 on the Service Quotas page).
Chat, agents, flows, and automations can all write to external systems through action connectors — the supported-integrations table lists ~60, from Slack and Jira to SAP, plus generic MCP, OpenAPI, and REST API connectors (supported-integrations, verified 2026-08-26). Governance handles:
aws quicksight *-action-connector*.The complete meter list for the AI layer, with sources from this module's lessons (all verified 2026-08-26):
| Meter | Rate | Trap |
|---|---|---|
| Seats | $0 / $20 / $20 / $40 per user/month | Plans ≠ editions; Topics still gate on Enterprise Edition |
| Infrastructure fee | $250/account/month (Pro/Ent) | Exists at one user or one thousand |
| Index storage | 25/50 GB/user pooled; $5/GB/month over | Counts source bytes; extra-Region capacity is all overage |
| Agent hours | 4/8 per user pooled; $3/hour over | Research ≈ 0.3–0.7 hr/run; scheduled automations run unattended |
| Dashboard Q&A | "the associated enablement fee" (dashboard-qa page) | Per-feature fee, easy to miss in a proposal |
| Custom-model (Bedrock) inference | billed to the Bedrock account | Invisible on the Quick bill |
For an AI-layer case, the evidence bundle (pattern from Q2/Q8, adapted):
account ID + home Region · plan/edition · the feature and its doc-tree URL (say which of the
three trees, given the inconsistencies) · for flows/automations: automation-group ID, automation
ID, JobId from start-automation-job, job status, and the run's log artifacts · for
index/space issues: index capacity setting, INDEX_USAGE_LOGS excerpt, KB_FILE_SYNC_LOGS for
sync failures · for chat quality issues: the conversation from CHAT_LOGS and the Explanation
panel's generated SQL · CloudTrail event IDs for the API calls involved.
CHAT_LOGS) — not CloudTrail. The catch: vended logs must have been
enabled; there's no retroactive capture, and chat conversations are retained 90 days in-product.quicksight:QuickSuiteUsageMetrics.describe-automation-job status; the run's structured input/output artifacts from the
Runs page logs panel; CloudTrail events for the action-connector calls (plus whether the target
app's credential is valid).CHAT_LOGS entry if you have
one; end on the usage dashboard.