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.
These are written to be said. Practice out loud; anything you can't finish in 90 seconds, split.
"What is Amazon Quick, and how does it relate to QuickSight?"
It's AWS's AI workspace — chat, agents, automation, research — built around what used to be QuickSight. The BI product still exists as Quick Sight, a component inside it. The key operational fact: the rename was console-and-docs only. Every API, IAM action, and ARN still says quicksight, deliberately, so existing integrations didn't break. So I read docs under the new names and write code under the old one.
"What's a space, and why would a team use one?"
A curated container — files, dashboards, topics, knowledge bases, actions — that grounds the AI for a domain. Instead of chat answering from everything in the company, you pin it to the finance space and it retrieves from what finance curated. One design rule I always flag: linked resources keep their own permissions, but uploaded files are readable by everyone in the space. So uploading a file to a shared space is an access decision, not a convenience.
"How does natural-language querying actually work now?"
Two tiers. For one dashboard, there's a one-click "Allow data Q&A" checkbox — no modeling. For real cross-dataset questions there are topics: up to twelve datasets with defined join keys, and the agent parses intent, walks the relationships, and generates actual SQL — joins, unions, subqueries — which you can inspect in the Explanation panel. And that's the trust story: every numeric answer can show the SQL that produced it.
"Flows versus Automate — when do you use which?"
Flows run as a person: up to thirty-five steps, built in natural language, shared like a document — personal productivity. Automate runs as the organization: projects, typed variables, version control with commit-and-deploy, stored credentials, human-in-the-loop approvals, and an API trigger. My one-line router: if it should keep running when the author quits, it's Automate.
"What does the AI layer cost?"
Three meters beyond seats. A flat $250 per account per month infrastructure fee on the paid tiers. Index storage — 25 or 50 gigs per user pooled, five dollars per gig over, and it counts the source file size, not extracted text. And agent hours — four or eight per user pooled, three dollars an hour over, shared by Research, Flows, and Automate. Plus two sneaky ones: dashboard Q&A has its own enablement fee, and bring-your-own-model automations bill inference to the Bedrock account, off the Quick bill entirely.
"Design the knowledge setup for a support chat agent."
Split by function. The refund policy and tone guide go in reference documents — those live permanently in the agent's context, so they're always applied; the limits are tight, ten docs and a hundred thousand characters, which is fine for policy. The ticket archive goes in a space the agent searches at question time. If I invert that, the policy sometimes doesn't get retrieved and the archive blows the context limit on day one.
Follow-up they ask next: "And how do you stop it answering off-topic?" — Persona instructions, plus building it "preconfigured with resources" so it answers only from its wired-in sources, and I test the refusal behavior before sharing it.
"Why can't you index Salesforce content, and what do you do instead?"
The supported-integrations matrix gives Salesforce actions support only — there's no Salesforce knowledge base, so its content never lands in the Quick Index. Only six sources are indexable: S3, SharePoint, OneDrive, Google Drive, Confluence, and the web crawler. So for Salesforce I use actions — the agent queries it live, which honestly is often better: no sync staleness. If someone truly needs Salesforce content in the corpus, it has to be exported to an indexable source like S3 first — with the access implications thought through.
Follow-up: "What's actually under the index?" — Amazon Q Business machinery. The tell is that CMK-encrypted knowledge-base sync needs both the quicksight and qbusiness service principals in the key policy, and you can attach an existing Q Business index directly.
"An auditor asks: show me what users asked the AI last quarter."
The split is: CloudTrail has the actions — API calls for flows, agents, automations — but chat content is CloudWatch vended logs, the CHAT_LOGS type. And the catch I lead with: vended logs capture only from when they were enabled. If nobody turned them on, that history doesn't exist, and in-product retention is ninety days. Which is why enabling them is on my day-one checklist, next to CloudTrail itself.
Follow-up: "And how would you catch runaway agent-hour spend?" — The agent-hours usage logs in CloudWatch plus the built-in usage dashboard — gated by its own IAM action, QuickSuiteUsageMetrics — with an EventBridge-driven alert before overage, not after.
"You inherit an account full of topics. First week?"
Inventory legacy versus new. Legacy topics — anything pre-redesign — keep their metadata inside the topic object and answer from one fuzzy-matched dataset. New topics keep metadata on enriched datasets and generate multi-dataset SQL. They're even separate API families — topic versus topic-v2 — so any automation has to know which it's touching. Migration isn't automatic; it's a deliberate Dataset Enrichment exercise, so I'd sequence it by which topics get the most questions.
Follow-up: "How many fields can a topic have?" — Not documented. That's the honest answer — the nearest published number is two thousand fields per dataset, and I'd point at the quota page rather than quote folklore.
"What would make you nervous about rolling out Automate's UI agent?"
It's an AI operating a browser under stored credentials — so the blast radius is whatever that login can do. I'd want three things: the HITL gates configured with named resolver groups for anything irreversible, the credential scoped to a least-privilege service account on the target site, and the run artifacts reviewed — every run keeps structured inputs and outputs in the logs panel. And I'd treat MCP connectors the same way: that's third-party tool code in the loop, so approving one is a dependency review.
"Design 'ask our data anything' for a 500-person company on Quick, governed and budgeted."
Requirements first: governed answers with traceable sources, per-department scoping, an audit trail, and a predictable bill.
Architecture: per-department spaces as the grounding unit — curated files plus knowledge bases over SharePoint and S3, remembering only six source types index. Enrich the core datasets once — synonyms, descriptions — because that investment is reused by every topic and chat session. Two or three real topics for genuinely cross-dataset questions; the one-click dashboard Q&A checkbox everywhere else — cheaper and instant. A custom agent per department, preconfigured with its space, policy in reference documents.
The trade-off I'd surface: all-data-mode chat is magical in demos and ungovernable at scale, so default users into department agents and leave open chat opt-in.
Cost: Enterprise tier for fifty-gig index pooling; the meters are index gigs at source size and agent hours — Research at roughly half an hour per run is the volatile one, so I'd cap expectations at launch.
Monitor: CHAT_LOGS enabled before user one, agent-hours usage with an alert at 80% of pool, INDEX_USAGE_LOGS for capacity, and a monthly look at unanswered-question patterns to drive enrichment.
"Space uploads are failing across the org. Go."
One — index capacity: is it full, and is auto-scale at its configured max? The docs say space uploads are rejected when index storage is full — the error surfaces in the wrong layer. Two — file specifics: over 50 megs for a non-document type? Unsupported extension? Three — per-space quotas: file count near ten thousand, or total size at the documented cap — remembering the 1 GB versus 10 GB doc conflict, plan on the lower. Four — if it's knowledge-base content rather than uploads: sync status, the deletion-threshold safeguard, and KB_FILE_SYNC_LOGS. Five — if I open a case: account, home Region, index capacity setting, the exact error, and the log excerpts.
"A deployed automation 'ran but nothing happened.' Go."
One — describe-automation-job with the JobId: did it actually succeed, or is it queued, failed, or stopped? Two — the run artifacts on the Runs page: what inputs went in, what outputs came out? Three — version check: is the deployed version the one containing the change — live, committed, deployed are three different states. Four — credentials: website credential still valid? Action credential authorized? Five — the target system: CloudTrail for the connector calls, then the app's own audit log. Six — HITL: is the run actually parked on a human task in the task center nobody's resolving?
"Chat gives a wrong number from a dataset. Go."
One — Explanations: read the generated SQL. That usually ends the mystery — wrong filter, wrong join, wrong interpretation of the question. Two — scoping mode: was it answering from the dataset or from a stale document in all-data mode? Three — semantics: does the dataset's enrichment define the term the user asked about, or did the model guess a synonym? Four — the data itself: same query by hand against the dataset — is SPICE stale? That's a Q2 refresh question, not an AI question. Five — fix at the right layer: enrichment or a topic reviewed answer for recurring questions — not a prompt workaround.