The question that still takes two tools
A sales leader asks whether FY2025 results in North America actually lined up with the strategy memo, or whether the memo got ahead of the numbers. In a lot of companies, that is still two systems, a folder hunt, and someone who remembers which PDF is current.
Tables hold the metrics. Documents hold the story. Amazon Quick is AWS’s conversational analytics product for that job: a custom chat agent that business users can point at a structured dataset and an indexed set of documents, and then question in plain language.
That gap, numbers in one system and narrative in another, is the actual business problem these agents are being sold to close. The useful version is narrow. It is a scoped agent over a shelf of company files and figures, and the answer is only as good as what you put on that shelf.
A general assistant can also read a PDF. What this setup adds is a governed dataset beside a document index, answers that can cite those sources, and the files staying in AWS.
What the walkthrough shows, and what it does not measure
AWS’s walkthrough uses a fictional company, AnyCompany. Three documents sit in Amazon S3 and are indexed as a knowledge base: a business review, a market analysis, and a product strategy memo. A sales file, Product_Sales.csv, is loaded separately as an enriched dataset covering FY2024-FY2025 across North America, EMEA, Asia Pacific, and Latin America.
Both sources are gathered in a Space, the workspace that binds the knowledge base, the dataset, and the agent. They are then handed to a custom chat agent. The questions are the ones a revenue team already asks: how sales moved by region in 2025, what the strategy memo treats as the priorities, and whether those results match the trends in the market write-up.
The same walkthrough has the agent generate Excel and Word files, plus a downloadable eight-slide PowerPoint named AnyCompany_FY2025_Sales_Performance.pptx. A deck a manager can open is more useful than a transcript that dies in a browser tab. It is still a sample artifact from a fictional company.
AWS’s pitch is that people no longer need to know where the data lives or how to write a query. The walkthrough does not measure whether that is true. It describes regional breakdowns and a strategy summary in the replies. It does not publish the underlying sales figures or the memo’s priorities. If you are buying on the strength of those replies, you are buying a screenshot.
The enrichment step is the actual work
Uploading a CSV does not make it legible to an agent. In the walkthrough, several fields arrived as plain text and had to be recast before the file was usable: Order Date and Date Key to dates, Customer ID to an integer, Sales, Discount, and Profit to decimals, Quantity and Unit Price to integers.
The team then wrote the business rules the columns do not carry themselves. Revenue is in US dollars. Comparisons group by those four regions. Order Date drives time trends. For this sample, fiscal year matches the calendar year. Sales means quantity times unit price times one minus discount.
Leave that formula unstated and the same file yields two regional stories. If Discount is already inside Sales and the agent applies it again, a region that cleared plan looks like it missed. If the column is still gross and the agent reads it as net, that region looks stronger than the books. Treat the fiscal year as something other than the calendar year and the “2025” trend is a different set of orders than the one finance will defend. Same CSV, fluent voice, wrong quarter.
The persona is written the same way, as a job description rather than a hope:
You are a Business Intelligence Assistant designed to help organizations break down information silos.
The walkthrough also tells that agent to cite sources, explain its conclusions, and stay on practical business value. Citations are the right instinct. They are not, by themselves, a number a controller will sign.
What should change the buying decision
AWS documentation defines a knowledge base as an indexed collection built for generative-AI retrieval and question answering. You connect a source, choose content, index it, and make it available to spaces and agents. From indexed S3 content, people can ask across file types and get file metadata such as modification dates. S3 is one supported source. The docs also list Google Drive, OneDrive, Confluence, SharePoint, and a web crawler.
The walkthrough’s split (dataset for the sales file, knowledge base for the PDFs) is a sensible design. It is not a wall. Documentation says a knowledge base can index text, documents, and structured data. CSV and Excel are allowed as tabular knowledge-base content. An index is still not a live warehouse query, and it is not a folder the chat agent should be rearranging.
License. AWS documentation makes an Amazon Quick Enterprise subscription a prerequisite for S3 knowledge-base integration, along with bucket access, a populated bucket, integration permissions, and an administrator grant on the buckets. The walkthrough, separately, treats Enterprise as required for the person creating datasets. Confirm both before you staff a pilot. Cross-account S3 access is documented only within the same AWS Region. The walkthrough also tells you to keep the bucket in the same Region as Amazon Quick.
Who can see what. Documented access is coarse and knowledge-base-level: people receive information only from knowledge bases they are authorized to use. Knowledge bases can be shared through spaces and reused by more than one space and agent. That is not row-level security on the sales file. The walkthrough says answers follow the permissions of the person asking. The docs do not say what the agent does when rights differ across a document and a dataset.
Freshness. Synchronization is daily by default and can be edited. A safeguard can skip a sync that would delete too much content. A Monday pipeline review can easily be talking to Friday’s index. Say that out loud. Do not brief the board as if the chat were the warehouse.
Size. Standard documents are documented up to 500 MB, with at most 30 MB of extracted text per document. CSV and Excel files in a knowledge base stop at 50 MB. A decade of board decks will not all fit on the first try.
Read, not delete. Knowledge-base ingestion is for retrieval. AWS documents upload, download, and delete against S3 only for Quick Automate, a different capability. An agent that can read a strategy memo should not, on that basis, be deleting objects in a bucket.
Under AWS’s shared responsibility model, you still own how access is configured for your data, knowledge bases, and agents. The product will not invent those boundaries for you.
Where a revenue team should use this
The useful test is specific. Regional comparisons, product-line revenue, quarterly trends, and a check of those trends against a market memo are exactly the questions the AnyCompany demo is built to ask. That is a fair pilot for AI for sales work: one cross-source question the team already argues about in meetings.
Send any deck upstairs only after a person who owns the number has read it. A confident slide with the wrong region on it is worse than the old two-tool process, because the old process at least looked slow enough to doubt.
Clean the sales file and write down what Sales means. Index a small, trusted set of documents, not the entire shared drive. AWS documentation notes that several knowledge bases can come from one source, share an index, and stay distinct (a policy library in one, customer feedback in another), so split domains when the questions are different. Decide who may use each one. Then ask the question, and see whether the citations hold.
AI agents will keep absorbing the chores that sit between a spreadsheet and a PDF. The companies that get value from this one will treat the agent like a new analyst with a badge: a defined scope, a known data cutoff, and a manager who still signs the number.
Questions worth asking first
Can one AI agent answer from both a sales file and a stack of memos?
Yes. The AWS walkthrough connects an enriched sales dataset and an S3 knowledge base, then questions both through one custom chat agent inside a Space, the workspace that binds those sources to the agent.
Do we need Enterprise to try the S3 piece?
AWS documentation lists an Amazon Quick Enterprise subscription as a prerequisite for S3 knowledge-base integration, along with bucket access and an administrator grant. That is a buying decision, not a toggle you discover at the end of a pilot.
Who sees which documents?
Access is documented as coarse-grained and knowledge-base-level: users get information only from knowledge bases they are authorized to use. Do not assume per-document or row-level controls from the demo.
How big can the files be?
Standard documents are documented up to 500 MB, with at most 30 MB of extracted text per document. CSV and Excel files in a knowledge base stop at 50 MB. A long archive of board decks will not all fit on the first try.
How fresh is the index?
Synchronization is daily by default and can be changed. A Monday review can be talking to Friday’s index, so do not treat the chat as if it were the warehouse.
What if the memo and the spreadsheet conflict?
The documentation does not say what the agent does when they disagree. Until you test it on your own files, treat a conflict as something you design for, not something the agent will resolve quietly.
Can this agent also move or delete files in S3?
Not through knowledge-base ingestion. AWS documents upload, download, and delete only for Quick Automate, which is a different capability from asking questions of an index.
What has to be true before we trust a cross-source answer?
Sales has to be defined, the document set has to be small and trusted, separate questions should sit in separate knowledge bases, and a named owner has to sign the number. The agent can draft the comparison. It should not be the last signature.