Your data. More useful with every question.

Which paid orders still need shipping?

For someone keeping a small business’s orders moving. The answer is spread across three files. Matching them once also prepares the data for later questions.

Start with: 12 orders, payments, and shipment records. All amounts are euros. The files share an order number; a payment covering the order total counts as fully paid. A shipment record means the order is complete in this simplified example.

“Match these files by order number. Which fully paid orders have no shipment? Save the combined data so we can use it again.”

Sample answer: Five paid orders still need shipping, worth €12,200.

Order	Country	Amount
1007	Spain	€5,400
1008	France	€3,600
1010	Spain	€1,000
1011	France	€1,200
1012	Portugal	€1,000

Ask next: “Which of these are going to Spain?” The same prepared data gives two orders, worth €6,400.

Keep using: The source files, combined records and saved matching procedure. You can ask another question without assembling the files again. Your assistant can also export the open orders as a CSV file.

Use the same definitions next month

For a consultant preparing recurring client analysis. Agree on the meaning of “open order” once, then reuse that preparation when another batch arrives.

Start with: The saved matching procedure and a previous-month order file. For this fictional comparison, the payment and shipment lists above cover both batches. The previous batch has nine orders; the new batch has twelve.

“Here is the next batch. Use our existing definition: fully paid with no shipment. Replace the prepared results and tell me what changed.”

Checked sample results:

Same definition	Previous batch	New batch
Orders in the source	9	12
Paid orders awaiting shipment	2	5
Value awaiting shipment	€9,000	€12,200

Ask next: “Which records account for the €3,200 difference?” Orders 1010, 1011 and 1012 were added, worth €1,000, €1,200 and €1,000 respectively.

Keep using: The definition and preparation steps, alongside each retained batch. Your assistant reruns the saved procedure when you ask; new files do not trigger an automatic update.

An answer for you. A view for the team.

For a business owner who needs both a quick answer and something others can explore. The prepared data can serve both purposes.

Start with: The twelve combined order records from the first story.

“Tell me the value still waiting to ship. Then make a view where the team can filter those same records by status and see the totals by country.”

Sample answer: “Five orders are open, worth €12,200. Spain accounts for €6,400, France €4,800 and Portugal €1,000.”

Explore the same sample results in the operations dashboard →

That working dashboard uses the same fictional prepared records. Its filters run read-only sample queries; they do not run an AI assistant or refresh a source.

Ask next: “Give me just the open orders as a CSV.” File export already exists. Later questions and outputs can reuse the preparation.

Keep using: One prepared dataset for answers, exports and a dashboard. In your account, websites start private. You choose who can see a published result through Website access; the assistant does not grant that permission.

Try a different rule before adopting it

For an analyst testing a classification. Compare an alternative while keeping the accepted data available for ordinary work.

Start with: The five open orders and a saved rule that marks orders worth at least €3,000 for review.

“Try raising the review threshold to €5,000 in a separate job. Show what would change, and leave the accepted classification alone.”

Checked sample results:

Rule	Orders marked for review	Value of those orders
Accepted: at least €3,000	1007 and 1008	€9,000
Experiment: at least €5,000	1007 only	€5,400

Order 1008 moves out of the review group in the experiment. The accepted classification still marks both orders for review.

Ask next: “Keep the original rule. Save a note explaining the alternative.”

Keep using: The accepted data and the separate result for comparison. Your assistant checks the job and copies selected results back only when requested. Jobs are bounded tasks, not recurring schedules; unfinished jobs stop on a server restart.

Ask a new question about last year

For someone investigating a question they did not anticipate. Retaining the underlying records lets you calculate a new result later.

Start with: Two retained fictional snapshots, labeled September 2025 and September 2026. The earlier order file contains nine orders. The current file contains twelve. Both use the same matching rules and the illustrative payment and shipment lists above.

“We never made this report: how many fully paid, unshipped orders worth at least €3,000 did we have in each snapshot?”

Checked sample results, calculated with the same question:

Retained snapshot	Matching orders	Value
September 2025	1: order 1007	€4,000
September 2026	2: orders 1007 and 1008	€9,000

Ask next: “Use €2,000 instead.” The earlier snapshot has two matching orders worth €6,000; the current snapshot has two worth €9,000.

Keep using: The retained records, not just a report with the questions you thought to ask at the time. No report needs to have been prepared in advance.

The data must have been retained. An old definition of a live remote query does not preserve its old rows. This example uses saved workspace data and does not promise unlimited history or reconstruct missing external records.

Recover the part the agent got wrong

For anyone handing data changes to an assistant. Repair selected records while keeping unrelated work done afterward.

Start with: A retained version of the order data. An incorrect edit replaced the countries on orders 1007 and 1008 with “Unknown”. A later, intentional edit changed order 1009’s country to Ireland.

“Compare the countries with the saved version. Restore only orders 1007 and 1008. Keep the correction to order 1009.”

Checked sample result:

Order	Before recovery	After recovery
1007	Unknown	Spain
1008	Unknown	France
1009	Ireland	Ireland

Ask next: “Show me the differences again.” Only the intentional change to order 1009 remains relative to the earlier version.

Keep using: The repaired current data and the retained versions. Recovery adds a new version; it does not erase the work in between. It cannot undo an external action or recall a result already delivered elsewhere.