Onevium
All storiesAutomation

A small team's first AI workflow: a report you can check

Build a first AI reporting workflow with three tiny CSV files. Find duplicates, trace every count, and spot the missing data before trusting the summary.

7 min read

Choose a report someone already has to check

Friday afternoon arrives, and someone on a small team still has three exports to compare before writing the weekly update. This is a useful first AI task: the inputs are concrete, the answer is small, and another person can check it without rebuilding the whole workflow.

In its September 10, 2026 article, Anthropic describes workshops with more than 1,000 small business owners and operators. Reporting came up frequently, along with concern about accuracy. These are workshop observations, not a representative study or a controlled test of productivity gains.

The exercise below is Onevium's original teaching example. It uses entirely fictional data and includes deliberate defects. You will produce a draft report with traceable counts and unresolved questions; no business system or customer account is needed.

Further readingRead Anthropic's workshop account

Prepare three small files

Create an empty practice folder named weekly-report-practice. Save each labeled block below as its own plain-text CSV file, leaving the filename label out of the file. Keep the headers and blank cells exactly as shown. These fictional exports cover September 7–11, 2026; the date range is supplied here because the tiny samples omit date columns.

An order row records one order and its total item quantity. A refund row records returned items, and a support row records one ticket. Quantities are item counts, not money. Refund records in this period may refer to orders from an earlier period. That means these files cannot establish net sales for the week.

Fictional practice data · three separate files
orders.csv
order_id,status,units
O101,paid,2
O102,paid,1
O102,paid,1
O103,cancelled,3
O104,paid,

refunds.csv
refund_id,order_id,units
R201,O101,1
R202,O999,1

support.csv
ticket_id,order_id,status
T301,O101,open
T302,O102,closed
T303,,open

Give the assistant the counting rules

Decide the meaning of each number before requesting a polished paragraph. For this exercise, count unique order IDs, keep cancelled orders separate, and collapse a fully identical duplicate only when calculating. A blank quantity remains unknown. Never change the input files to make the totals easier to produce.

Ask for references precise enough to inspect: filename, record ID, and line number, counting the header as line 1. A matching order ID shows that two records can be linked; it does not prove why an item was returned or why a customer opened a ticket.

Copy this task into your project conversation
Read only these files in this practice folder:
orders.csv, refunds.csv, support.csv.
They are fictional exports for Sep 7–11, 2026.

Return a draft weekly report in this conversation.
Show each file's data-row count, excluding its header.
Count unique orders by status.
Collapse fully identical duplicate rows for counting.
List duplicates without editing the files.
If one ID has conflicting records, flag the conflict;
do not choose a record or total that group.

Sum known units for unique paid orders only.
Keep missing units unknown, not zero.
List refunds without a matching order.
Count open tickets and flag missing order IDs.
Do not subtract refund units from order units.
Do not infer trends or reasons from these files.

Cite filename, ID, and line for every count or issue.
Count each file's header as line 1.
Separate verified results from unresolved questions.
Do not edit files, fetch other data, or send messages.

Check the answer against this key

Check the counts yourself before reading the assistant's summary. There are five order rows but four distinct orders. O102 appears twice with identical values, on lines 3 and 4. After accounting for that duplicate, three orders are paid and one is cancelled.

The known quantities on paid orders total three items: two on O101 and one on O102. O104 is also paid, but its quantity is blank. Therefore, three is only the known subtotal; the complete paid-order quantity is unknown. A report that confidently calls three the weekly total has hidden the most useful finding.

Expected checks · calculated from the fictional files
Data rows: orders 5; refunds 2; support 3.

Unique orders: 4.
Paid: 3 — O101, O102, O104.
  orders.csv lines 2, 3, 6.
Cancelled: 1 — O103, line 5.
Duplicate: O102, line 4 repeats line 3.

Known paid units: 2 + 1 = 3.
Complete paid units: unknown.
  O104 has no units, orders.csv line 6.

R201 matches O101.
  refunds.csv line 2; orders.csv line 2.
R202 has no matching order in this export.
  O999, refunds.csv line 3.

Open tickets: 2 — T301 and T303.
  support.csv lines 2 and 4.
T303 has no order ID, support.csv line 4.

Turn missing data into useful questions

A good report can be incomplete and still help someone act. It could end with: “Confirm the quantity for O104, check whether O999 belongs to an earlier export, and identify the order for T303.” Those are requests for evidence, not instructions to invent values or delete records.

If the assistant returns a wrong number, ask it to show the included IDs and arithmetic. If an answer changes, check what changed in its reasoning. Repeating the same conclusion with more confident wording does not resolve a disagreement.

  • A missing required file or column blocks the affected calculation. Report the gap and keep independently verified results separate.
  • Two records with the same ID but different values need a decision from the data owner. Do not silently take the first or last one.
  • An unmatched refund is a question about this export's coverage. It is not evidence of fraud, an invalid refund, or a broken business system.
  • Two open tickets do not measure customer satisfaction. A single reporting period cannot establish growth or a trend.

Keep the files and review together in Onevium

After configuring a model connection, use Add Project in Onevium's sidebar to open the practice folder. Create a conversation from that project row, confirm its directory, and send the task above. Keep the exercise scoped to the three fictional files.

When the answer arrives, use the conversation toolbar's Browse all files button to open the CSVs and check the cited rows. Keep corrections in that conversation so the report and its supporting discussion remain easy to find. The quickstart and file guide below explain these workspace controls. Use the answer key above to check the result of your own practice run.

For a later real-data pilot, use approved copies containing only the necessary fields. A local file location does not determine what information your configured model provider receives. Check that data use with your organization and provider before introducing customer information.

Further readingOpen your first project and conversation Browse project files and review their content

Decide whether this task earns a second week

Once the sample works, choose one existing report with an owner who can verify it. Preserve the team's real definitions, export window, and expected checks. Start with a draft that the owner reviews before sharing; resolving ambiguous records still belongs to the person who understands the data.

For a few comparable runs, record the time spent preparing exports, asking for the report, checking it, and correcting mistakes. Include model usage cost if available. Compare that complete effort with the previous process. A fast first draft is useful only if the checking and corrections still fit the team's day.

Keep the workflow if it repeatedly produces accurate, traceable results with manageable review effort. Narrow or pause it if missing evidence keeps getting disguised as a complete answer. The first useful outcome is modest: next week's reviewer can see exactly where every number came from.