SQL shows the drop. Does the document explain it?
A SQL result can establish that revenue fell. A supplier memo can describe a delay. Putting both into an AI answer does not establish that the delay caused the decline.
Pablo’s “Chatting with your Database” prompted this example. The session streamed October 11, 2026; its archive upload metadata is October 12 Eastern time. Its description proposes combining BigQuery metrics with document retrieval. We reviewed the description, not a loaded transcript, and did not reproduce the implementation. No endorsement is implied.
Establish the measured change
Use a fictional table with one monthly revenue row per product. Amounts are integer cents in one currency. Assume the two months cover the same population and are complete under the same accounting definition. In real data, verify those assumptions, including refunds and recognition timing, before comparing totals.
SELECT product,
SUM(CASE WHEN month = '2026-08' THEN revenue_cents ELSE 0 END) AS august_cents,
SUM(CASE WHEN month = '2026-09' THEN revenue_cents ELSE 0 END) AS september_cents,
SUM(CASE WHEN month = '2026-09' THEN revenue_cents ELSE -revenue_cents END) AS delta_cents
FROM sales
WHERE month IN ('2026-08', '2026-09')
GROUP BY product
ORDER BY product;
Product A falls from $100 to $70, a $30 decline. Product B rises from $100 to $110, a $10 increase. Total revenue therefore falls from $200 to $180: down $20, or 10%. A’s own decline is 30%; that is not the company-wide percentage. The product deltas reconcile to the total, but they do not explain their causes.
Read the memo as evidence with a scope
Now separately supply a fictional memo: “September 18: Product A deliveries are delayed. We expect normal deliveries next week.” The product and month overlap with A’s decline. That makes the delay a hypothesis worth investigating.
The memo does not say which orders were affected, whether inventory ran out, whether customers cancelled, or when revenue was recognized. It cannot by itself explain all of A’s $30 decline, much less the net company decline after B’s increase.
Record the memo’s author, date, product scope and exact passage. Distinguish a contemporaneous observation from a forecast or a later explanation. If another memo claims “company revenue fell 30%,” flag that statement as inconsistent with this fixture’s 10% result. Do not silently substitute the product percentage.
Ask what would distinguish competing explanations
The same revenue rows could also accompany a memo saying that A’s discount campaign ended. Revenue alone does not distinguish delayed fulfilment from a change in prices, order volumes or demand. The next investigation needs appropriate records: promised and actual delivery dates, stockouts, cancellations, quantities, prices and campaign history.
Even those records may establish an association rather than a causal effect. State what the evidence supports and what remains unknown. A useful answer here is: “Revenue fell 10%. A contributed a $30 decline, offset by B’s $10 increase. A supplier memo reports a delay affecting A; whether that delay caused the decline is unverified.”
What we tested
The offline Python/SQLite fixture passed seven checks: product deltas, aggregate totals, company percentage, product percentage, reconciliation, identical query results under two supplied explanations, and SQLite query-only write refusal. The two explanations are authored examples; the test does not evaluate document retrieval or prove causal inference.
We did not run this through BigQuery, Bufflehead’s UI/MCP interface or a live AI model. The query-only check applies to the disposable SQLite connection. This is a deterministic SQL example, not an AI accuracy benchmark.
Keep the answer’s evidence visible
Ask your AI to separate the measured result, the document’s claim, conflicting evidence and the next question. Bufflehead provides read-only SQL access to existing structured data. Preparing tables, retrieving documents and assessing causality remain separate responsibilities; Bufflehead does not supply a document-retrieval or causal-analysis pipeline.
Try Bufflehead with a known-answer query. Keep the SQL result alongside any document passage so a plausible explanation does not become a measured fact.