Hello, this is issue 0 of Neurobuttr, a sample of what lands in your inbox every Sunday. Issue 1 arrives on Sunday 8 November.

Each week I pick one way artificial intelligence (AI) can make your data work easier, explain it in plain English, and give you one thing to try in 15 minutes.

The one thing

Google now lets people ask questions of company data in BigQuery, its cloud data warehouse, in plain English from a Google Sheet. Google announced it on 19 October, and the BigQuery release notes marked it fully released on 27 October. In a sheet connected to BigQuery, you open the panel for Gemini, Google's AI assistant, and type a question. Gemini writes the SQL (Structured Query Language, the code that pulls data from databases), runs it and sums up the trends. It can also build a tab of charts and pivot tables that refresh on a schedule.

It is on by default for several paid Google Workspace plans, with no admin setting to control it. And it queries the whole BigQuery dataset, not just the rows in the sheet, for anyone allowed to query that data.

What it means for your work: people no longer need you to write the query. Your job shifts from writing SQL to making sure each number means one thing. If "sales" includes refunds in one table and not another, Gemini has to guess. Google's guide to conversational analytics warns that Gemini can give answers that sound right but are wrong, and says to check every output. It also says clear table and column descriptions help.

What to do: if your company uses Workspace and BigQuery, expect someone to show you a Gemini number soon. Add plain descriptions to key tables now. Nobody knows yet how it copes with messy tables, so test it on three numbers you already trust. Google's announcement does not say whether you can see the SQL behind each answer, so check that too.

What to safely ignore: if your team runs on Excel and Power BI, this feature will not reach you. The lesson about definitions still applies.

Try this

A good data prompt names the table, defines the terms and asks the tool to show its assumptions. Give this 15 minutes in the free version of ChatGPT, Claude or Gemini. First type only "What were sales last quarter?" Then paste the prompt below, which uses a made-up table, and compare the two answers.

You are helping me write SQL for BigQuery.
Table: shop.orders (made-up example)
Columns: order_id, order_date (DATE), region (STRING),
amount_gbp (NUMERIC), status ('paid' or 'refunded')

Definitions:
- Sales = sum of amount_gbp where status = 'paid'
- Last quarter = 1 July to 30 September 2026

Task: one query showing sales by region for last quarter,
highest first. Then list every assumption you made and
one way I could check the result.

Notice what the first answer had to guess. Never paste real work data or customer details into a public AI tool.

Word of the week: prompt

A prompt is the instruction you give an AI tool: your question plus any context, rules and the format you want back. The tool only knows what you tell it, so a vague prompt gets a guess. "Show me sales" is a weak prompt. "Sum paid orders in shop.orders for July to September 2026, by region, and list your assumptions" is a strong one. It is like writing a clear brief for a colleague.

Three quick ones

• On 22 October Anthropic, which makes Claude, gave its Agent Skills standard to the Agentic AI Foundation, part of the Linux Foundation; a skill is a folder of instructions that teaches an AI agent (a tool that carries out tasks step by step) one job.
• dbt Labs, maker of the data transformation tool dbt, ran a small test where feeding an AI agent sales call summaries from the warehouse, not full transcripts, roughly halved the cost of each run, but the shortest summaries lost dates, numbers and decisions.
• On 20 October Databricks launched Agent Memory in beta, which lets AI agents on its platform remember facts and user preferences between conversations.

The UK number

30%. That is the share of UK businesses using at least one kind of AI in September 2026, up from 9% when the Office for National Statistics first asked in September 2023. In information and communication firms it was 60%. With use up more than threefold, the useful question at work is how to use it well. Source: Office for National Statistics, 8 October 2026

Your turn

If someone in your team could ask your data a question in plain English tomorrow, which number would you most worry about them getting wrong, and why? Hit reply and tell me.

Ishan

Know someone in data who would find this useful? Forward it. They can sign up at ishanyash.com/newsletter

Your one-pager to keep: 20 AI words for data people