Query Google Sheets with Natural Language: An n8n Template Walkthrough

Small and midsize teams almost always have one critical spreadsheet: a customer list, project tracker, or order table. All the answers are in there, but pulling them out usually means opening the sheet, filtering, sorting, maybe building a pivot table. There's a popular workflow in the n8n template library that puts a chat box in front of Google Sheets so team members can just ask questions in plain English.

The problem: the answer is in the sheet, but getting it out takes 15 minutes

Picture a Monday morning standup. A partner asks, "Which clients are at risk?" "Who's renewing in November?" Someone has to open the sheet, filter, sort, maybe build a pivot table, then paste the result into chat. Account managers repeat this a dozen times a week. And the person asking is often the least familiar with the filter menu, so they either wait for someone to reply or make a decision from memory.

The real cost isn't the spreadsheet itself, it's the interruptions around it. Every "quick question" pulls someone out of their work, and by the time the answer arrives it's already stale.

The shift: ask the sheet like you'd ask a colleague

Template number 7639 in the n8n template library puts a chat box in front of a Google Sheet. You type your question in plain English, an AI agent reads the rows, and you get an answer in seconds. On the sample customer list used for teaching, these questions work well:

• "Which clients are At Risk?" returns Northwind Insurance and Riverbend Accounting.
• "Which clients renew in November?" returns Atlas Logistics Co and Northwind Insurance.
• A follow-up like "Which of those two is at risk?" also works, because the template includes a memory node that remembers the conversation.

No filters, no formulas, no waiting on a colleague. The partner asks, the sheet answers.

The AI agent in n8n template 7639 is a tool-calling agent. Tool calling means the large language model can, while generating a response, decide on its own to invoke an external function or data source and fold the returned result into its reasoning. In this workflow, the Google Sheets Get Row(s) operation is registered as a tool named Analyze Data. When the agent sees a question like "Which clients are At Risk," it first decides it needs to read the sheet, issues a Get Row(s) call, and only then composes a natural-language answer from the returned rows.

This matters: the model doesn't "know" the sheet contents on its own, it just looks them up when needed. That means the agent's capabilities are bounded by how many tools you attach to it. This template only attaches the read-only Get Row(s), so it can read but not modify. If you also attached write operations like Append or Update as tools, the agent would call them too, and the risk would rise accordingly.

How it's built (five nodes)

The template has five working nodes:

1. Chat with Your Data — the chat trigger that receives your messages.
2. Talk to Your Data — the AI agent, which decides how to handle the question.
3. OpenAI Chat Model — the brain the agent thinks with.
4. Memory — stores the conversation so follow-ups make sense.
5. Analyze Data — a Google Sheets tool set to Get Row(s), which the agent uses to read the sheet.

Note that Analyze Data is read-only. The agent in this template cannot modify your sheet, which is exactly what you want the first time you deploy it on real customer data.

The Memory node in n8n is used to preserve context across multiple conversation turns. Without it, each question is isolated, and the agent can't understand what "those two" refers to in "Which of those two is at risk?" Memory is typically stored partitioned by session ID, and the chat trigger assigns an identifier to each conversation, so follow-ups from different users don't bleed into each other.

Worth noting: Memory only stores conversation history, not sheet data. Each time you ask a question, the agent still has to call Get Row(s) again to read the sheet, so the latest data is always visible, but the context window gets consumed by historical messages. If a conversation runs long, you can configure it to retain only the most recent few turns, avoiding prompt bloat that drives up cost and latency.

Setup: two connections

A template can't carry anyone's keys, so credentials come up empty when you open it. Two steps finish the job.

Connect OpenAI. Open the OpenAI Chat Model node and select your OpenAI credential. If you don't have one, create a key on the OpenAI API Keys page, top up under Billing, and paste the key into a new credential in n8n. The template uses gpt-4.1-nano, the cheapest model, so practice costs are essentially negligible. If the model list won't load, switch the little "From list" box to "By ID" and type the model name in manually.

Connect your sheet. Open Analyze Data, select the Google Sheets credential for the Google account that owns the spreadsheet, then choose the Document and Sheet tab. Two easy traps: credentials from a different Google account won't list your spreadsheet, and the sheet needs column names in row 1 with data starting at row 2. Click Execute step. If the output panel shows one record per row with column names attached, the agent will be able to read it too.

gpt-4.1-nano is OpenAI's small model aimed at high-throughput, low-latency scenarios, positioned similarly to the earlier gpt-3.5-turbo or gpt-4o-mini: low unit price, fast, well-suited to tasks like tool calling and structured extraction that don't demand deep reasoning. Running a read-the-sheet Q&A like the one in this article on it typically costs under a cent per call, essentially negligible during practice.

The trade-off is weaker performance on complex reasoning and long contexts than larger models, so when questions involve multi-step calculations or cross-sheet joins, switching to a stronger model is often more effective than rewriting the prompt. n8n's OpenAI Chat Model node lets you toggle between "From list" and "By ID": list mode pulls available model names from OpenAI and may come back empty if there's a network or account-permission issue, in which case switching to "By ID" and typing the model identifier manually gets around it.

The gotcha you must know: verify your totals

This is the part I'm glad I actually tested rather than assumed. I asked the agent to total the monthly fees of one account manager's Active clients. It listed the correct three rows: 2,500, 1,800, and 1,900. Then it said the total was 5,200. The real total is 6,200.

Language models are good at finding the right rows and unreliable at arithmetic. The model can also vary from run to run, so my answer might be wrong and yours might be right, or vice versa. The rule for business is simple: use the agent to find and list, and add up yourself (or with a spreadsheet formula) before repeating a total to a client or writing it into a report. If you have the agent show its rows, verification takes ten seconds.

Counting and listing questions ("who," "which," "when") are the sweet spot. Monetary totals deserve a human check.

The root of language models' arithmetic unreliability is that they process tokens, not numbers. The "2500" the model sees gets split into a handful of tokens, and it isn't doing decimal addition, it's predicting the next token based on patterns seen during training. So the model can accurately "copy" amounts out of a sheet, yet stumble on sums, averages, and percentages, and the same question can produce different results across runs.

There are three categories of mitigation: first, have the agent list the rows it used so a human can verify in ten seconds; second, offload the summation to a spreadsheet formula or a dedicated Calculator tool, leaving the model responsible only for selecting rows and not for arithmetic; third, explicitly require in the prompt that "any monetary total must include the calculation steps." But the safest approach remains treating AI as a find-and-list assistant, with the final numbers confirmed by a human or a deterministic program.

Cost and expectations

• Setup time: about 8 minutes once you have keys.
• Running cost: the nano model is very cheap; typical questions cost very little.
• Risk: low, because the agent is read-only. Don't share the chat with anyone who shouldn't see the full list, since the agent can read every row it's pointed at.

To let other people use it, you need to publish the workflow. That's a separate step, and it's worth deciding who gets access before you publish.

Where to practice

The workflow above has 22 steps across three parts: connect OpenAI, connect the sheet, ask questions of your data. The author turned it into a free interactive course that runs on a practice copy in the n8n editor with fictional data, so it never touches a real account and never incurs charges.

Full course here: n8n: Talk to Your Google Sheets with ChatGPT.

If you already have a customer list, task list, or order table, this is one of the fastest ways to make it queryable by the whole team. Start with the sheet you get asked about the most.

One-line Summary

n8n template 7639 uses a five-node AI agent to let teams query Google Sheets in plain English, read-only, low-cost, buildable in eight minutes.

Paragraph Summary

This article covers workflow number 7639 in the n8n template library, which puts a chat box in front of Google Sheets so team members can ask questions in plain English about customer lists, renewal dates, and more without filtering, formulas, or waiting on colleagues. The workflow has just five nodes: chat trigger, AI agent, OpenAI model, memory node, and a read-only Google Sheets tool. Setup only requires connecting OpenAI and the sheet, about eight minutes total, using gpt-4.1-nano at very low cost with a read-only agent for contained risk. The author specifically warns that language models are unreliable at arithmetic, with a real test misreporting a total of 6,200 as 5,200, so monetary totals must be human-verified. Counting and listing questions are the AI agent's sweet spot, making this a good starting point from the sheet you're asked about most.