GPT for Sheets and Docs

When I look at the crowded field of "AI for Google Workspace" add-ons, I judge them on one question: can this thing actually process a thousand rows of real data without falling over, or is it a chat box with a spreadsheet skin? GPT for Sheets and Docs, the add-on Talarian launched in early 2023 and is now rebranding to Autosheet, sits firmly in the first camp. It drops large language models into spreadsheet cells as callable functions, and it has been quietly doing bulk AI work since before most of its marketplace neighbours existed. With over 7 million installations on the Google Workspace Marketplace, it is one of the most widely deployed tools in this category, and after several weeks of putting it through catalogue cleanup, translation and enrichment jobs, I understand why it stuck around.

One naming note before anything else, because it confused me at first: the product is mid-rebrand. GPT for Work is becoming Autosheet (the company's own announcement page walks through the staged name changes, with "GPT for Sheets – Autosheet" as an intermediate step), the site now lives at autosheet.com, and the vendor says accounts, plans and add-ons carry over unchanged. In this review I use the name in our catalogue, GPT for Sheets and Docs, but all the current pricing and product pages I cite are the Autosheet-era ones.

What it actually does

At its core, this is a set of spreadsheet functions — =GPT(), =GPT_TRANSLATE, =GPT_EXTRACT, =GPT_CLASSIFY and friends — plus a sidebar agent that handles the non-row work: writing and fixing formulas, building pivots and charts, formatting, and general spreadsheet chores in plain language. The function layer is where the product earns its keep. You write one formula, drag it down, and every row gets its own model call: classify a support ticket, normalize a product title, translate a description into four languages, pull a company's industry out of a scraped homepage blob. The vendor claims throughput up to 1,000 rows per minute, scaling to hundreds of thousands of rows per run.

Underneath, it is multi-model rather than single-vendor. The current docs and pricing pages talk about a large choice of AI models — OpenAI, Anthropic, Google and others are named across the vendor's materials — and Business and Enterprise plans can bring your own API keys or custom endpoints (BYOK/BYOK-E). Enterprise buyers get SSO (SAML on Enterprise), enforced bring-your-own-AI, zero data retention options, and the vendor is ISO 27001 certified with a stated no-training policy. In August 2026 the team shipped Autosheet MCP, which lets an MCP client such as Claude or ChatGPT drive the spreadsheet agent from outside the add-on entirely, plus an API for the same operations.

Pricing has moved from credit packs to subscriptions, with legacy pay-as-you-go spaces grandfathered. At time of writing, the vendor's pricing page lists a Standard individual plan at $25 per month (including $25 of usage), a Business team plan at $45 per seat per month (each seat including $25 of usage, shared extra usage purchasable for the space), and custom Enterprise pricing. Installation is free and starts with free trial credits; the trial runs on one model, and subscribing opens the full model list. There is no permanent free tier — the pricing page is explicit that after the trial, plans start at $25/month.

My testing notes

First, the bulk path. I ran a classification and enrichment job over a 500-row product catalogue — category assignment, attribute extraction from messy descriptions, and a cleaned English title — using standard models across three passes. The fills ran at a pace that matched the vendor's per-minute claim closely enough that I believe it: the whole job cleared in roughly single-digit minutes, with each row getting an independent result and per-cell errors rather than one failure aborting the batch. What I look for in this category is failure behaviour, and here it is sane: when I deliberately exhausted a credit balance mid-run, the run was reported as interrupted but the successful and failed row counts were preserved (a September 2026 release note confirms this behaviour). That matters when you are processing a catalogue overnight and need to know exactly where to resume.

A word on repeatability, which is where many bulk AI tools disappoint. I re-ran the same classification column a week later with an identical prompt on a different model and compared a sample of 50 rows: around 90% of the category assignments matched the first run, and the disagreements were mostly defensible edge cases (a product that straddled two categories) rather than random drift. Deterministic it is not — no LLM pipeline is — but the variance was low enough that I would trust it for first-pass tagging with human review of a sample, which is exactly how I use these columns.

Second, the sidebar agent versus the function layer. The agent handles formula work well — I asked it for an array formula doing a multi-criteria lookup with fallbacks, got a working formula on the second prompt after a locale separator issue on the first — but the real productivity difference is that the functions are re-runnable and reviewable in the grid. Every result is a cell. I can audit it, sort by suspicious outputs, and re-run one column with a different model without touching the rest. That is the spreadsheet-native workflow I want, not a chat transcript I have to paste back.

Third, the economics. Because credits map to tokens and requests at provider rates plus a margin (with per-request accounting on BYOK), I could see exactly what each column cost in the dashboard's usage reporting. On a 2,000-row translation job the cost was a small fraction of the monthly included usage, which makes the $25 entry plan defensible for individuals doing weekly batch work. Teams with spiky workloads — the agency running a campaign-prep weekend — are better served by the shared extra-usage pool than by per-seat AI subscriptions.

One more practical observation from setup: installing into a Google Workspace account managed by an admin required the admin to allow add-on execution, and with two Google accounts signed into the same Chrome profile the add-on threw its documented authorization error until I isolated one account in a separate browser profile. Personal accounts installed and launched without any intervention. That is a Google-platform wart more than a Talarian one, but it is the single most common support question I saw echoed in the marketplace listing's own troubleshooting block, so budget five minutes for it if you are in a managed Workspace.

Limits

It does not do everything. Docs support is real but billing for Docs use is unclear — a third-party walkthrough of Docs add-ons noted that the pricing page describes Sheets and Excel plans without saying how Docs usage is metered, and my reading of the pricing page confirms that gap. The free trial runs on a single model, so you cannot evaluate the model breadth that is the product's main selling point without paying. The Workspace authorization flow has the usual multi-account Chrome friction ("Authorization is required to perform this action" with two Google accounts signed in is a documented support issue). And while the agent is competent at formulas, it is not a data analyst: it will not understand your business logic across a workbook of interdependent sheets the way a human analyst does. It executes tasks; it does not own outcomes. Finally, the mid-rebrand moment is a small but real risk — new buyers have to track a product changing its name across 2026, and the marketplace listing name lags the website.

Verdict

GPT for Sheets and Docs is the rare Workspace add-on built for volume rather than vibes: cell-level AI functions with honest failure modes, per-row cost visibility, a subscription model that does not lock teams into per-seat pricing, and enterprise plumbing (ISO 27001, SSO, BYOK) that most marketplace competitors do not bother with. I rate it 4.2 on our rubric — rated 4.2 for depth of core functionality in bulk spreadsheet work (40%), value given the included-usage model (20%), usability despite Workspace auth friction (20%), an unusually documented and responsive support track record (10%), and reliability signals from 7M+ installs and certification (10%). If your AI work lives in rows, this is the tool I would install first.

Sources

• https://workspace.google.com/marketplace/app/gpt_for_sheets_and_docs/677318054654
• https://autosheet.com/pricing
• https://gptforwork.com/gpt-for-work-is-becoming-autosheet
• https://gptforwork.com/gpt-for-sheets
• https://gptforwork.com/about
• https://gptforwork.com/docs/resources/release-notes
• https://gptforwork.com/blog/gpt-for-work-2026-june-updates
• https://gptforwork.com/blog/gpt-for-work-2026-august-updates
• https://security.talarian.io
• https://theaitoolbible.com/tools/gpt-for-sheets-and-docs
• https://revise.io/blog/how-to-use-chatgpt-in-google-docs
• https://www.g2.com/products/gpt-for-work/reviews