Sourcetable heralds 'self-driving' spreadsheet

Spreadsheet and data solutions provider Sourcetable launched a "self-driving" spreadsheet that allows users to simply tell the spreadsheet what they want done through natural language commands.

Sourcetable developed the solution as a way to bring advanced spreadsheet functionality to people who might struggle with basic functions like VLOOKUP or creating a pivot table. The "self-driving" autopilot capabilities give the AI complete write access and edit control to complete multi-step operations.

"AI is the biggest platform shift since the browser, with a bigger opportunity for disruption," said Sourcetable CEO and co-founder Eoin McMillan. "Sourcetable is building the AI spreadsheet for the next billion users, be they human or AI. As AI makes analysis easier, everybody will become an analyst. Sourcetable's AI automation ushers in a new era of productivity and human cognition."

Robot spreadsheet AI
Summit Art Creations - stock.adobe.com

Sourcetable's autopilot mode can complete a wide range of complex tasks, including creating and editing financial models, generating spreadsheet templates, building pivot tables, cleaning data, creating charts and graphs, editing formatting, enriching data and analyzing entire workbooks. The AI can understand data context without requiring users to pre-select ranges, interpret multiple ranges across different tabs, work with messy data, and seek human clarification when instructions are unclear.

The AI is capable of accessing anything that is publicly available on the internet, McMillan said in an email, and it can also extract data from URLs if instructed to do so. This includes Federal Reserve Economic Data, stock ticks and trading data, Yahoo finance, futures, geopolitics, market sentiment, macroeconomic analyses, Wikipedia data and much more. "There's even a full fund manager Easter egg included in this release," he added. 

This ability to access tools outside itself also means that users, via a virtual machine with hundreds of libraries and AI tools available, can ask the autopilot to find a more advanced tool to serve their needs by requesting the system to "download data" or "use Python" to solve a task. McMillan said Sourcetable plans to make this feature more user-friendly in the future as the technology ultimately moves toward becoming a full agentic platform and operating system. 

To discourage the AI from providing false information, the solution is built around a code-driven evaluation loop developed internally that verifies AI response in real time. Without this foundation, according to Sourcetable, self-driving spreadsheet automation would be too slow and unreliable to be trusted. McMillan said the company uses a combination of techniques to optimize results while minimizing latency. First, there's AI-driven process supervision of inputs, outputs and prompts, effectively AI watching AI. This is combined with a code-driven audit of quantitative outputs (e.g., Python, SQL and spreadsheet output evaluation) and, finally, thought-driven techniques (e.g., Chain of Thought Reasoning and Deliberate Reasoning) to drive better results, particularly for multi-step processes. 

The new solution uses not one but many models to deliver results. While certain companies are locked into their own proprietary AI models, Sourcetable's AI selects the optimal model for each task–including OpenAI, Anthropic, Groq, Meta (Llama), Nvidia, Prior Labs, DeepSeek and Hugging Face—and even combines multiple models for better results. McMillan explained that different models are better suited to different tasks and run better on different kinds of hardware. For example, he noted, Claude is currently best at coding, TabPFN at interpreting tabular data, Groq at fast inference, etc. Sourcetable's AI knows model specifications and strengths, so i's able to understand what a user is trying to do and find the best tool for it. 

While accessing public models can sometimes come with a per-prompt cost, McMillan said the company has established relationships with many service providers to ensure high rate limits and the ability to handle a large number of requests. He added that, right now, Sourcetable use a combination of manual and automated controls to prevent abuse of the system that could conceivably create large fees, though he believes the long-term cost curve indicates that AI will essentially become free, with the price of software being more aligned with value than cost of goods sold.

Prior to this release, Sourcetable did offer an AI copilot similar to many in the market that was more for formula assistance, charting and answering questions, according to McMillan. This was initially included as a SQL assistant to retrieve database data to help users who didn't know how to write SQL, and this is how the company learned that users really wanted to use the AI for their regular spreadsheet workflows, leading Sourcetable to develop this current solution. 

"Ironically, solving the database retrieval problems forced us to build our own Chain of Thought equivalent before OpenAI released theirs publicly," said McMillan. "That taught us how to leverage processes like CoT for multistep processes and automation, and this gave us a big head start once we shifted gears toward full spreadsheet automation via AI. Today's autopilot moves us from answering questions to thinking and agency. It's a big leap forward."

Sourcetable offers both a free tier and a pro tier, which costs $20 a month. All Sourcetable users get the first two weeks free on the Pro tier and can continue using the system on a rate-limited free tier. All the regular spreadsheet and charting features are free and unrestricted. McMillan added that Pro users are Sourcetable's revenue source. Free tier users generate no revenue, he said, "although happy users spread the word, which is the best form of marketing."

For reprint and licensing requests for this article, click here.
Technology Artificial intelligence Data visualization Data Analytics Hardware and software
MORE FROM ACCOUNTING TODAY