Decide AI doesn’t want to be ChatGPT. It just wants to fix your spreadsheets

Spreadsheets can be humbling. They look simple on the outside, with their rows and columns. Type in a few numbers, arrange them in ascending or descending order, maybe use a summation equation to calculate a total, and that is ‘proficiency in spreadsheets’ going on my resume. 

Until the spreadsheet is one that has hundreds of rows and columns that spill across the screen, and then it starts to look like a puzzle.

Abiodun Adetona noticed problems with spreadsheets and analysis during his four years as a software engineer at Flutterwave, Africa’s largest payments infrastructure startup. His colleagues from non-technical teams often needed help drawing insights from datasets or spreadsheets, which often required technical intervention. 

“I used to assist them [his colleagues] with pulling data from various sources every day for almost a year… it was hectic, and there was no easier way to do it than the technical way,” he recalled in an interview with TechCabal.

In 2025, Adetona and his three-person team built Decide AI, an artificial intelligence (AI) spreadsheet analyst, to help users analyse data in spreadsheets via prompts.

Inside Decide AI

When I opened Decide AI for the first time, the interface felt really familiar, almost identical to modern AI assistants, such as ChatGPT or Gemini. The interface allowed me to choose between two AI agents to run my tasks.

The Fast agent is designed for quick analysis that prioritises speed, while the Pro agent is intended for heavier tasks that require deeper analysis, such as performing complex calculations across multiple datasets.

When I curiously clicked the ‘connections’ drop-down menu, I noticed that Decide AI is also designed to work with data that lives outside local spreadsheet files and allows for a connection to external sources such as Google Sheets, Metabase, Google Analytics, and Google Ads. 

Since I only have a Google Sheets account, I gave the AI agent permission to access my Google Drive upon request and connected my account.

Use cases for the other connections could be a user managing marketing campaigns connecting their Google Ads account and asking the agent to analyse campaign performance, or a team using Google Analytics asking the agent to analyse traffic patterns or identify trends in user behaviour.

Adetona said the agent can conduct basic tasks like cleaning datasets or performing calculations. It can also handle more analytical tasks, including scenario analysis, where users explore how different variables might affect a business outcome, and market size opportunity analysis, which involves estimating the potential value of a market based on available data.

How Decide AI analyses a spreadsheet

To see how Decide AI behaves with a real spreadsheet, I uploaded a dataset from a course registration programme from my Google Sheets. This spreadsheet contained entries from participants across different parishes and deaneries in Lagos, along with other details such as when participants registered and whether they had uploaded proof of payment. 

Using the Fast agent, I asked the system to clean the dataset, count the number of participants who paid in each month, exclude certain entries, and group the results by benchmark. 

Once submitted, the agent displayed a reasoning trace that appeared as lines of its activity in a darker, code-like font. That visible execution trace reflects what happens behind the interface when an analysis is requested.

According to Adetona, the system runs a preprocessing stage that attempts to interpret the file structure to identify sheets, headers, tables, formulas, and relationships between parts of the dataset, which are then converted into a structured representation that the system can analyse more reliably.

“The exact implementation is proprietary, but this step is critical because spreadsheet AI often fails when it misreads the structure of the file,” he said. He explained that once the document structure is mapped, it is passed to the AI agent along with the user’s request, which then plans how to carry out the analysis. 

The agent relies on frontier language models for reasoning, drawing on providers such as OpenAI, Anthropic, and Google. 

Adeotona said Decide AI is designed to be model-agnostic; the system can switch between models while its orchestration layer handles spreadsheet interpretation and computation.

The system generates Python code for spreadsheet analysis and calculations, and runs it in a secure sandboxed environment. 

“We do not rely on prompt reasoning alone for calculations because that is more prone to hallucination,” Adetona added.

He further explained that the generated code runs against the dataset, executing tasks like filtering entries, grouping results, calculating totals, and reorganising the spreadsheet. If the initial output does not fully answer the prompt, Adetona said the system can iterate through additional computational steps until it produces a complete result. 

Once those calculations are complete, the platform runs a verification stage that checks the output against the provided spreadsheet data before returning the final response.

For my prompt, the system produced a structured summary of the dataset in the chat window within two minutes. This summary highlighted the number of participants who had paid in January, grouped the results by deanery, and flagged registrations that did not include proof of payment. It also identified which deaneries had the highest number of incomplete payment records.

Along with the summary, Decide AI generated an Excel workbook that can be downloaded, but I exported it to my Google Sheets. This workbook contained five different worksheets derived from the original dataset. One sheet summarised January registrations, another compiled a breakdown of the data as I requested, another isolated entries where proof of payment had not been uploaded, and the others listed detailed records of participants whose registrations were missing documentation.

I pushed the analysis further by asking Decide AI to produce a more detailed report of the dataset. This time, I asked the system to calculate monthly totals and averages, compare participation between months, rank deaneries by total contributions, and identify the highest and lowest revenue sources. 

After close to four minutes of thinking, the AI agent generated tables and summaries. It further calculated percentages, estimated the potential revenue tied up in registrations that lacked proof of payment, and developed key insights based on the request. 

The final output included a reorganised Excel workbook and a written PDF report summarising the findings and highlighting patterns in the dataset.

How Decide AI makes money

Decide AI is still an early-stage product, but the company has already begun experimenting with ways to monetise the platform. The tool was initially available for free while the team focused on attracting users and understanding how people interacted with the system. 

Today, Decide AI operates on a hybrid pricing model that combines subscriptions with a pay-as-you-go option. A $7 payment allows users to send 20 prompts to the system, while the $15 package increases that allowance to 100 prompts monthly. 

For users who rely on the platform more heavily, the $25 monthly subscription tier offers no prompt limits and unlimited access to the system’s analysis tools.

Adetona said the platform has attracted between 3,000 and 4,000 users, with a smaller subset converting into paying customers since its February 2026 public launch. He noted that many of those users are experimenting with the system to analyse spreadsheets and financial reports.

Decide AI has also raised a $15,000 friends-and-family round, which helped support early development while the team worked on refining the product, according to him.

Decide AI appears to prioritise product development over monetisation. However, operating in the AI assistant sector can become expensive, particularly when systems rely on external models and cloud computing resources.

The limits of AI spreadsheet analysis

Like many AI systems, Decide AI is not immune to errors. Adetona acknowledged that the system does not always produce perfect outputs, because datasets often contain missing context or inconsistent formats that are not explicitly stated in the file itself. 

When those cases occur, Adetona explained that the team examines the system’s execution path to identify where the breakdown happened and then adjusts the system architecture accordingly, refining how the agent processes similar tasks in the future.

Large AI labs, including OpenAI, Google, and Anthropic, are already building systems capable of analysing spreadsheets and structured datasets, with products such as ChatGPT, Gemini, and Claude. Adetona believes Decide AI’s advantage lies in focusing narrowly on spreadsheets rather than attempting to build a broad AI assistant.

Against these broad AI assistants, Decide AI points to its performance on SpreadsheetBench, a benchmark designed to evaluate how well AI agents perform real-world spreadsheet tasks. SpreadsheetBench contains over 900 spreadsheet analysis challenges drawn from Excel workflows, and 400 expert-annotated instances. 

These tasks test whether an AI agent can correctly interpret spreadsheet structures, perform calculations, and produce accurate outputs. 

According to Adetona, Decide AI’s system achieved 82.5% accuracy on the SpreadsheetBench Verified benchmark, placing it in fourth place among the top-ranked agents globally.

As much as benchmarks are a measure of product performance, competing with large AI companies that are rapidly improving their analysis capabilities may prove a greater challenge for Decide AI than achieving a strong benchmark score.

Decide AI is still in its first month of operations as the team continues to refine the types of workflows the system can handle. An area of focus is deeper integration with the tools people already use to store and analyse data, including BigQuery, Meta ads, and allowing users to connect their own Application Programming Interface (API). 

Adetona says the goal is to allow the system to connect directly with business data sources to analyse spreadsheets and reports within existing workflows. Whether a three-person team can compete in a space where some of the world’s largest AI companies, such as ChatGPT and Microsoft Copilot, are building similar tools remains to be seen.

If spreadsheet analysis becomes just another feature inside general-purpose AI tools, Decide AI will need to prove that a specialised spreadsheet AI agent can produce better or more reliable analysis than tools embedded inside platforms users already rely on. This will ultimately determine if Decide AI becomes a niche tool for spreadsheet analysis or if its capability gets absorbed by the wave of general-purpose AI assistants.