Tech Giants' New Rival? Decide Shifts to In-House AI Development
Nigerian startup Decide has launched DAX-1, its first specialized AI model for spreadsheet automation, aiming to deliver faster and significantly more cost-effective solutions than larger general-purpose models. By training an open-source model specifically for tasks like fixing formulas and cleaning data, Decide seeks to become a critical infrastructure while expanding its business through increased efficiency. This strategic move allows Decide to enhance its product performance and reduce operational expenses for repetitive spreadsheet tasks.
Nigerian startup Decide, specializing in AI tools for spreadsheet automation, has unveiled DAX-1, its first proprietary specialized AI model. This development marks a significant shift for the company, which previously relied on a combination of larger general-purpose models from entities like OpenAI, Anthropic, and Google to power its product. Decide's strategic move aims to enable faster and more cost-effective spreadsheet editing, addressing the limitations of general-purpose AI models for highly repetitive, narrow tasks.
DAX-1 is specifically engineered to handle common spreadsheet operations such as fixing formulas, cleaning duplicate entries, correcting errors, and formatting workbooks. According to Abiodun Adetona, Decide’s co-founder, the ambition is to establish Decide as a critical infrastructure for spreadsheet automation across the internet. This specialized approach counters the inherent cost inefficiencies of using versatile, but often overkill, general-purpose models for straightforward, recurring tasks.
The decision to build DAX-1 stemmed from the understanding that while general-purpose AI excels at a broad range of tasks, its cost can escalate rapidly when applied to high-volume, repetitive functions. Rather than developing a model from scratch, Decide initiated its development with an existing open-source model from Qwen, Alibaba’s AI research arm. This base model was then rigorously trained and fine-tuned specifically for spreadsheet editing. Through this specialized training, DAX-1 significantly improved its accuracy, moving from an initial 9.17% on Decide's internal benchmark to an impressive 91.7% post-training.
Decide emphasizes that the production version of DAX-1 focuses on deterministic spreadsheet editing, meaning it produces executable edits rather than merely suggesting actions to users. Internal testing has demonstrated DAX-1's superior performance on its designated spreadsheet editing tasks compared to several larger models, all while offering substantial cost savings. Notably, Decide reports that DAX-1 is five times faster than Anthropic’s Fable 5 for the tasks it was designed for and operates at 97% less cost. While these benchmarks were internally conducted, Decide has published its methodology for verification. Furthermore, Decide's broader AI agent previously scored 82.5% on SpreadsheetBench Verified, ranking fourth at the time of evaluation.
This shift to a cheaper, more efficient AI model is seen by Decide as a business advantage, not a detriment. Adetona likens it to the Jevons paradox, an economic theory suggesting that increased efficiency of a resource can lead to greater, rather than reduced, overall consumption. By making spreadsheet automation more affordable to run, Decide can offer greater value to existing subscribers, scale its operations, and potentially unlock new demand by making the technology accessible for broader use cases. This strategy highlights a complex question facing AI companies: the goal is not always to maximize per-task cost but to make technology cheap enough to significantly increase its adoption and usage.
Decide's adoption of DAX-1 does not signify an abandonment of external model providers like OpenAI or Anthropic. Instead, the company is pursuing a