A startup can build an AI product without spending years training a model from scratch. It can plug into an existing model, build a product around it, and focus on getting customers. But using someone else’s intelligence can get expensive.
Decide, the Nigerian startup building AI tools for spreadsheet automation, has now taken a step towards building some of its own. The company has introduced DAX-1, its first specialised AI model designed to edit spreadsheets faster and more cheaply than the larger models it previously relied on.
The startup had been using a mix of models from companies including OpenAI, Anthropic, and Google to power different parts of its product. Now, it has built a model of its own. DAX-1, Decide’s first specialised AI model, is designed to handle tasks such as fixing formulas, cleaning duplicates, correcting spreadsheet errors, and formatting workbooks.
“Our goal is to be a critical infrastructure for spreadsheet automation operations in general on the internet,” Abiodun Adetona, Decide’s co-founder, says.
A smaller model for a smaller problem
General-purpose AI models are designed to do almost everything. They can write, reason, analyse documents, answer questions, and increasingly perform complex tasks across different tools. That versatility is useful, but it can also be unnecessary.
A spreadsheet user asking an AI to clean duplicate entries or fix a broken formula does not necessarily need a model capable of writing a research paper or generating a marketing strategy. And when those narrow tasks are performed repeatedly across thousands of users, the cost of relying on a large model can quickly add up.
“There’s a place where general-purpose models fall short, and that’s in the place of cost,” Adetona says. “For some repetitive tasks that require doing them over and over again, they can get very expensive quickly.”
Instead of training a model from scratch, Decide started with an existing open-source model from Qwen, Alibaba’s AI research arm, and trained it specifically for spreadsheet editing. The company is not necessarily trying to build the next GPT or Claude. It’s taking on a much narrower bet.
According to Adetona, the base model Decide started with achieved 9.17% accuracy on the company’s benchmark before the additional training. DAX-1 achieved 91.7% after the work was completed.
The company says the production version of DAX-1 is designed for deterministic spreadsheet editing and produces executable edits rather than simply suggesting what a user should do. It can handle tasks including formula fixes, lookups, duplicate cleaning, and table repairs.
Its internal testing found that DAX-1 performed better on its specialised spreadsheet editing tasks than several larger models while being significantly cheaper to run. According to Decide, the model is five times faster than Anthropic’s Fable 5 on the task it was designed for and 97% cheaper to run. Adetona says Decide also tested it against models from OpenAI, Google and DeepSeek, with larger models performing strongly but at a higher cost for this particular type of work.
The company acknowledges that the benchmark was conducted internally and has published its methodology and evaluation approach so others can verify the results.
However, Decide had previously been independently assessed against other spreadsheet AI agents. The company’s agent scored 82.5% on SpreadsheetBench Verified, placing it fourth on the leaderboard at the time of its evaluation.
Cheaper AI could mean more business, not less
For a business that earns money from subscriptions or usage, making tasks cheaper could appear to create a problem. If users can do more for the same amount of money, doesn’t that mean the company is giving away more while earning the same?
Adetona doesn’t think so. Because DAX-1 is cheaper for Decide to run, he says users can get more value from what they are already paying for. At the same time, the company has more room to serve more customers and scale its operations.
“It’s good business for us,” he says. “We want to satisfy users, and satisfied users will likely bring more people in.”
He compares the idea to the Jevons paradox, an economic theory that suggests making a resource more efficient can sometimes increase rather than reduce its overall consumption. His view is that lower costs could make spreadsheet automation easier to use at scale, creating more demand rather than simply reducing the amount Decide earns from existing customers.
It is a theory that Decide will have to prove as it grows. But it points to one of the more complicated questions facing AI companies today. The goal is not always to charge more for every AI task. Sometimes the bigger opportunity is to make the technology cheap enough to be used far more often.
Decide is changing how it thinks about AI
DAX-1 does not mean Decide is abandoning OpenAI, Anthropic, or other model providers. Adetona says the company will continue to use external models where they make sense.
“We can build agents and provide solutions in those areas,” he says. “And we can also build models that provide intelligence.”
For much of the current AI boom, startups have had to decide whether they want to build models or build products on top of them. The first option is expensive and difficult. The second can leave companies dependent on technologies they do not control.
Decide is taking a middle path. It is still using the best models available for broader and more complex tasks, but it is starting to own the parts of its technology stack where specialisation can improve the economics and performance of its product.
That approach may ultimately make more sense for a company focused on a specific type of work like spreadsheets than trying to compete directly with companies building models for everything. DAX-1 is Decide’s first model, but it is unlikely to be the last if the strategy works.
