Basis saves accountants 30% of time using OpenAI's GPT-5

Basis, founded in 2023, isn’t chasing shiny demos. They’re building accounting agents that actually work, and they’ve locked in up to 30% time savings for top firms by automating the grunt work—reconciliations, journal entries, financial summaries—that eats up billable hours. The real story isn’t just the time saved; it’s the architecture. Basis treats accounting as a system of workflows managed by a multi-agent setup, and they’re betting heavily on OpenAI’s latest models to make it hum.

They’ve been working with OpenAI since day one, co-founder Mitchell Troyanovsky says, and they just migrated their supervising agent from o3 to GPT-5. That’s a significant jump. The team calls GPT-5 the strongest model they’ve evaluated for reasoning, consistency, and explainability. It acts as a traffic controller, routing tasks to specialized sub-agents and selecting the right model for the job—GPT-4.1 for fast, mid-review clarifications, and GPT-5 again for heavy lifting like interpreting weird transactions or managing a month-end close. An internal benchmark suite scores models on specific capabilities, so the switch isn’t faith-based; it’s data-driven.

But automation in accounting is useless if it’s a black box. No one’s signing off on a journal entry they can’t interrogate. Basis gets this. Their agents surface the assumptions, data sources, and logic behind every decision. “Everything we do depends on reasoning,” Troyanovsky notes, pointing to how GPT-5 scales test-time compute while still exposing its chain of thought. That visibility gives accountants actual control rather than a magic number they’re supposed to trust. Function calling has also pushed the system from proposing actions to executing multi-step processes directly, mirroring how a human accountant would think through the work.

What’s clever is that Basis is built as a system that compounds improvements. Each new model release isn’t just a patch; it expands the frontier of what agents can safely handle. The team runs detailed benchmarks on real-world accounting workflows, evaluating accuracy alongside explainability. GPT-5’s standout performance in parallel tool calling—coordinating multiple structured actions in one go—is exactly the kind of capability that moves the needle from a cool demo to a tool that eats into the service backlog. The question isn’t whether AI can do accounting tasks; it’s how quickly the scope of autonomous work expands from here.