Scaling accounting capacity with OpenAI

一些初创公司利用人工智能（AI）解决特定时点的问题，而另一些则构建随着AI进步而不断优化的系统。Basis（https://www.getbasis.ai/）属于后者。

Basis成立于2023年，专注于为顶级会计师事务所打造AI代理，旨在承担结构化的会计工作，满足这些任务所需的可靠性和深度。团队使用OpenAI的o3、o3-Pro、GPT-4.1和GPT-5模型，驱动AI代理帮助会计师事务所自动化重复性任务，如对账、分录和财务摘要，同时确保会计师能够全面了解决策过程并掌控流程。这样可节省高达30%的时间，并提升处理高价值工作的能力，如客户咨询和拓展新业务。

随着OpenAI模型的不断演进，Basis也不断叠加这些改进。每次新版本发布，都扩展了代理能处理的任务范围，提高推理质量，加快审核速度，并解锁更复杂的工作流程。

Basis联合创始人Mitchell Troyanovsky表示：“我们从一开始就与OpenAI合作。每次模型升级都拓宽了代理的能力。随着推理能力的提升，我们能够解锁更复杂、持续时间更长的工作流程，并赋予代理更大的自主权。”

将会计任务分配给合适的OpenAI模型

Basis将会计视为一套工作流程系统，每个流程都有其特定的上下文和复杂度。为此，团队构建了多代理架构，将最合适的OpenAI模型分配给对应任务。

每项任务由一个主管代理开始，最初基于OpenAI o3，现已迁移至GPT-5，负责协调整个流程——根据任务类型、复杂度、延迟需求和输入类型，将步骤分配给专门的子代理。GPT-5在推理、一致性和可解释性方面表现最佳，非常适合在高上下文工作流程中以最少监督引导代理。

子代理则由多种模型驱动，团队通过内部基准测试套件，根据关键能力和特性为每个模型打分。对于速度要求高的交互，如审核中澄清问题或快速反馈，Basis依赖GPT-4.1。

在更复杂的场景中，如解读异常交易模式、解决模糊分类或管理多步骤流程（如月末结账），Basis代理再次依赖GPT-5的深度推理能力。

这种协调机制使Basis能够随着模型能力的提升，持续改进任务覆盖率和准确性。

（附图：会计多代理系统流程图，展示多个AI代理执行发票处理、交易分类、对账和报告生成等任务，箭头表示模块间数据的顺序流动。）

利用OpenAI推理验证代理输出

在会计领域，自动化最有价值的是可审查的。Basis代理虽能独立工作，但通过中央层共享上下文，展示假设、数据来源及每个决策背后的逻辑。Basis最初依赖OpenAI o3-Pro来扩展工作流程中的推理，后来随着GPT-5发布，迁移至该模型，因其能推理结构化流程并解释结果生成过程。

以分录为例，主管代理会审查支持材料、检索数据、参考共享上下文和最佳实践，并协调子代理完成工作。会计师能看到分录及清晰的解释，包括使用了哪些数据、为何如此映射，以及系统对建议的置信度。

Troyanovsky指出：“我们所做的一切都依赖推理。这就是为什么OpenAI的模型，尤其是GPT-5，如此关键。通过大幅提升测试时计算能力，同时展现模型推理，我们能提供解释，让客户了解并掌控发生的事情。”

这种推理能力也支持主管代理以上下文和精准度分配任务。随着系统成熟，Basis从单纯的任务自动化迈向真正的工作流程委派。功能调用推动了这一进展，使代理能完成多步骤流程，如对账和分录，而不仅仅是提出建议，方式与会计师的思考和工作方法相符。

通过推理和可审查性推动模型基准测试

每次新模型发布，Basis团队都会对真实会计工作流程进行详细基准测试，不仅评估准确性，还考察模型解释推理的清晰度。这帮助团队决定不同任务应使用哪些模型，以及代理何时能安全承担新工作流程。GPT-5是Basis迄今为止最强的模型，因其在并行工具调用和高级推理方面的表现，非常适合需要深度和精确度的工作流程。

GPT-5在并行工具调用方面表现尤为突出——这是一项关键能力，使Basis代理能在单一工作流程中协调多个结构化操作。在Basis的工具调用基准测试中，GPT-5在启用代码解释器和网络搜索的情况下，实现了100%的成功率，并在所有推理基准中领先其他模型。

GPT-5能满足他们在规模上的性能需求，部分得益于与OpenAI团队的紧密合作。开发过程中，Basis分享了真实案例和边缘情况，并提供反馈，帮助塑造模型在生产环境中的表现。

Troyanovsky说：“OpenAI的模型在性能和部署速度上一直领先。这种推理能力与易用性的结合，使我们的架构成为可能。这种进步也让OpenAI成为极具价值的合作伙伴。我们不仅是被动响应模型改进，更是在推动它们的发展。”

与OpenAI共同构建信任，而不仅仅是完成任务

如今，Basis支持美国众多大型会计师事务所。使用Basis的事务所平均节省30%时间，并随着信任的增长，持续扩大代理的职责。更重要的是，他们重新获得了服务客户、探索新业务领域和深化咨询关系的能力。

Troyanovsky表示：“OpenAI在这一转变中发挥了关键作用。他们的模型不仅性能出色，还帮助塑造了我们的产品。随着模型的演进，代理的能力范围也在扩大，进而提升了会计师的工作能力。”

想了解更多关于ChatGPT在商业中的应用？

请联系我们的团队：https://openai.com/contact-sales/

Some startups use AI to solve a point-in-time problem. Others build systems that get better as AI improves. Basis⁠ is the latter.

Founded in 2023, Basis builds AI agents used by top accounting firms—designed to take on structured accounting work with the reliability and depth those tasks require. The team uses OpenAI o3, o3‑Pro, GPT‑4.1, and GPT‑5 to power AI agents that help accounting firms automate repetitive tasks like reconciliations, journal entries, and financial summaries while giving accountants full visibility into how decisions are made and control over the process. The result is up to 30% time savings and increased capacity for high-leverage work, like advising clients and taking on new business.

As OpenAI’s models evolve, Basis compounds those improvements. Each new release expands what agents can handle, boosting reasoning quality, speeding up reviews, and unlocking more sophisticated workflows. 

“We’ve worked with OpenAI from day one,” says Mitchell Troyanovsky, co-founder of Basis. “Each model improvement broadens what our agents can take on. As reasoning improves, we unlock more complex, longer-running workflows and grant our agents greater autonomy.”

Routing accounting tasks to the right OpenAI model

Basis treats accounting as a system of workflows, each with its own context and complexity. To support that, the team built a multi-agent architecture that assigns the best-fit OpenAI model to the right job.

Each task begins with a supervising agent, originally built on OpenAI o3 and now migrated to GPT‑5, which coordinates the full process—routing steps to specialized sub-agents based on task, complexity, latency needs, and input type. GPT‑5 is the strongest model Basis has evaluated to date in reasoning, consistency, and explainability, making it well-suited to guide agents across high-context workflows with minimal oversight.

Sub-agents are powered by a range of models, selected by an internal benchmark suite that scores each model on key capabilities and traits. For speed-critical interactions, like clarifying questions mid-review or surfacing quick feedback, Basis relies on GPT‑4.1.

In more complex scenarios, such as interpreting unusual transaction patterns, resolving ambiguous classifications, or managing multi-step processes like month-end close, Basis agents again rely on GPT‑5 for its deep reasoning capabilities.

This orchestration allows Basis to continuously improve task coverage and accuracy as model capabilities grow.

Validating agent output with OpenAI reasoning 

In accounting, automation is most useful if it's reviewable. Basis agents act independently but share context through a central layer, surfacing assumptions, data sources, and the logic behind each decision. Basis originally relied on OpenAI o3‑Pro to scale reasoning across workflows, and later migrated to GPT‑5 upon its release for its ability to reason through structured processes and explain how outcomes were reached.

Take a journal entry, for example. The supervising agent reviews supporting materials, retrieves data, references shared context and best practices, and coordinates sub-agents to prepare its work. The accountant sees the entry along with a clear explanation of what data was used, why it was mapped that way, and how confident the system is in its recommendation.

“Everything we do depends on reasoning,” notes Troyanovsky. “That’s why OpenAI’s models, especially GPT‑5, are so critical. By scaling test-time compute well beyond what earlier models could support, while still exposing the model’s reasoning, we can surface explanations that give customers visibility into and control over what is happening.”

This reasoning also powers the supervising agent’s ability to route tasks with context and precision. As the system matured, Basis moved beyond task automation into real workflow delegation. Function calling pushed that forward, enabling agents to complete multi-step processes like reconciliations and journal entries, not just propose them, in ways that mirror how accountants actually think and approach their work.

Driving model benchmarking with reasoning and reviewability

With each new model release, the Basis team runs detailed benchmarks on real-world accounting workflows, evaluating not just accuracy, but how clearly the model can explain its reasoning. This helps the team decide both which models to rely on for various tasks and when agents can safely take on new workflows. GPT‑5 is the strongest model in Basis’ stack to date and a strong fit for workflows that require depth and precision, thanks to its performance in parallel tool calling and advanced reasoning.

One area where GPT‑5’s performance stands out is parallel tool calling—a critical capability that enables Basis’ agents to coordinate multiple structured actions within a single workflow. On Basis’ tool-calling benchmark, which tested the model’s ability to use multiple tools in parallel with both code interpreter and web search enabled, GPT‑5 achieved a perfect 100% success rate while also leading all other models across reasoning benchmarks.

GPT‑5 delivers the performance they need at scale, thanks in part to close collaboration with the OpenAI team. Throughout development, Basis shared real-world examples and edge cases and contributed feedback that helped shape model behavior in production.

“OpenAI’s models have consistently led the way in both performance and speed of deployment,” says Troyanovsky. “That combination of reasoning power and accessibility is what makes our architecture possible. And that kind of progress is what makes OpenAI such a valuable partner. We’re not just reacting to model improvements, we’re helping drive them.”

Scaling trust, not just tasks, with OpenAI

Today, Basis supports a significant share of large accounting firms across the U.S. Firms using Basis report 30% time savings on average, and continue expanding agent responsibilities as trust grows. More importantly, they’re reclaiming capacity to serve clients, explore new practice areas, and deepen advisory relationships.

“OpenAI has been instrumental to that shift,” says Troyanovsky. “Their models don’t just perform, they’ve helped shape how and what we build. As the models evolve, so does the scope of what our agents can do, and therefore what accountants can do.”

Interested in learning more about ChatGPT for business?

Generated by RSStT. The copyright belongs to the original author.

Source