Xero and Anthropic have entered a multi-year partnership to integrate the Claude large language model (LLM) into the global small business platform, aiming to transform accounting from a manual process into an automated, agentic workflow.
In an exclusive intevriew with the Fintech Times, we speak to Diya Jolly, Cheif Product & Technology Officer at Xero.
This collab is a reflection of the industry shift to embedding financial tools directly into a third-party AI interface, allowing users to manage payroll, payments, and tax obligations through natural language.
The move follows a period where Xero utilised multiple models, including those from OpenAI and Gemini, to achieve a 97 per cent accuracy rate in bank reconciliations. By partnering with leading AI developers like Anthropic, Xero intends to tackle the remaining three per cent of complex financial tasks, such as split invoice payments and unstructured data sources that traditional ledgers struggle to process.
Moving Beyond Structured Data
Historically, accounting software has relied on structured data to maintain accurate ledgers. Diya Jolly, Chief Product and Technology Officer at Xero, explained that the integration with Anthropic allows the platform to reason across unstructured information. This capability enables small businesses to upload diverse data sources, such as spreadsheets or information from legacy payroll systems, and map them directly to their financial records.
“Accounting has sat within a web app forever,” Jolly noted during the discussion. “With natural language, you can actually make accounting available wherever the user is and meet the user”. This shift towards “headless apps” means that financial data is no longer siloed within a single interface, but can be interrogated via Claude.ai or other communication channels.
The Rise of Agentic Workflows
The partnership introduces a reasoning layer to Xero’s financial infrastructure through its AI superagent, JAX (Just Ask Xero). Unlike standard automation, which follows pre-defined rules, agentic AI can orchestrate complex tasks from start to finish. For a small business owner, this might involve asking Claude to identify a supplier, check real-time cash flow to confirm affordability, and then generate the invoice or payment without leaving the AI chat interface.
This evolution prompts questions regarding the future of the bookkeeping profession. While some fear that AI will replace human roles, Jolly argued that the technology is designed to eliminate “grunt work” rather than human oversight.
“Bookkeepers will become more advice givers than the doer of grunt work,” Jolly said. She added that while AI can automate repeated actions, it lacks the subjective understanding of a business’s seasonality or specific context—such as whether a lunch expense should be
classified as sales or entertainment. As the industry shifts, the role of the bookkeeper is expected to move from a volume-based profession to one focused on high-value judgment.
Deterministic Records vs. Probabilistic AI
A fundamental challenge in applying LLMs to finance is the conflict between the probabilistic nature of AI and the deterministic requirements of accounting. Xero functions as a system of record where $1 + 1$ must always equal $2$. In contrast, LLMs predict the next likely token in a sequence, which can lead to variations in output.
To mitigate this, Xero acts as the deterministic anchor. The platform serves as a financial operating system that pipes in bank data and manages compliance rules, while Claude provides the analytical layer on top. Jolly emphasised that for critical tasks like tax submissions, a human expert remains essential to sign off and take responsibility for the filings.
“I don’t think you’re ever going to not review something that has to be deterministic and where the repercussions for it being wrong are high,” Jolly commented, noting that users should treat AI agents like junior employees that require supervision.
Security and Data Sovereignty
A primary concern for financial institutions (FIs) and small businesses when adopting AI is data privacy. The partnership is built on a framework where financial data remains within the Xero ecosystem. When a user queries their financials through Claude, the LLM calls a Xero Model Context Protocol (MCP) server.
This setup ensures that proprietary business data is never used to train Anthropic’s models. Instead, Xero serves an applet within Claude, maintaining a private space for the data exchange. This “trusted intelligence” model is designed to provide the benefits of advanced reasoning without compromising the security of the underlying financial records.
Scaling for the Future
As small businesses grow from lean operations to larger organisations, their need for integration increases. Xero intends to remain an infrastructure layer that connects to various point-of-sale (POS) systems, CRM software, and payroll providers. While Claude provides the reasoning capabilities, Xero provides the API surface and connections to thousands of ecosystem apps.
Looking ahead, Xero’s roadmap includes expanding its presence across more interfaces, including potential integrations with popular accounting tools such as spreadsheets and other office applications. The goal is to move toward a “default automated” state for books closing and tax preparation, utilising the specific strengths of different AI models for different tasks.
The collaboration between Xero and Anthropic represents a significant step in the fintech sector’s adoption of agentic AI. By combining the vast datasets of a global accounting platform with the reasoning power of a leading LLM, the partnership seeks to provide small businesses with the level of financial intelligence previously reserved for companies with dedicated CFOs or analysts.
