AccountantGPT

Problem

Small accountancy firms spend a significant fraction of their time on repeatable, rule-bound tasks: reconciliation, VAT prep, generating management account summaries. These tasks are automatable, but existing accounting software doesn't offer LLM-native workflows.

Approach

Build a LangGraph agent with access to the Xero API as its primary tool set. The agent can plan and execute multi-step workflows (e.g. "pull all transactions from last month, categorise any uncategorised ones, flag anything over £1k, generate a summary") using MCP for tool extensibility.

Key technical decisions

LangGraph chosen over a single-prompt chain because reconciliation workflows are stateful — the agent needs to branch, retry, and aggregate across multiple API calls. MCP used to make the tool set extensible without rebuilding the agent.

Outcome

Early architecture phase. Targeting a working demo against the Xero sandbox. Success = a full month-end workflow completed with zero manual steps.