You just asked Claude "what did we spend on travel last quarter?" and it answered — not by guessing, not by asking you to export a CSV first, but by reaching directly into your expense reports and reading the real numbers. No copy-paste. No spreadsheet round-trip. Just a question and an answer, sourced live from your actual books.
That's not a hypothetical. On June 8, 2026, Expensify launched a Model Context Protocol (MCP) server that lets AI assistants — ChatGPT, Claude, Cursor, and any other MCP-compatible client — connect directly to a business's expense data and answer questions like "which expense reports need my approval?" or "what did I spend on software subscriptions in June?" in plain English, pulled live from the account. It's a small feature announcement with a much bigger implication: the walls between "your books" and "your AI assistant" are coming down, one integration at a time, and small business owners are going to have to decide how much access they're comfortable granting.
What MCP Actually Is (in Plain English)
Model Context Protocol is an open standard, originally published by Anthropic in late 2024, that defines a common way for AI models to connect to external tools and data sources. Before MCP, every company that wanted its AI assistant to read data from, say, QuickBooks or Expensify had to build a custom, one-off integration. MCP replaces that with a standard "connector" — build one MCP server, and any MCP-compatible AI client (Claude, ChatGPT, Cursor, and a fast-growing list of others) can plug into it the same way.
Think of it as roughly analogous to what USB did for peripherals, or what a universal remote does for your TV, streaming box, and soundbar. Instead of a different cable or app for every device, you get one interface that everything speaks. For accounting and expense software, that means an AI assistant can, in principle, read your invoices, categorize your transactions, flag pending approvals, or summarize spend — using the same protocol whether it's talking to Expensify, QuickBooks Online, Xero, or Zoho Books.
Expensify's implementation connects through OAuth 2.1, the same authorization framework banks and financial apps use to grant limited, revocable access without ever handing over your actual password. Setup is available to anyone with a validated Expensify account, and once it's connected, a member can ask their AI client natural-language questions and get answers sourced from their real expense data — no manual export required.
Why This Matters Even If You Don't Use Expensify
The specific product here is almost beside the point. What matters is the trend it represents: 2026 is the year "AI-readable financial data" stopped being a nice-to-have and started becoming table stakes. QuickBooks, Xero, and Zoho Books have all shipped or are shipping their own MCP servers. Gartner projects that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from under 5% a year earlier. The accounting software you use — or evaluate switching to next year — is very likely to have an "ask your AI about your books" feature within the next 12 months, whether or not you ever click the button.
For a small business owner, that shift changes a quiet but important assumption: your financial software used to be a closed box that only you (and your bookkeeper) could query. Increasingly, it's a data source that a third-party AI model can read, summarize, and reason over on request. That's genuinely useful — instant answers to "are we over budget on marketing this month?" without opening a report — but it's also a new category of access to a system that has your bank details, vendor relationships, and spending patterns.
What to Actually Check Before You Connect an AI Assistant to Your Books
If you're a small business owner using Expensify, QuickBooks, or a similar tool that now offers an AI/MCP connection, a few questions are worth answering before you flip it on:
Is the connection read-only, or can the AI take actions? There's a meaningful difference between an assistant that can answer "what did I spend on travel" and one that can approve expense reports or move money. Read-only access to summarize and search is low-risk. Write access — approving, categorizing, or submitting on your behalf — deserves a much closer look at what guardrails exist before you grant it.
What's actually revocable, and how fast? OAuth-based connections (like Expensify's) can typically be disconnected from the software's own settings panel at any time, which is the right architecture. Confirm you know where that toggle lives before you need it in a hurry.
Does your AI client's own data policy matter here? The data isn't just flowing to Expensify — it's flowing to whichever AI model you're chatting with. Check that model provider's policy on whether prompts and retrieved data are used for further training, and for how long results are retained.
Who else on your team can connect it? MCP access is typically tied to an individual user's account and permissions within the software, but it's worth confirming that a junior employee with expense-approval rights can't accidentally grant an AI assistant the same reach.
This caution isn't paranoia for its own sake. Security researchers tracking MCP adoption through 2026 have flagged prompt injection — where a malicious instruction hidden inside a document, email, or web page tricks an AI agent into taking an unintended action — as a structural risk of any system where an agent reads external content and can also act on connected tools. One industry survey this year found that 88% of organizations had experienced a confirmed or suspected AI agent security incident in the past twelve months. None of that means "don't use it." It means treat AI-to-books connections the way you'd treat any new piece of software with access to your financial accounts: read the permission scope, start with read-only where it's offered, and don't grant more access than the task actually requires.
The Upside Is Real, Not Just Hype
It's worth saying plainly: this isn't just a security story. For a small business owner who doesn't have a full-time bookkeeper on call, being able to ask "which vendors did we pay twice this month?" or "what's our burn rate trending toward" and get an instant, accurate answer sourced from real transaction data is a genuine time-saver. It closes the gap between "I have the data somewhere in my accounting software" and "I actually know the answer right now," which is exactly the gap that causes small businesses to make decisions on stale or half-remembered numbers.
The businesses that benefit most from this shift will be the ones whose underlying financial records were already clean, well-categorized, and easy to query — because an AI assistant can only summarize what's actually in your books accurately. If your chart of accounts is a mess or your categorization is inconsistent, connecting an AI won't fix that; it'll just answer questions faster with the same underlying noise.
Keep Your Books Clean Enough to Trust an AI With Them
Whether or not you connect an AI assistant to your expense software this year, the underlying lesson holds: financial records that are structured, transparent, and easy to query — by a human or a machine — are simply more useful. Beancount.io provides plain-text accounting that keeps your data in an open, auditable format you fully control, with no vendor lock-in and no black box between you and your numbers. Get started for free and see why developers and finance-savvy business owners are switching to plain-text accounting built for the AI era.