AI tools for bookkeeper workflows and client reporting.

AI tools for bookkeeping split into four categories that are easy to confuse and behave completely differently. AI inside the ledger is what QuickBooks itself ships, working in one company file while you are in the product. AI bookkeeping products and services such as Booke AI, Docyt, Zeni or Truewind do the coding for you and hand back results. AI review and close layers such as Keeper, Xenett or Numeric check work that already exists. AI access layers, including Intuit’s own QuickBooks Online MCP server and Numbers Game, connect an AI client you already have to the ledger and let it do the work under your supervision.

If you are a firm that wants the AI to run your procedures across many clients with your people approving every write, you want the fourth category, and Numbers Game is the multi-client option in it. If you want somebody else to do the bookkeeping and hand you the result, you want the second, and you should buy one of those instead of us. The rest of this article is the eight questions that tell those apart, in the order they matter.

• 1. Whose model is it, and does client data reach a model vendor? This is the first question because it is the one with a compliance conversation attached. Numbers Game does not run inference. There is no Numbers Game model and no AI provider sitting between a firm and its clients’ data: the firm brings its own Claude or ChatGPT subscription and that account does the thinking. It also means there is no second usage bill, and no vendor quietly accumulating a corpus of your clients’ financials.
• 2. Can you read the instructions it follows, and can you change them? Most products in this space treat their prompt as the product, which means the instructions your work runs under are a trade secret you are not allowed to audit. Ask to see them. The answer tells you whether you are buying a tool or a subscription to somebody else’s judgment.
• 3. Does anything write without a person? Ask what a human sees before a posting lands, and what happens by default rather than what is possible with settings. Defaults are the real policy.
• 4. Is there an audit log, and can you read it yourself? Not a support ticket that produces a log. A screen your own staff can open. When a dispute lands, the questions are who approved this, what instructions were in force that day, and can you produce the version.
• 5. Does it scale across clients, or per file? A tool that is wonderful in one company and identical work in the next sixty is not a firm tool. Ask specifically how a single request reaches many clients, and where the per-client boundary is enforced.
• 6. Where does confidence come from? If the product shows a score, ask whether the model produced it. A model asked to grade its own certainty returns something that reads calibrated and is not. Confidence worth trusting is computed outside the model, from rule agreement, coding history, and source-data quality.
• 7. What can it structurally not do? A vendor who can list their product’s incapabilities has thought about blast radius. Ours cannot permanently delete QuickBooks records, cannot change Intuit account settings, cannot move money or initiate ACH, and does not touch payroll filings. Those are tool-level exclusions rather than policy promises.
• 8. What does the client actually receive? Reporting is where the firm’s brand lives. Ask whether the output carries your name, your language and your format, or the vendor’s.

Every accountant evaluating AI is asking a question vendors mostly avoid. Not can it categorize this transaction, which is table stakes. The real question is: when it gets it wrong, whose name is on the return. It is the accountant’s name. It is always the accountant’s name, and that fact should shape the product.

So Numbers Game ships a library of standard operating procedures, currently three dozen or so across seven categories: clearing the uncategorized queue, vendor mapping, bank and credit card reconciliation, pre-close review, locking the period, branded financials, board reporting, handling a new vendor, class and department tagging. Each one has the same shape. What the work is, in plain accounting language, including the preconditions and the tie-outs you check before approving anything. The prompt, as actual text you can read and copy. And the inputs and outputs: what the AI needs to see, and what artifact you should end up holding.

We call these prompt recipes rather than step paths, deliberately. A language model is probabilistic. Selling an accountant a deterministic workflow is selling something that does not exist, and the first time it deviates they will never trust it again. The reproducible artifact is the prompt, not the path.

The strongest signal we get is a firm telling us our SOP is too thin. One CPA firm on the platform read our month-end close procedure, decided it did not match how they close, and wrote their own: twenty pages, seven steps, their own materiality floors, their own rule for dormant accounts, their own capitalization threshold per client. That is not a support ticket. That is the product working, so we shipped the layer underneath it. A firm can now write its own section of the AI’s instructions, with its own trigger phrase and its own rules, spliced into the skill file its team downloads. Their instructions take precedence over ours, and because that precedence is real it is governed like a change to a control: every publish is versioned, requires an explicit acknowledgement from a named person, is recorded in a publish log, and can be rolled back to any earlier version.