The Future of AI in Accounting: Why Firms Need to Own Their Prompts

Firms get more from AI when they own the instructions, keep accountants in the review loop, and use deterministic automation for work where the rule is already known. 

Most accounting software pitches start with a demo of a transaction being categorized. That is fine. It is also the least interesting part of the conversation. 

The question I hear from firm owners is harder: when the system makes a bad call, who explains it to the client? 

The accountant does. The partner does. The person whose name appears on the work does. 

That is why I have never been interested in building a black box that asks a firm to trust its judgment. Firms already have judgment. It sits in the heads of partners and senior managers, in month-end checklists, in the way one client treats contractors differently from the next. Good firms have spent years building that context. They should not have to hand it over to a vendor and hope it comes back intact. 

What NumbersGame is for 

NumbersGame connects Claude or ChatGPT to the QuickBooks Online and Xero books a firm manages through a secure Model Context Protocol connector. An accountant can ask for a reconciliation, a close review, a set of adjusting entries, or client reporting against the live ledger. The work is prepared against the books, also could be enriched with additional datasets like spreadsheet export pasted into a chat window. 

Three choices matter to me. 

Nothing writes before a person approves it. A categorization, journal entry, reconciliation, or report may be prepared in detail, but the accountant still decides what gets posted. The product should make the work faster without making the reviewer disappear. 

The model belongs to the firm. We do not run a NumbersGame model or ask firms to send their client data through another AI provider we selected. The firm uses its own Claude or ChatGPT account. We provide the connection and the instruction layer. That keeps the firm in charge of its AI account, its prompts, and its costs. 

Access is scoped and auditable. Connections use OAuth for QuickBooks Online and OAuth for Xero. A firm authorizes the specific companies an AI client can reach. There is a read-only tier, an audit log, and a quick way to revoke access. The platform also has hard boundaries: it cannot permanently delete ledger records, change the accounting platform’s settings, initiate payments, or touch payroll filings. 

That last list is not an afterthought. It is part of the product. 

The useful thing is not the button. It is the instruction behind it. 

Ready-made AI bookkeeping products often hide the most consequential part of the system: the instructions that tell the model what to do. If a firm disagrees with the result, it can correct the output, but it cannot inspect or change the policy that produced it. 

That does not work for accounting. 

A firm needs to be able to say: here is our materiality floor. Here is how we treat capitalization. Here are the vendors that require a 1099 review. Here is what this client expects to see in a board report. Those rules are not edge cases. They are the firm. 

NumbersGame includes a growing SOP library, currently around three dozen SOPs across close work, vendor mapping, reconciliation, financial reporting, and new-vendor handling. Each SOP explains the job in accounting language, gives the actual prompt, and names the inputs and the artifact the accountant should expect back. 

A language model is probabilistic. Calling the process a deterministic workflow would be misleading. The prompt is the reproducible artifact. A firm can read it, change it, version it, and teach a new hire how to use it. 

That is much more useful than asking someone to press a button and trust a confidence score. 

Two products, two kinds of work 

Some finance work should run the same way every time. A payout run, a revenue schedule, an invoice pipeline, and a dunning sequence all belong in deterministic process automation when the rule is already known. That is the work Loopfour is built for: repeatable finance processes with explicit rules, controls, and reliable execution. 

NumbersGame has a different job. It sits with the accounting firm at the ledger, where the work often begins with a question rather than a known rule. Is this new vendor payment a prepaid expense, a capital purchase, or an ordinary expense? Which intercompany balances should be eliminated for this client group? What needs to be explained in the board package? Those are judgment calls against a client’s facts and the firm’s own instructions. 

The products are complementary. Loopfour automates the repeatable finance process. NumbersGame gives an accounting firm controlled access to the books, its own AI instructions, and a reviewable way to handle work that needs professional judgment. One should not be stretched to imitate the other. 

The firms that get value write back 

The best product feedback we receive is not praise. It is a firm telling us that one of our SOPs is too thin. 

One CPA firm read our close guidance and said it was not how they close. They wrote their own version with their materiality floors, dormant-account rules, and capitalization thresholds. That was exactly the right response. The point is not for every firm to copy our operating model. The point is for a firm to put its own model somewhere its team can use and inspect. 

That led us to build the layer underneath the SOPs. A firm can add its own instructions, that take precedence over ours. A published change is versioned, acknowledged by a named person, recorded in a publish log, and reversible. 

Over time, a useful separation has emerged: 

• The SOP describes what a person does. 

• The firm instruction describes what the AI does. 

• The client file holds facts about that client, such as thresholds, accounting basis, and quirks. 

When all three live in one enormous prompt, a failure is hard to diagnose. Was the procedure wrong? Was the instruction unclear? Was a client fact stale? Keeping them separate makes the work easier to review and improve. 

Training should start with real asks 

Accountants do not learn this work from a feature grid. They learn it by seeing a useful request and trying it on a real client. 

We hold monthly AI training sessions for our customers and their teams. The sessions focus on the work in front of them: writing clearer asks, reviewing what came back, and building habits they can use in their own client work. 

That is why we publish our Prompt Library. The prompts come from work our own accountants do, with client details removed. They cover close work, reconciliations, deferred revenue, client onboarding, vendor review, and client communication. 

The practical habits are simple: 

1. Name the client. 
2. Name the period. 
3. Say whether this is read-only work or a proposed write. 
4. Describe the artifact you need at the end. 
5. Say explicitly when nothing should post. 

Those habits are more valuable than prompt-engineering tricks. They give the model a boundary and give the accountant something concrete to review. 

Scheduling follows the same principle. NumbersGame does not need to be its own scheduler. Firms can use Claude’s scheduled tasks to run a recurring weekly review or a month-end prompt through the same connector and instructions their team uses manually. The proposals are there the next morning for someone to review. 

The client with two ledgers 

Many firms have a client group that does not fit neatly into one platform. One entity is on QuickBooks Online. Another is on Xero. The usual answer is an export, a spreadsheet, and a monthly stitching exercise. 

NumbersGame can connect to both. The firm groups the entities, reviews the elimination rules, and asks for one consolidated view. The source books stay where they are. The accountant still decides how to treat intercompany balances and mismatched accounting bases. The system should show those differences, not quietly smooth them away. 

That is the broader reason I think the connector matters. An AI feature built into one ledger will naturally stay inside that ledger. Firms need an AI layer that can work across the systems their clients actually use. 

What changed for us 

Nicole Jordan-Dahdal of PrismEdge put the point well after using NumbersGame for an opex audit across three entities. The audit would normally have taken days. It took under two hours. 

The value was not that the system made the CFO’s decision for her. It saved the time spent assembling the numbers so she could spend more time reading them. 

AG Accounting now uses the product on more than 200 QuickBooks Online clients before new capabilities reach other firms. That is why the roadmap is full of unglamorous details: applying sales receipts to open invoices, converting purchase orders to bills, voiding rather than deleting, handling classes and departments at the point of creation, and bank-feed controls that hold work for review when the context is not there. 

That is the real accounting product. It is not a magic button. It is better access to the books, clearer instructions, and a reviewer who still has the final say. 

Give the accountant the prompt. Keep them in the chair. 

Anelya Grant is founder of NumbersGame, the AI connector that lets accounting firms run bookkeeping across QuickBooks Online and Xero through Claude or ChatGPT. She also runs AG Accounting and works on Loopfour, deterministic finance automation for CFOs and controllers.