Finance AI That Ties to the Books
Listen to this story
I spent years at Workiva on one problem: making a reported number trace back to the facts behind it. A number in a 10-K had to say what it was, which period it covered, and how it rolled into the lines above it, or it wasn't done.
Finance AI has brought that problem back. An agent will happily tell you that software spend rose 18% last quarter. The question every controller, auditor and CFO asks next is the same one we asked of every filing: where did that come from? If the answer is "a spreadsheet someone exported on Tuesday," the number is a guess with good grammar.
This is the case for finance AI that ties to the books: every figure an agent gives you opens to the entries behind it, and the agent runs where your books are allowed to live.
What does "ties to the books" mean for finance AI?
It means an agent's answer can be checked the way you'd check a workpaper. Each figure points to a period, an account, and the posted entries that make it up, and a person can follow that path without asking the agent to explain itself.
Accountants already use the phrase. A schedule "ties" when its total agrees to the general ledger. The bar for an AI agent should be no lower than the bar for a staff accountant's schedule: show the number, show where it came from, and make the two agree.
Why do AI agents need the ledger's structure, not an export?
Because the structure is where the meaning lives. A trial balance export gives an agent account names and numbers; it doesn't give it the chart's hierarchy, the fiscal calendar, which months are closed, how accounts map to reporting lines, or which entries are drafts that haven't posted.
Without those, an agent guesses. It sums an account that belongs to two reporting lines, or compares an open month to a closed one, or treats a draft accrual as booked. The arithmetic looks right and the answer is wrong, and nothing in the export tells you so.
In RoboLedger, the general ledger built on RoboSystems, the arithmetic is calculated, not generated. Entries come from events by rules, statements are built from the posted ledger through your mapping, and closing a month saves its statements so they read the same next week. The agent reads those calculated figures and does its reasoning on top of them. It never has to invent the math.
How can an AI agent generate a balance sheet or general ledger summary by period?
It should ask the ledger for one rather than build one. A ledger that exposes statements by period, from posted entries through a defined mapping, lets an agent request "the balance sheet for October" and get the same answer a person would get from the same books.
RoboSystems exposes the ledger to AI over MCP, so Claude, ChatGPT or any MCP client can request a live statement for a period, an account's register with the balance after each line, or a closed month exactly as it was saved. The agent's job becomes explaining and comparing, which is what it's good at, rather than summing, which is where it slips.
What does it look like for FP&A AI agents to cite their sources?
Every number in the answer carries a path to its source, and the path is something a person can open. For your own books, that path is the account, the entries, and the event behind each entry: the invoice, the bill, the payment. For a public company, it's the fact in the filing, with its concept, its period and a link to the document on EDGAR.
That second half matters more than it sounds. Most FP&A questions end in a comparison: our gross margin against peers, our days sales outstanding against the industry. RoboSystems holds a knowledge graph of 8,000+ SEC filers' XBRL filings in the same schema as your ledger, so a peer's figure arrives with the same kind of citation as yours. One ontology underneath, labelled in XBRL's standard terms wherever XBRL has one.
What software provides traceable root-cause analysis from consolidated numbers?
The kind that keeps the group's structure, so a consolidated figure can be walked down to the company, the account and the entry that moved it. A spreadsheet roll-up can't do that, because the roll-up is the end of the trail.
In a knowledge graph, the path is the data. A consolidated line connects to each company's contribution, each contribution to its accounts, each account to its entries, and each entry to the event that caused it. Ask why operating expenses jumped and an agent can follow the edges down to the three vendor bills that did it, instead of offering three plausible stories.
What platform can track and log every action an AI agent takes in your financial systems?
One where the agent can't change the books without leaving a record, and can't post at all without a person. Two controls do most of the work: a change log on the books themselves, and a rule about who posts.
On RoboSystems, every run of its own AI on a graph is recorded with the graph: the question, what the model read through the graph's tools, what it answered, and what it cost. A change made by the AI points back to the run that made it, and the record is kept for thirteen months. Agents work through API keys you issue and can revoke. And on the QuickBooks side, nothing writes back until a person posts the entry.
Where can AI agents retrieve verified financial data?
From a source that carries its own context, so the agent doesn't have to guess what a number means. For public companies, that's the XBRL filings themselves: each fact already says its concept, its period, its units and how it rolls up.
There are two ways to use them with RoboSystems. The hosted SEC repository serves the 8,000+ filers as a graph your agent can query over MCP. And xbrlkit, free and local, loads any single filing on your own machine and answers from the filing itself, no account needed. We also measured how much the form matters. In the Filing Ladder, a public benchmark on SEC filings, every representation that carried the filing's text read the same on the same questions, and the cheapest was a plain search tool over that text. Structure doesn't make the model smarter. What it buys is the trace: a fact that says what it is, and a path from it to everything it connects to.
Why should finance AI run where your books are allowed to live?
Because some books can't leave. A CPA firm holding client ledgers, a company under a confidentiality agreement, or a group whose board wants its numbers on its own infrastructure can't paste the general ledger into someone else's product and hope.
RoboSystems is open source, Apache 2.0 and MIT. You can run it on the RoboSystems managed platform, in a Dedicated Deployment RoboSystems runs, or in your own cloud, and you keep the code and the data whatever happens to any vendor, including us. When you host it yourself, you can point its AI at the model you choose, including an open-weight model on your own hardware. RoboSystems' own AI doesn't train on your data.
What proof should a finance AI vendor provide?
Proof you can open, not adjectives. Five things are worth asking for before you trust any finance AI with your numbers:
- The trace. Pick a number in a demo answer and ask to see the entries behind it, live.
- Where the data lives. Which cloud, which region, who can reach it, and whether you can run it yourself.
- The licence. If the vendor disappears, what do you keep: the code, the data, or neither?
- The log. What the AI read and changed, recorded where you can read it.
- Their own books. Does the vendor run its own company on it?
Our answers: the trace is the product; the data lives where you choose; the licence is Apache 2.0 and MIT; the log is kept with the graph; and our own companies' books are kept on it, closed through September.
The bar
Finance has always had a bar for numbers: they tie, they trace, and someone can check them. AI doesn't get to lower it. The agents are good enough now that the limit isn't the model; it's whether the model is reading your books or a copy of them.
If you own finance data or AI for a team, a firm or your clients, I'd like to show you the trace on a demo company. Book 20 minutes, or read the code first. RoboSystems is the open-source knowledge graph for finance.
Joey French
RoboSystems, the open-source knowledge graph for finance, is built by an inactive CPA and ex-Workiva knowledge engineer. Our own companies' books are kept on it.
Finance AI that ties to the books
Talk it through in 20 minutes, or read the code first.
More from RoboSystems Blog
Same Answers, Cheaper: Four Versions of a Benchmark on One SEC Filing
Nov 10, 2026
On one SEC filing, every representation that carries the text reads the same, and the facts without it score 15 to 53%. What the tools buy is cost: the hosted knowledge graph went from 9.1 to 5.4 cents per correct lookup across three versions while the plain text read whole went from 13.9 to 12.4, and accuracy did not move. The three findings that cut against us come first.
What Is Financial Intelligence? For AI, It Means Tying to the Books
Sep 21, 2026
Financial intelligence usually means a person's skill with financial statements, or a government unit that tracks money laundering. For AI working on a company's books, what matters is narrower and easier to test: every number opens to the entries behind it. Shown on a 3M filing anyone can check.
Semantic Sovereignty: The Missing Layer of Sovereign AI
Jul 30, 2026
Sovereign AI tells you where your intelligence runs. Semantic sovereignty asks a harder question: who owns what your numbers mean? A black box in your own data center is still a black box.