If you review technical accounting memos, you know the conclusion is the easy part to check. The hard part is the path. Did the preparer land on the right ASC topic, read the paragraph that actually governs, check the related topics the Codification points to, and cite every step? Get the first step wrong and every citation after it is beside the point.
This post explains how our AI walks that path, why we stopped asking the AI model to work it out one search at a time, and what changed when we let the Codification's own structure do the routing.
Key takeaway: Our AI now follows the cross-references already written into the Codification, instead of guessing at them. On 18 real questions, each run as one search across the ASC text and three other sources, the ASC text itself now appears in every search (18 of 18, up from 2) and the paragraph a real answer cited is found more often (10 of 18, up from 6), and each answer needs half as many searches. Every link it follows points back to a paragraph a reviewer can open.
What is Fina?
Fina is Finrep's AI analyst for SEC reporting and technical accounting teams. You ask it questions in plain English, the way you would ask a colleague:
- "How should we account for this contract modification under ASC 606?"
- "How did our peers disclose segment changes in their last 10-K?"
- "Does this footnote draft cover what ASC 842 requires?"
Fina answers from primary sources: SEC filings, the US GAAP Codification, SEC rules and staff guidance, IFRS standards, and US tax law and regulations. Every statement in an answer carries a citation to the filing or paragraph it came from, so you can open the source and check it before relying on it. For more on why that matters, see our guide to AI hallucination in financial reporting, which puts it plainly: "If the output cannot be traced, it cannot be verified; if it cannot be verified, it should not appear in a filed document."
Why is technical accounting research hard for AI?
Because the answer is rarely in one paragraph. A technical accounting question usually involves more than one ASC topic. The governing paragraph has to be found by its address, not by matching words. The Codification also tells you which other topics to consider before you conclude. A general-purpose AI that searches by keywords finds passages that mention your words. It does not find the chain of guidance a reviewer expects to see.
What does a reviewer check on a typical question?
Take a question a reporting team might ask:
"We received $4 million to settle a patent lawsuit. Where does the receipt go in the cash flow statement, and can we recognize the gain this quarter? Cite the standard."
A careful preparer works through it in steps:
- Find the governing topics. Topic 230 (Statement of Cash Flows) for classification, Topic 450 (Contingencies) for the gain.
- Find the governing paragraphs. ASC 230-10-45-16(c) includes amounts received to settle lawsuits among operating cash receipts, unless they stem from investing or financing activities. ASC 450-30-25-1 bars recognizing a gain contingency before it is realized. Our ASC 450-30 guide covers what "realized or realizable" means in practice.
- Check the facts that could change the answer. Is the settlement final, or is any part still contingent? Does the agreement also grant a license to use the patent from here on? If it does, part of the $4 million may need its own analysis under another topic, such as Topic 606 if licensing is part of the company's ordinary activities.
- Follow the cross-references. A paragraph that says "see paragraph 450-20-55-36" means the analysis isn't finished until you've read that paragraph.
- Cite by paragraph. Every conclusion needs an address an auditor can open.
In the usual case the answer is an operating inflow, with the gain recognized once the settlement is final and the cash is in. That isn't the hard part. The hard part is knowing you've checked everything that could change it. A keyword search for "lawsuit settlement cash flow" doesn't show you that path.
Why did our first design fall short?
It made the AI model do every lookup itself, which was slow and often left the standard out. On every turn, the model read a routing map our accounting team maintains, which lists the topics that relate to each topic. The map ran to about fifteen thousand tokens. The model then searched one source at a time and read the standard one paragraph per search.
That worked, but users paid for it in two ways:
- Slow answers. A typical research chat took six searches and twelve model calls, and each call re-read the whole conversation. Most of those calls were lookups that had only one right answer.
- The standard dropped out. When a single search asked for the ASC text alongside several other sources, one source crowded out the rest. The ASC text itself appeared in only 2 of 18 test searches. No reviewer signs off on a memo that cites commentary on a standard but not the standard.
How does the Codification link its own topics?
Every topic has two sections that say where else to look: Section 15 (Scope and Scope Exceptions) and Section 60 (Relationships). Every ASC paragraph also has an address: Topic, Subtopic, Section, Paragraph. For example, 450-20-25-2 is Topic 450 (Contingencies), Subtopic 20 (Loss Contingencies), Section 25 (Recognition), paragraph 2.
Section 15 says what a topic covers and which topic to use instead. Section 60 lists other topics that bear on the same transactions. For instance, 450-10-60-3 points to Topic 606 on variable consideration, and 450-10-60-6 points to Topic 740 on uncertain tax positions.
We read the Scope and Relationships sections of the 97 US GAAP topics in our library and recorded every link along with the paragraph it came from, 774 links in all. They match about half of the links in our accounting team's routing map. The team's other links are judgement calls that no paragraph states, such as when a contingency question turns into an exit-cost question under Topic 420. We keep those as a reviewed layer on top. The links taken from the text update automatically when the standard changes.
The same idea works beyond US GAAP. An Accounting Standards Update names its topic in its title ("Update 2016-13 (Topic 326)"). IFRS standards, Treasury regulations and PCAOB standards carry their numbers. And the text itself is full of pointers like "see paragraph 450-20-55-36", which we store as links too.
How does our AI research a question now?
The model makes one search call naming the topics it thinks govern, and code does the lookups from there:
- Route. Each topic goes to the sources that cover it, such as the ASC chapter and the related ASUs. The related topics come back with the paragraph that links them, for example "450; also 340 (450-10-60-2) and 740 (450-10-60-6)".
- Search twice. Once inside those sources and once across the whole library, so a wrong guess about the topic doesn't hide the right answer.
- Balance the results. The ASC text and every other source the model asked for are all represented, instead of one source filling the list.
- Follow cross-references in advance. A passage that says "see paragraph 450-20-55-36" arrives with that paragraph attached.
- Fetch named paragraphs. When the model already knows which paragraphs it needs, it gets them all in one go.
The library holds about 5,500 sources, including the US GAAP Codification, every ASU since 2009, IFRS standards and their Basis for Conclusions, SEC rules and staff guidance, and the Internal Revenue Code and Treasury regulations. IFRS and tax questions route the same way ASC topics do. A question on IAS 37 goes to the standard and its Basis for Conclusions. A question on Treas. Reg. §1.409A goes to the regulation and the IRS bulletins that cite it.
The whole search takes under half a second.
What changed, in numbers?
| Typical research chat | Before | Now |
|---|---|---|
| Searches | 6 | 3 |
| Model calls | 12 | 8 |
| Cost per chat | 1× | about half |
| Time per search | under half a second | under half a second |
Measured on our test set (18 real questions from production chats, each run as one search across the ASC text and three other sources):
- ASC text in the results: 2 of 18 searches before, 18 of 18 now.
- The paragraph a real answer cited, found in the top ten results: 6 of 18 before, 10 of 18 now.
- Different sources in the top ten: 1.6 on average before, 9.7 now.
How did we test it, and what doesn't it prove?
We took questions from real chats and used the paragraph each answer went on to cite as the target. Then we checked every target by hand against the standard. The test is strict: a passage one paragraph away from the target counts as a miss. Before any change shipped, we also ran the full AI on a separate set of graded questions, so a better search could never hide a worse answer.
The limits, plainly:
- Eighteen questions is a small set. We are adding more as we check them.
- Eight of the eighteen still miss.
- Questions about a single standard did not improve. Search already found the right chapter there. The remaining misses are about ranking within one long chapter, which is a separate problem.
- Search uses the current edition by default. Whether that edition applies to your reporting period is still your call, and a search can ask for an earlier one.
What does this mean for your review?
- You can check the path, not just the conclusion. Every link the AI followed points back to a paragraph.
- The standard comes first. The ASC text itself is part of every search, not just commentary on it.
- The AI does the lookups. The judgement stays with you.
Key Takeaways
- Technical accounting answers depend on finding the governing ASC paragraph and the related topics, not on matching keywords.
- The Codification already links its own topics through Section 15 (Scope and Scope Exceptions) and Section 60 (Relationships).
- Following those links in code halved the searches per research chat and put the ASC text in every result on our test set.
- Every link carries the paragraph it came from, so a reviewer can trace and verify each step.
FAQ
What is Fina? Fina is Finrep's AI analyst for SEC reporting and technical accounting teams. You ask questions about filings, standards or your own drafts in plain English, and every answer cites the paragraph or filing behind it.
How does AI find the right ASC paragraph? Fina routes each question to the ASC topics that govern it, then follows the Codification's own Scope and Relationships sections to related topics. It searches the ASC text by paragraph address as well as by meaning, so the governing paragraph is found even when its wording doesn't match the question.
Can I verify the AI's answer? Yes. Every statement cites a specific paragraph or filing, and every related topic the AI checked comes with the paragraph that links it.
What if the AI starts from the wrong ASC topic? Every topic-scoped search runs alongside a search of the whole library, so the right guidance can still surface.
Does it cover IFRS and tax, or only US GAAP? IFRS standards, SEC rules, the Internal Revenue Code and Treasury regulations all route the same way ASC topics do.
Does it replace our accountants' judgement about which topics relate? No. Links taken from the Codification's text cover about half of them. The rest come from our accounting team and are reviewed by hand.
Related Reading
- ASC 450-30 Gain Contingencies: 2026 Guide for Finance Teams
- AI Hallucination in Financial Reporting: A 2026 Practitioner Walkthrough
- AI Tools for SEC Filing Research Compared: 2026 Guide
We're hiring engineers who want to build AI that accountants can check. See open roles at Finrep.
