BalanceProof vs Financial Datasets AI: Sampling vs a Test on Each Filing

One of the few rivals that publishes its verification method in detail, so the comparison is unusually concrete. Their own page describes sampling 1,000 companies across 75 sectors and checking 20,000 data points per audit cycle against SEC EDGAR. That is a serious process, and it is a different one.

At a glance

Prices and features on the right-hand column change, and this page does not — check their site before you decide anything on it. The rows about method are the ones that stay true.

BalanceProof compared with Financial Datasets AI
  BalanceProof Financial Datasets AI
Positioning A verification layer over SEC balance sheets “The first platform designed for AI agents” — their words
Verification method Each filing tested against A = L + E when it reports the lines needed Sampling: 1,000 companies, 20,000 data points per audit cycle, checked against EDGAR — their figures
What the check covers Each balance sheet served that reports the lines needed, every load; the rest are marked not testable A sample, per cycle
Result published per company Yes — pass, or the exception and its reason Check their site
Stated scope 6,257 SEC filers, 1.3M facts 27,530 stocks, 75 sectors, 30+ years — their figures
Statements covered Balance sheet, income statement and cash flow (US SEC filers) Balance sheets, cash flow, earnings, insider trades, segments
Webhooks On Business (CA$99/month) On their top self-serve tier (Professional), per their pricing page
Free tier 1,000 calls/month, no card Check their site

Why transparency matters more than a number

Every provider in this category reads the same filings. The difference is what happens when a filing is hard to read. The usual answer is that you get a number anyway, with nothing attached to say how confident it is — and a figure that is quietly a segment instead of a company looks exactly like one that is right.

BalanceProof checks each balance sheet against Assets = Liabilities + Equity when it reports the lines needed, and marks it not testable instead of passing it when it does not. A filing that reconciles is published with its figures. A filing that does not is published with the reason: a noncontrolling interest reported as a separate line, mezzanine equity outside permanent equity, rounding inside one percent, a component we could not read, or a filing whose own totals disagree with each other. The exceptions are counted in public and named individually on how we verify.

That is the whole claim. Not that nothing is ever wrong — that when something is, you are told which number and why, instead of finding out from your own reconciliation three weeks later.

The difference is which tag gets picked

Every provider in this category reads the same source. SEC EDGAR publishes XBRL for every filer, free, and nobody has better raw material than anybody else. What separates one API from the next is not access. It is the selection step, and that step is almost never documented.

Here is the problem it has to solve. Open JPMorgan's 10-Q and search for Assets and you get twenty-three facts. Not twenty-three values — twenty-three tagged instances, one for the consolidated bank and one for each segment and subsidiary that has to be broken out separately. Every one is valid XBRL. Exactly one is the number on the face of the balance sheet, and the only thing marking it is an absence: it is the fact with no dimensions attached.

An extractor that takes the first match, or the largest, or the most recently filed, will be right most of the time and wrong in a way that leaves no trace. No exception, no null, no warning. Just a number that is a segment instead of a company. SEC's companyfacts API already drops the dimensioned facts, so there the failures move elsewhere: in our 100-company test (September 2026) a naive script found no total liabilities tag for 17 companies, a stale equity tag for 8, and headline totals that did not satisfy A = L + E in about a third of the companies it could check.

BalanceProof resolves it with arithmetic rather than a heuristic: pull every candidate for assets, liabilities and equity, and keep the combination that satisfies Assets = Liabilities + Equity. The consolidated figures balance against each other. A segment's assets do not balance against the whole company's liabilities. The identity is a test, not a guideline, and it is the reason a figure here is checked rather than guessed at.

When nothing balances, the answer is that nothing balances. The filing is served as-reported with a warning on it rather than adjusted until the columns agree, because a filing that does not add up is a fact about the company, and you should get it as one.

Why this comparison is worth reading carefully

Most pages like this compare a published method against silence, which is an easy argument to win and not a very useful one. This one does not. Financial Datasets AI publishes what it does, in numbers, and those numbers describe real work.

So the difference is worth stating precisely rather than spun. A sampled audit against the source document is deeper than an identity test: a person comparing a figure to the filing catches a misread label, a wrong period, a footnote that changes the meaning — none of which arithmetic notices. An identity test is wider: it runs on every filing, every load, at no cost, and it catches the one failure that is invisible to inspection at scale, which is a plausible figure selected from the wrong tag.

If you are choosing between them on this axis alone, the question is whether you need to know the general quality of a dataset or the specific status of the filing in front of you. Those are different questions and they have different right answers.

What we are not going to pretend

We are not going to put Financial Datasets AI's prices in a table on our own website. They change, this page would not, and you would be reading a number we had no way to verify at the moment you read it. Go and look at their pricing page. It is the only copy that is current.

We are also not going to tell you their data is bad. We have not audited it and we are not in a position to. What we can tell you is what this service does and how to check it, which is the part we are actually responsible for.

The check that settles it costs you nothing either way: take a company where you already know the answer, call both, and compare each against the filing on EDGAR. Not against each other — against the filing. That is the only comparison that means anything, and it is why the free tier here needs no card.

When to choose each

Choose BalanceProof

If the number has to be right.

  • A sample tells you a rate; a test tells you about YOUR company. Auditing 1,000 companies establishes that the data is good in general. It cannot tell you whether the filing you are about to publish a number from reconciles, because your company may not be in the sample. Each filing here is tested when it reports the lines needed, and the answer travels with the response as a check object saying whose gap a failure is; one that cannot be tested says so.
  • The identity needs no sampling because it is free. Checking A = L + E costs one subtraction per filing, so there is no reason to do it on a subset — the economics that make sampling sensible for a human-audited comparison do not apply.
  • You need the failures named. This site publishes what it cannot reconcile, by ticker, with the reason.
Choose Financial Datasets AI

If any of these is you.

  • You are building for agents and need more than the three statements. Earnings, insider trades, segment breakdowns, prices. This covers the statements only, and that is a real limit.
  • Their verification is against the filing itself. That is a stronger check than an identity test on the points it covers — a human reading a 10-Q catches things arithmetic cannot. The trade is coverage for depth, and it can go either way.
  • You want one vendor for an AI product.

Common questions

What is the difference between BalanceProof and Financial Datasets AI?

BalanceProof: A verification layer over SEC balance sheets. Financial Datasets AI: “The first platform designed for AI agents” — their words.

When should I choose Financial Datasets AI instead of BalanceProof?

You are building for agents and need more than the three statements. Earnings, insider trades, segment breakdowns, prices. This covers the statements only, and that is a real limit. Their verification is against the filing itself. That is a stronger check than an identity test on the points it covers — a human reading a 10-Q catches things arithmetic cannot. The trade is coverage for depth, and it can go either way. You want one vendor for an AI product.

When should I choose BalanceProof instead of Financial Datasets AI?

A sample tells you a rate; a test tells you about YOUR company. Auditing 1,000 companies establishes that the data is good in general. It cannot tell you whether the filing you are about to publish a number from reconciles, because your company may not be in the sample. Each filing here is tested when it reports the lines needed, and the answer travels with the response as a check object saying whose gap a failure is; one that cannot be tested says so. The identity needs no sampling because it is free. Checking A = L + E costs one subtraction per filing, so there is no reason to do it on a subset — the economics that make sampling sensible for a human-audited comparison do not apply. You need the failures named. This site publishes what it cannot reconcile, by ticker, with the reason.

The method, written out

These pages compare on METHOD because method is the part that stays true. Both of these are the method itself rather than an argument about it, so you can judge the claim rather than take it.

Compare

The same question from the other directions. Every one of these compares on method, for the reason at the top of this page.