How far is your AI investing agent from being the next Buffett?

The rise of AI these past few years has shown almost every industry what disruption could look like, and finance is where countless people want to innovate. Yet so far, AI can do little more than assist a human in researching a company. It can't actually invest on its own, unsupervised, and the reports it produces, however polished, still need a human to check them.

We've watched a wave of projects chase the "AI fund", but a handful of prompts can't carry something as serious as investing. We tried it ourselves. We're a team out of finance (CFAs, fund managers, accountants, XBRL specialists, people who've spent years inside filings), and the conclusion was blunt: to this day, no truly autonomous investing agent exists.

But the model was never the problem. We're increasingly convinced that a frontier model already out-reasons any human, even the likes of Buffett and Munger, whose methodology and judgment can be reproduced. The wall everyone hits sits one layer down: the data layer.

Our AI was confidently wrong

Our agent lied. It cut corners. And above all, it hallucinated numbers. It would state a company's revenue, a segment's margin, a footnote detail with total confidence, and be wrong.

Here's the flavor of it. Ask for a company's operating margin and the agent answers a confident "38.2%", even walking you through the line items it used. Open the filing: 3.82%. One slipped decimal, and every conclusion stacked on top of that number is quietly ruined.

And the underlying model is rarely the culprit. Give it clean, structured data and it reasons just fine. Hand it a raw filing, though, and the hallucinations come in three kinds:

• Numeric: A misplaced decimal, thousands read as millions, the scale quietly off.
• Dimensional: Not knowing what a number even is (segment or consolidated, GAAP or non-GAAP).
• Temporal: The right figure from the wrong period, delivered with total confidence. The most common, and the most dangerous.

And beneath all three sits a quieter problem: every model has a training cutoff. A filing published last week has never been seen by any model; ask about it, and what you get is a number from its old memories, dressed up as an answer. Financial data runs on freshness, and a stale quarter can flip a conclusion.

You can't prompt your way out of that. If your agent can't actually understand the numbers going in, every conclusion coming out is wrong, no matter how good the model is. The only fix is at the data layer: turn the filing into something an agent can genuinely read. So we went looking for financial data clean, deep, and traceable enough to trust. We assumed it existed. It didn't.

What's wrong with the data everyone else ships

We worked through most of the financial-data APIs on the market. Almost all fail in at least one of three ways.

They're too shallow. The income statement comes back as a handful of standard lines (revenue, gross profit, net income), and nothing underneath. But the insight isn't in the standard lines. It's in how a company breaks itself down: iPhone vs. Mac vs. Services, Greater China vs. the Americas, the segment quietly shrinking while the headline grows. That granularity is exactly what gets flattened away.

They hand you naked numbers. Even when a figure is right, it usually arrives stripped bare: no link back to the SEC filing it came from, so you can't verify it and your agent's decisions can't be audited. And a number on its own means nothing. Is "7,961" revenue or cost? For which company, which product? Millions or thousands? Which period? Without that, a model can only guess. What it needs is the layer of metadata behind the number: what line item this is, on what basis, for which period, from which filing. Fill that in, and it can finally read.

They drop the footnotes. Most APIs hand you the big tables and throw out the footnotes and the MD&A. But the footnotes are where each company is most itself. Most analysts skim past them. And here's the thing: footnotes are exactly what Buffett and Munger read every single day.

What we built

In investing, the headline numbers have never bought anyone an edge. The edge comes from connecting all that detail into judgment, the work Buffett and Munger have done by hand for sixty years. We can't hand an agent that judgment. But we can hand it Buffett's eyes: a complete, faithful, machine-readable record of what every company actually said.

So we built the data layer we wished we'd had. Akkru turns company filings into data an AI agent can actually read and actually trust:

• As filed, not flattened. Full statements plus the footnote and MD&A detail others drop, down to individual product lines and per-region segments. And not just 10-Ks and 10-Qs: fund holdings (13F), insider transactions (Forms 3/4/5), material events (8-K), and more are coming into the same structured, sourced shape.
• Every number, one click to the source. Each data point links to the exact spot in the SEC filing. Every call your agent makes can be audited, no black boxes.
• Built for agents to read. Structured, dimensional, consistent, and every number and line of text carries a thick layer of metadata behind it, so your agent truly understands a figure instead of guessing at raw HTML or flat JSON.
• Fresh, not from memory. Models have training cutoffs, and anything filed after that date simply isn't in their heads. Akkru ingests new filings as they're published, structured and sourced, so your agent reads what the company just said rather than what the model half-remembers.
• Metrics computed, math left open. We calculate the ratios and show every formula. Verify ours, or swap in your own.
• Institutional-grade, without the institutional price. The depth desks pay six figures for, at a price a solo builder can run on.

And connecting is the easy part: our MCP server is live, so any MCP-compatible agent (Claude included) can plug into Akkru and start pulling sourced, dimensional financials in minutes. No scraping, no parsing pipeline to build.

See it for yourself

The headline numbers are the easy part. The depth (the part that separates a good analysis from a wrong one) lives in the footnotes, which is exactly what flat APIs throw away. Here are four simple questions, put to three very different companies (a chipmaker, a drugmaker, a software giant), and what Akkru hands your agent for each. Every answer links back to the source.

"How concentrated is NVIDIA's business, really?"

A flat API gives you one number: $215.9B of FY2026 revenue. Akkru breaks it down by market and shows that Data Center alone is $193.7B, close to 90% of the entire company. The single most important fact about NVIDIA never shows up in the total. It's sitting right there in the footnote.

"Which products actually drive Pfizer's $62.6B?"

A flat API gives you total revenue. Akkru gives you the whole product table from the footnotes. That's roughly 40 individual products, each one sourced: Eliquis $8.0B, the Prevnar family $6.5B, Vyndaqel $6.4B, Comirnaty $4.4B, Ibrance $4.1B, Paxlovid $2.4B, and on down the list.

Show the full table, all 39 products, as filed (FY2025, $ millions)

Product	Segment	FY2025 ($M)
Eliquis	Primary Care	7,961
Prevnar / Prevenar family	Primary Care	6,494
Vyndaqel family	Specialty Care	6,380
Comirnaty	Primary Care	4,367
Ibrance	Oncology	4,122
Other Hospital	Specialty Care	4,030
Other Primary Care	Primary Care	2,860
Paxlovid	Primary Care	2,362
Xtandi	Oncology	2,194
Padcev	Oncology	1,940
Other Specialty Care	Specialty Care	1,588
Nurtec ODT / Vydura	Primary Care	1,424
Pfizer CentreOne	—	1,338
Oncology biosimilars	Oncology	1,301
Other Oncology	Oncology	1,127
Xeljanz	Specialty Care	1,087
Abrysvo	Primary Care	1,033
Lorbrena	Oncology	1,023
Inlyta	Oncology	923
Adcetris	Oncology	907
Braftovi / Mektovi	Oncology	716
Sulperazon	Specialty Care	653
Inflectra / Remsima	Specialty Care	646
Zavicefta	Specialty Care	638
Enbrel	Specialty Care	627
Bosulif	Oncology	611
Tukysa	Oncology	463
Aromasin	Oncology	450
Genotropin	Specialty Care	446
Orgovyx	Oncology	421
Octagam	Specialty Care	418
Zithromax / Zmax	Specialty Care	399
Cresemba	Specialty Care	349
FSME-IMMUN / TicoVac	Primary Care	319
Elrexfio	Oncology	304
Cibinqo	Specialty Care	284
Talzenna	Oncology	182
Tivdak	Oncology	147
Pfizer Ignite	—	41

"How much future revenue has Microsoft already locked in?"

A flat API can't tell you: it's a forward-looking footnote disclosure. Akkru gives you Microsoft's remaining performance obligations: $375B of contracted revenue not yet recognized, plus $67.3B of unearned revenue on the balance sheet.

"And what does NVIDIA actually say about its credit risk?"

Footnotes aren't only tables; they're text, and Akkru returns that too. Ask about NVIDIA's concentration of credit risk and you get the actual words of the disclosure: "Financial instruments that potentially subject us to concentrations of credit risk consist primarily of cash equivalents, marketable securities, lease guarantees, and accounts receivable…", quoted from the filing, linked to the exact spot.

Tables and text, across companies and sectors, every figure and every sentence traceable to the source. That's what separates an agent that parrots a revenue number from one that can actually read a company.

But can't the AI just read the filing itself?

Fair question. Try this first: take any 10-K, hand it to your favorite model, and ask it to pull out every table, completely and correctly. We'll bet you $100 it can't. Not as of the day we're publishing this.

The reason is simple. A filing is hundreds of pages where the meaning of a number lives in its relationships: number to number, number to text, text to text, across dozens of dimensions and a web of arithmetic. Models still can't hold all of that in context reliably, let alone reason over it without dropping or inventing pieces.

And even when it sort of works, three problems remain:

• It's expensive. Feeding whole filings into a model and making it reason over them, for every question and every company, burns a staggering number of tokens. The bill adds up fast.
• You can't see its work. You don't know whether each calculation was right. It answers a slightly different question than the one you asked. It digs in and won't back down when it's wrong. (Anyone who's used AI for serious research knows the feeling.)
• It was never trained on this. The deep, granular disclosure we provide (the footnotes, the segment detail, the latest filings) mostly isn't in any model's training data. So even the best model can't "just know" it.

That's the whole point of Akkru. We do the hard part once (parse the filing into clean, dimensional, fully-sourced data), so your agent reads numbers it can trust instead of guessing at a PDF. Every figure it uses links back to the original. Every step is checkable, which means it's supervisable. And what you can supervise, you can correct, all the way to 100% accuracy. The black box you had to take on faith is gone.

Who Akkru is for

It doesn't matter whether you're building an AI investing agent or a quant strategy; running an AI platform company that audits public companies or operates a fully autonomous hedge fund; shipping any fintech product that needs fundamentals (a brokerage, a prediction market, anything); or simply doing research and analysis. You can build on Akkru.

Because what we provide is one thing: public-company fundamentals that are AI-native and AI-friendly; comprehensive and fully auditable, down to every number; and genuinely affordable.

This kind of data used to sit behind a steep wall: tens of thousands of dollars per seat. We don't think the AI era should let a price tag stand between people and innovation. We'd rather grow with the ecosystem than gate it.

Don't take our word for it

You don't have to take any of this on faith. Plug your agent into Akkru (over the MCP server or the API) and see whether it gets sharper on the questions that actually matter.

Better still: connect it to Akkru and to whatever data API you use today, and ask both the same question. Break down NVIDIA's revenue. List what's inside Apple's investment portfolio. Quote what a company says about its own credit risk. The difference shows up in seconds: the segments, the footnotes, the source links that one of them simply doesn't have. Side by side, the things that are unique but actually matter become impossible to miss.

Where we are, and where we're going

We're in closed beta today. As of this writing, we cover the full US Russell 3000, with filings from Korea, Japan, the UK, and several European markets on board as well. Every number is sourced, every statement as-filed.

That's the start. The plan is to cover everything a company tells the public: Forms 3/4/5; 13D, 13F, and the rest; press releases, earnings calls, expert calls, all turned into clean, machine-readable, agent-friendly data. And not just US companies. Every company, everywhere.

Here's the bet behind all of it. People like to say Buffett's later returns faded because he lost his touch. What he actually ran out of was two things: enough opportunities for the size of his capital, and enough hours in a single human life. Now imagine an agent with that same discipline and effectively infinite bandwidth, reading every footnote of every company, every day, without tiring. Like a river finding every stream, it could reach the thousands of small, overlooked situations no one person ever has time for. The gap between great investing and average investing starts to close.

Our vision is simple: the next generation of investing (AI funds, solo builders, anyone who invests with an agent) should stand on fundamentals they can completely trust, for every company on earth. That's the data infrastructure the AI era is missing. We're building it.

If you're building an AI investing agent, we'd love for you to try it, and to tell us where it falls short.

Your agent is only as smart as its data. Let's make it smarter.

— The Akkru Team