How AI Is Changing the Speed and Accuracy of Financial Analysis
I spend a good part of my week building AI systems that read company filings, so I hear some version of the same question constantly: how much of financial analysis can the machines actually do now? The honest answer is more than the skeptics think and less than the vendors claim. The gap between those two lines is where the practical value sits, so let me map where things actually stand.
The old workflow was mostly transcription
A thorough company analysis used to take days, sometimes a week per name, and the breakdown of those hours was a little embarrassing. Pull the 10-K off EDGAR. Copy numbers into a spreadsheet. Tie out the totals. Normalize three to five years of history. Build the model, write the memo, send it up the chain. Very little of that time went to judgment. Most of it went to moving numbers from one document into another and confirming they arrived intact.
The slowness had a defensive logic. Manual work forced you to touch every number, and touching every number caught mistakes. But it also capped how many companies one person could cover, and it meant the expensive part of an analyst's brain sat idle while the cheap part did data entry. That tradeoff is the thing AI has genuinely changed, and it's worth being precise about how.
Where AI actually helps right now
Strip away the vendor decks and there are four or five places where the technology is already earning its keep.
Extraction. Pulling structured data out of unstructured documents was the tax you paid before analysis could start. Language models are good at this now. Hand one a 10-K and ask for three years of segment revenue, the debt maturity schedule, or the exact wording of the revenue recognition footnote, and you get it back in seconds instead of an afternoon. The catch, which I'll come back to, is that you still have to verify what comes back.
Screening. Running multi-factor screens across thousands of listed companies used to mean an expensive terminal subscription plus a lot of manual cleanup. AI-backed platforms run those screens in seconds, and the better ones let you refine the results in plain English instead of a query language.
Tone and language. There's real academic weight behind this one. Loughran and McDonald showed back in 2011 that general-purpose sentiment dictionaries misread financial text (a word like "liability" is neutral in a filing and negative almost everywhere else), so they built finance-specific word lists that are still in wide use. Current models go further. They can flag when management's language turns hedged, when a CFO stops answering a question directly, or when the risk factors section quietly gained a paragraph since last year. None of that is a trading signal by itself, but it tells you exactly where to dig.
Anomaly flags. The classic tools here are decades old and still useful. The Altman Z-Score, where readings below 1.8 sit in the distress zone, and the Beneish M-Score for earnings manipulation both came out of academic research, and both are simple enough to compute by hand. What machine learning adds is breadth. A model can watch far more variables at once than any person and surface the two or three relationships that look off, like receivables growing much faster than revenue, or margins expanding while inventory swells.
First drafts. A model can turn a financial dataset into a competent first-pass writeup. It won't have a view, and it shouldn't, but it kills the blank page and it rarely forgets a section.
A worked example: one 10-K in an afternoon
Here's how this plays out on a real filing. Say you're looking at a mid-cap industrial you've never touched. The old first day went entirely to data collection. The new sequence looks like this.
- Pull the last three 10-Ks and the most recent proxy statement from EDGAR. This is free and takes about five minutes.
- Have the model extract the financial statements plus the footnotes that matter most: revenue recognition, segments, debt, leases, commitments and contingencies, related-party transactions. Ask it to cite the page or section for every number it returns.
- Ask for the basic diagnostic ratios, and more importantly the gaps between them. Say the company reports revenue up 9% while receivables are up 30%. That gap is either a timing story or a channel-stuffing story, and working out which one is a genuinely good use of your afternoon.
- Run the same extraction on two or three competitors so your ratios have context instead of floating in space.
- From the proxy statement, pull how executives actually get paid. If the bonus plan keys on revenue growth and receivables are ballooning, you've just connected two dots that each looked boring on its own.
The mechanical steps that used to eat the first day now take under an hour. The judgment steps take as long as they always did, which is fine, because now they're where all your time goes.
The accuracy question
Speed only matters if the numbers hold up, and here the picture is genuinely mixed.
On structured work like extraction, calculation, and cross-referencing, models are strong and getting stronger. But they fail in a specific, dangerous way: when a model doesn't know a number, it will sometimes produce a plausible one, confidently formatted, with no hint that it's guessing. A fabricated figure that looks hand-checked is worse than a blank cell. The fix is boring process rather than clever prompting. Require citations back to the source document, spot-check anything that feeds a decision, and never let a model's number into your spreadsheet without seeing it in the filing first.
On judgment work, accuracy is harder to even define. A DCF is only as good as its growth, margin, and discount-rate assumptions, and no model can tell you whether a company deserves a decade of double-digit growth in your terminal value. What a model can do is keep you honest. It can show you historical base rates, show you what comparable companies actually achieved, and make it awkward to type in a hero assumption without noticing you're doing it.
The practical sweet spot is breadth. An analyst working with these tools covers more companies, tests more scenarios, and checks more sources than one working by hand. The judgment stays human, but the information base underneath it gets much wider.
Why so many AI projects in finance disappoint
Plenty of AI initiatives inside finance teams quietly miss their targets. Having watched a number of these up close in consulting work, I find the failure patterns boringly consistent. Teams try to automate the judgment instead of the mechanics. Or they buy a capability first and go looking for a problem second. Or, most often, they underestimate the unglamorous data work: cleaning inputs, wiring up verification, deciding exactly what a human reviews and when.
The projects that work tend to be narrow and measurable. Filing extraction with citation checks. Screening against predefined criteria. Transcript summaries linked back to the source audio. Clear inputs, clear outputs, and an obvious way to tell whether the thing is earning its cost. If a proposed use case can't be scored, that's usually a sign it's aimed at judgment, and judgment is the part you want to keep for yourself anyway.
Agents and multi-model analysis
The current frontier is multi-agent systems, where specialized models each handle a slice of the work: one extracts, one builds comparables, one reads the proxy, one checks the others' output, and a coordinating layer assembles the result. That division of labor mirrors a human analyst team, with a junior collecting data, a senior building the model, and somebody experienced signing off.
Standards like the Model Context Protocol, which gives models a common way to plug into data sources and tools, are making these systems much easier to wire together. What the demos tend to skip is that errors compound in a chain. An agent pipeline is only as reliable as its weakest verification step, so the teams doing this well pour most of their effort into the checking layer and treat the impressive choreography as the easy part.
What this means if you don't have a Bloomberg
Serious analysis used to be gated by tooling costs. A Bloomberg Terminal runs north of $20,000 a year, and institutional platforms like Capital IQ and PitchBook also price well into five figures. For decades, that pricing quietly decided who got to do professional-grade company analysis.
The primary sources were never the expensive part, though. EDGAR is free, and it's the same repository the professionals read. Pair it with AI tools that handle the extraction and screening work the terminals used to justify their price with, and a serious individual investor can now run analyses that would have needed a small team ten years ago.
The skill that matters is shifting accordingly, away from collecting data and toward interrogating it. Knowing how to pull numbers out of a filing is worth a little less every quarter. Knowing which questions to ask, which gaps between numbers deserve an afternoon, and when a clean-looking answer needs a second source is worth more.
Working rules for trusting the machine
Executives are right to stay wary of AI-generated analysis, mainly because the failure mode is so quiet. A model can produce a valuation that looks reasonable and is wrong for reasons you can't see, and many of the strongest models still won't show you a reasoning chain you'd be comfortable defending in an investment committee memo. Until you can reproduce the logic yourself, you don't really own the conclusion.
The stance that works in practice is to treat the model as a very fast junior analyst whose work always gets reviewed. A few rules worth making non-negotiable:
- Every extracted number carries a citation to the filing, and you click through on the ones that matter.
- Nothing generated goes into a decision document without a human pass. Drafts stay drafts until someone accountable has read them.
- Keep a running log of where the model got things wrong. Error patterns are surprisingly stable, and knowing yours tells you where to concentrate review time.
- Spend the recovered hours on the work that used to get squeezed: reading footnotes, comparing this year's proxy against last year's, talking to a customer. Saved time only counts if it goes somewhere useful.
The people getting the most out of these tools run that loop on everything. Delegate the mechanics, verify the output, spend the freed-up time on judgment. Models will keep improving, but that division of labor has held steady for a while now, and it's a sensible basis for planning your own workflow.