Company Financial Data for AI & LLM Applications: What Data Teams Need
A low score on a hallucination leaderboard does not mean a model knows a private company's revenue. This guide covers what the data behind your retrieval layer must contain, and what to test before you trust it.
The AI Data Quality Problem
Vectara measures grounded summarisation; FinanceBench measures answers about company filings
On FinanceBench, a set of 10,231 questions about public-company filings, GPT-4-Turbo with a retrieval system answered incorrectly or refused 81% of the time. That was a 2023 test of public companies with plentiful filings. Private companies are harder: MonetaiQ covers 400M+ private companies against 60,000 public listed ones, and private accounts sit mostly in national registries rather than in the text models read.
The catch is that the headline hallucination numbers measure something else. Vectara's leaderboard tests whether a model stays faithful to a document it is given. It says nothing about whether the model can supply a figure it was never shown.
For credit risk engines, compliance tools, research copilots and KYB automation, the practical order is: define the fields and provenance you need, test a candidate source on companies whose filings you can check yourself, then choose between bulk ingestion and live lookups. The sections below follow that order.
What changed in 2025-2026
- EU AI Act high-risk dates moved. Obligations for Annex III high-risk systems now apply from 2 December 2027, and for Annex I systems (AI in regulated products) from 2 August 2028, per the AI Act implementation timeline. The Digital Omnibus on AI behind this was reported as given final Council approval on 29 June 2026. I could not confirm Official Journal publication, so check it before relying on the dates.
- Vectara reset its benchmark. The leaderboard (updated 22 October 2026) now uses HHEM-2.3 and a private dataset of 7,700+ articles. The lowest rate is 1.8% and the highest 24.2%. Falcon-7B is no longer listed and the old 0.7% Gemini Flash figure is gone. Newer is not automatically better: gemini-2.5-flash-lite scores 3.3%, gemini-3-flash-preview 13.5%.
- MCP is a versioned spec to build against. The Model Context Protocol specification is at version 2026-07-28, with stateless requests, tools and resources, and opt-in extensions such as Tasks for long-running operations. It is a standard way to expose a company-data API to an LLM as a tool.
LLMs Hallucinate Company Financials — It's a Data Problem
Ask any leading LLM for the revenue of a mid-size private company. You'll get a confident answer. A specific number. Maybe even a growth trend.
The problem: that number is fabricated.
Two sources give a fair picture. Vectara’s leaderboard (October 2026) shows rates from 1.8% to 24.2% when summarising a supplied document, depending on the model. FinanceBench shows the harder case: questions about real company filings, where a retrieval-equipped GPT-4-Turbo was wrong or refused 81% of the time. They measure different things, so do not read a financial error rate off either.
Small rates still matter at scale. As arithmetic, not a measured figure: a 2% error rate across 10,000 daily queries is 200 wrong answers a day, and one of them can become a flawed credit decision or a missed compliance flag.
This isn't a model problem. It's a data problem. LLMs fill knowledge gaps with statistically plausible text. When they lack reliable financial data for a specific company, they generate numbers that look right — but have no factual basis. The fix isn't better prompting. It's better data.
Most company financial data available to AI teams today is aggregated — collected from third-party sources, scraped from websites, pulled from secondary databases. By the time it reaches an AI pipeline, the data has passed through multiple hands. Each handoff introduces lag, errors, and gaps.
Take a hypothetical timeline. A company files updated financials with a government registry in January. The aggregator picks it up in March. The data vendor updates their feed in April. Your AI model trains on it in May. That's 4 months of latency on a single data point — and the model has no way to flag it as stale. Now multiply that across millions of companies.
Why Financial Data Hallucinations Are Worse Than Other Domains
Company financials combine four properties that make errors likely, and the cost per error is high because the figure usually feeds a credit, compliance or investment decision.
Entity-specific. There's no general knowledge about what "revenue" should be. It's different for every company, every quarter, every jurisdiction. A model can't generalize — it has to know the specific number or it's guessing.
Time-sensitive. Last year's revenue is wrong if someone asks for this year's. Financial data decays fast. A model trained on 2023 data will produce 2023 answers to 2025 questions — without flagging the gap. Public company data updates quarterly; private company accounts follow each jurisdiction's filing calendar, so check the filing date on every record.
Private companies are invisible. MonetaiQ covers 400M+ private companies against 60,000 public listed ones. Private firms don't file with the SEC, and their accounts sit mostly in national registries, not in widely published text. Ask a model for the revenue of a private German manufacturer and it has little to retrieve, so it estimates. For how German filings work, see our guide to German company financial data from the Bundesanzeiger.
Confidence without evidence. LLMs don't hedge. They deliver a fabricated revenue figure with the same confidence as a verified one. A fabricated figure looks identical to a verified one, so every number needs a source and filing date attached.
RAG Alone Doesn't Fix It
Retrieval-augmented generation is the standard answer to LLM hallucinations. Instead of relying on the model's training memory, you retrieve relevant documents at query time and feed them as context. Grounding helps only when the retrieved text actually contains the answer.
But here's the catch most teams learn the hard way: RAG doesn't solve the financial data problem on its own.
The RAG paradox: Your retrieval layer is only as good as the data in your knowledge base. If that data is scraped, stale, or aggregated from unknown sources — you're grounding your LLM in someone else's errors. A Stanford study of commercial legal research tools (2024) found hallucination rates above 17% for Lexis+ AI and Ask Practical Law AI and above 34% for Westlaw AI-Assisted Research. On FinanceBench, a retrieval setup was wrong or refused 81% of the time.
Common failure patterns in financial RAG implementations:
- Stale data. Revenue figures from 2-year-old web scrapes sit in the vector store and get treated as current. The model has no way to flag them as outdated — it serves them with full confidence.
- Missing coverage. Many financial knowledge bases hold only the roughly 60,000 listed companies, which leaves private companies uncovered. Any query about a private entity falls back to the model's imagination.
- Inconsistent formats. UK Companies House, German Bundesanzeiger, and France's INPI all use different schemas and line items. When your retrieval pulls from multiple jurisdictions, the results don't align — and the model papers over the gaps.
- No source metadata. Retrieved chunks carry no provenance. There's no way to tell if a revenue figure came from an official registry filing or a blog post.
RAG is necessary but not sufficient. The quality, recency, and provenance of your underlying financial data determine whether RAG actually reduces hallucinations or just makes them more convincing.
What Data Teams Actually Need From a Financial Data Source
AI applications have requirements that go beyond what a traditional BI dashboard needs. Here's the minimum spec for financial data feeding an AI pipeline:
Core financial metrics. Revenue, net income, EBITDA, operating margins, growth rates. Balance sheet: total assets, liabilities, equity, debt ratios. Cash flow statements. These are the foundational inputs for any financial reasoning model.
Filing metadata. When was the statement filed? For which period? Under which accounting standard (IFRS, local GAAP, US GAAP)? Without this, the model can't distinguish between current and historical figures or compare across jurisdictions.
Entity identifiers. Registration numbers, LEI codes, tax IDs — so the model reliably matches financial data to the right company. Without clean identifiers, entity resolution fails and the model conflates companies with similar names. The LEI file of 6 November 2026 holds about 3.5 million records, so for most private companies the national registration number is the only reliable key.
Source provenance. Every data point should carry metadata showing: the original source, filing date, reporting period, extraction date, and verification status. This is the audit trail that regulators and enterprise buyers will ask about.
Machine-readable delivery. JSON, CSV, or Parquet — formats that plug directly into data pipelines. Not PDFs. Not dashboard-only access.
Bulk access. You can't train a model or build a knowledge base one API call at a time. Data teams need full datasets delivered for ingestion — millions of records across jurisdictions.
How to Integrate Financial Data Into Your LLM Pipeline
There are three primary patterns for using company financial data in AI applications:
Pattern 1: RAG Grounding
Ingest verified financial data into a vector store or search index. At query time, retrieve relevant records and pass them to the LLM as context. The model generates answers grounded in real data instead of its training memory. Re-ingest on the source's filing cadence, and store filing date and period as fields the model must quote.
Pattern 2: Fine-Tuning Datasets
Use bulk verified financial statements as training data. Teach the model what real company financials look like — correct value ranges by industry, relationships between metrics, standard reporting patterns. This reduces hallucination at the model level.
Pattern 3: Evaluation & Benchmarking
Use verified financials as ground truth to measure your model's accuracy. Ask the model about companies where you know the real numbers. Compare outputs against verified data. Track hallucination rates over time.
What to do first: build a gold set of 50 to 100 companies whose filings you have read yourself, with at least a third private and several jurisdictions. Score the model with and without your data source, and count refusals separately from wrong numbers. FinanceBench follows the same logic: answers are scored against figures in the filings.
Pattern 4: Live Tool Calls via MCP
Bulk ingestion goes stale; a tool call fetches the current record. With the Model Context Protocol, you expose lookups (search by registration number, fetch latest accounts) as tools the model calls at answer time, and each response can carry source, filing date and period. Use bulk plus RAG for screening and similarity search across many companies, and tool calls for single-entity answers that feed a decision. MonetaiQ lists MCP among its features, alongside its API and bulk feeds.
Registry-Sourced vs. Aggregated: Why the Source Matters
There are two fundamentally different approaches to sourcing company financial data. For AI applications, the difference has direct implications for model reliability.
| Registry-Sourced | Aggregated | |
|---|---|---|
| Origin | Direct from government registries | Compiled from third-party sources |
| Verification | Traceable to original filing | Source often unknown |
| Freshness | Updated when company files | Depends on aggregator's schedule |
| Audit trail | ✓ Full provenance | ✗ Usually opaque |
| AI suitability | Verifiable ground truth | Black box inputs |
| Regulatory compliance | Makes data origin easier to document | Origin harder to document |
For AI teams that care about model reliability, registry-sourced data gives you a verifiable foundation. Aggregated data gives you a black box. The EU AI Act's Article 10 requires high-risk systems to run data governance covering how data was collected, where it came from and how it was prepared, plus checks for gaps and bias. It does not mandate registries. High-risk obligations now apply from 2 December 2027 (see above). Penalties are tiered under Article 99: up to €35 million or 7% of worldwide turnover is for prohibited practices, while breaches of high-risk obligations carry up to €15 million or 3%. Whether your system is high-risk is a question for counsel.
The Private Company Gap
Most financial data APIs cover public companies well. SEC filings, stock exchange data, and earnings transcripts are structured and accessible. But on MonetaiQ's numbers, 60,000 listed companies sit against 400M+ private ones.
In the EU, accounting rules require many private companies to file annual accounts, and registries in the UK, Germany, France and elsewhere hold them. Filing duties and public access differ by country, and small companies often file reduced accounts. Our guide to finding private company financials and comparison of European financial data APIs cover what each source returns.
This gap has direct consequences for AI products:
- A KYB verification tool that only checks public companies misses most of the entities it needs to screen.
- A credit risk model trained only on public company financials can't score the private supplier your client is onboarding — and private companies represent the majority of global supply chains.
- A market intelligence copilot that only knows listed companies gives an incomplete picture. In many industries most competitors are private.
The gap isn't just coverage. Private companies behave differently — revenue patterns, capital structures, and growth trajectories don't mirror public markets. Training an AI model exclusively on public data introduces structural bias into every output. A model that's only seen $500M+ revenue companies will systematically misjudge a $5M private firm.
Data teams need a source that covers both public and private companies, with financials sourced directly from the registries where those companies file. Once you have the data, how to evaluate private company financials covers what to check in the numbers themselves.
Evaluating a Financial Data Provider for AI Workloads
Not every financial data provider is built for AI. Most were designed for human analysts querying a dashboard. Here's the evaluation checklist:
| Criteria | What to Ask |
|---|---|
| Bulk data access | Can you download full datasets, or only query individual records? |
| Incremental updates | Do they offer delta feeds or change logs for ongoing refreshes? |
| Source transparency | Can you trace every data point back to the original registry filing? |
| Private company coverage | Do they cover private companies, or only public/listed? |
| Structured formats | Is data available in JSON, CSV, Parquet — or only via dashboard? |
| Global coverage | How many registries do they source from? How many countries? |
| Schema consistency | Is data normalized across jurisdictions, or do you map each country separately? |
| Freshness | How quickly after a company files does the data appear in their system? |
| OCR accuracy | How do they extract data from non-digital filings? What's the accuracy rate? |
| Audit metadata | Does each record carry source, filing date, reporting period, and last verified? |
If a provider checks all ten boxes, they're built for AI workloads. If they check five or six, you'll spend more time cleaning data than building your product.
How MonetaiQ Solves This
Everything described above — the data quality challenges, coverage gaps, delivery requirements — is what MonetaiQ was built to address.
MonetaiQ provides structured financial data for 400M+ private companies and 60,000 public listed companies across 100+ countries. Income statements, balance sheets, cash flow statements, P&L breakdowns, and key financial ratios.
- API and bulk data feeds. Scalable API for real-time lookups. Bulk feeds for large-scale ingestion. JSON, CSV, XML, and custom formats.
- Private and public coverage. 400M+ private companies sourced from government filings. Public data from stock exchanges and regulatory filings, updated quarterly.
- Registry-sourced. Data from official stock exchange filings, regulatory bodies, and government registries. Every data point traceable to its origin.
- Validated and structured. All data undergoes validation and quality control. Non-digital filings are extracted, verified, and structured.
- Flexible delivery. Customizable by metrics, companies, industries, or regions. Daily, weekly, or monthly frequency.
MonetaiQ sources company data directly from 400 government registries worldwide. That registry-sourced foundation means AI teams get financial data with the accuracy, freshness, and audit trail that production AI systems demand.
Build AI Products on Verified Financial Data
400M+ companies. Registry-sourced. API and bulk delivery. Start with a free trial.
Frequently Asked Questions
LLMs predict the next token based on training patterns. When they lack reliable financial data for a specific company — which is common for private companies — they fill the gap with statistically plausible numbers. The output looks authoritative but has no factual basis. This is especially pronounced for entity-specific data like revenue, profit margins, and balance sheet items.
RAG is an architecture where the AI retrieves relevant financial records from a knowledge base before generating a response. Instead of relying on training memory, the model references actual data. This reduces hallucination — but only if the knowledge base contains verified, current, and well-sourced financial data.
Core requirements include income statements, balance sheets, cash flow statements, profitability metrics (EBITDA, operating margins), filing metadata (dates, reporting periods, accounting standards), and entity identifiers (registration numbers, LEI codes). All data must be structured, machine-readable, and traceable to its original source.
MonetaiQ covers 400M+ private companies against 60,000 public listed ones. AI products for KYB verification, credit risk, supply chain monitoring, and market intelligence are incomplete without private company coverage. Models trained only on public company data develop structural biases and can't serve the majority of real-world use cases.
Registry-sourced data comes directly from government business registries — the official repositories where companies file their accounts. Aggregated data is compiled from multiple third-party sources, introducing lag, opacity, and potential errors. For AI applications, registry-sourced data provides verifiable ground truth with full audit trails.
Yes, but not all providers support it. Many only offer API or dashboard access. For AI workloads — model training, fine-tuning, RAG knowledge bases — you need bulk delivery of full datasets in machine-readable formats (JSON, CSV, Parquet). MonetaiQ offers both API access and customizable bulk data feeds.
Article 10 of the EU AI Act requires high-risk systems to document how data was collected, where it came from and how it was prepared. Credit scoring is an Annex III high-risk use, and those obligations now apply from 2 December 2027 under the Digital Omnibus changes. Registry-sourced data with audit trails makes that documentation easier, but the Act does not mandate it. Confirm your classification with counsel.
It depends on the use case. Monthly refreshes are a baseline for general intelligence. Weekly or more frequent updates are needed for credit risk and compliance monitoring. MonetaiQ offers flexible delivery frequency — daily, weekly, or monthly — with incremental updates so you're not re-ingesting entire datasets.
MonetaiQ provides financial data for over 400 million private companies and 60,000 publicly listed companies across 100+ countries. Data is sourced from government registries, stock exchange filings, and regulatory bodies — and delivered via scalable API or bulk data feeds.
MonetaiQ data is available in JSON, CSV, XML, and custom formats. For RAG: ingest bulk data into your vector store, then query via API for real-time updates. For training: use bulk feeds as fine-tuning datasets. For evaluation: use verified financials as ground truth benchmarks. The API supports company name search, registration number lookup, website URL, and LinkedIn URL matching.
Sources
Checked November 2026.
- Vectara hallucination leaderboard (HHEM-2.3, updated 22 Oct 2026)
- FinanceBench: a new benchmark for financial question answering (arXiv 2311.11944)
- Stanford HAI: AI on trial, legal models hallucinate (2024)
- EU AI Act implementation timeline
- EU AI Act, Article 10: data and data governance
- EU AI Act, Article 99: penalties
- Council final approval of the AI Omnibus (secondary summary, 30 June 2026)
- Model Context Protocol specification (version 2026-07-28)
- GLEIF concatenated LEI files