--- title: "Why AI Agents Need a Trusted Data Layer" slug: why-ai-agents-need-a-trusted-data-layer date: 2026-07-16 excerpt: "Half of new web articles are now AI-written and their citations often don't hold up. A trusted data layer gives agents structured, cited SEC filings to read before the noise." category: Data Infrastructure tags: ["Data Infrastructure", "AI & Agents", "Research Workflows"] author: Jennifer Ma authorRole: "Co-founder & CEO" coverImage: /images/blog/why-ai-agents-need-a-trusted-data-layer/cover.png cta: product featured: true status: published --- An AI agent is only as trustworthy as the data it reads — and the data on the open web is getting worse. More than half of newly published articles are now AI-generated, and the citations inside machine-written text frequently don't support what they claim. A trusted data layer fixes the input side of that problem: it puts structured, cited SEC filings and company disclosures in front of your agent *before* it wanders into the open web, so its answers come out more accurate, cheaper in tokens, and faster to review. **TLDR:** - Connected, cited data measurably raises accuracy: with the right documents in context, a model's error rate on financial questions drops from [81% to 21%](https://arxiv.org/abs/2311.11944). - Structured data is cheaper and easier to check: agents load a cited *section*, not a 50,000-word filing, and every value keeps a link an analyst can verify in one click. - The web your agent reads is now mostly machine-written — [74% of new pages contain AI content](https://ahrefs.com/blog/what-percentage-of-new-content-is-ai-generated/) and only [51.5% of generative-search sentences](https://arxiv.org/abs/2304.09848) are fully supported by their own citations. Read the trusted disclosure first. ## The same question, two different answers Here is the difference a data layer makes, from our own [live demo](/demo): the same agent, asked the same question about Micron, answering first from connected FocusAlpha sources and then from a direct web search. ![Side-by-side comparison: an agent answering a question about Micron from connected FocusAlpha sources, with every figure cited, versus the same agent answering from a direct web search with uncited figures](/images/blog/why-ai-agents-need-a-trusted-data-layer/fig-1.png) *Figure 1: The same agent, the same question — connected FocusAlpha sources versus a direct web search. Source: [FocusAlpha live demo](/demo).* Both answers read fluently. Only one is auditable. On the left, every segment figure and driver traces back to a specific earnings call in a connected research database; on the right, the numbers are plausible, unattributed, and impossible to check without redoing the work. The gap between "sounds right" and "can be verified" is exactly what a trusted data layer closes — and it is the difference that determines whether an agent's output survives a second look. ## Trusted data raises accuracy This isn't a matter of taste; it's measured. The single biggest lever on an agent's answer quality is not the model — it's whether the right, trusted documents are in front of it. On [FinanceBench](https://arxiv.org/abs/2311.11944), the same GPT-4-Turbo fails 81% of financial questions when it has to retrieve from raw filings itself, and only 21% when the correct documents are placed in context. | Condition | Result | | --- | --- | | Naive retrieval pipeline over raw filings | GPT-4-Turbo answers **81% of questions incorrectly** | | Same model, correct documents in context | **21% incorrect** | | Free-text summary of a 10-Q MD&A | flips the bull / neutral / bear call in **21.3–53.0% of cases** (11.0% baseline) | | Whole documents dumped into long context | **all 18 frontier models tested degrade** as input grows | *Sources: [FinanceBench, arXiv:2311.11944](https://arxiv.org/abs/2311.11944), long-context configuration via [Patronus AI](https://www.patronus.ai/announcements/patronus-ai-launches-financebench-the-industrys-first-benchmark-for-llm-performance-on-financial-questions); [When Summaries Distort Decisions, arXiv:2606.29251](https://arxiv.org/abs/2606.29251); [Context Rot, Chroma Research](https://www.trychroma.com/research/context-rot).* ![Bar chart: a naive retrieval pipeline over raw filings answers 81% of financial questions incorrectly, versus 21% when the correct documents are already in context](/images/blog/why-ai-agents-need-a-trusted-data-layer/fig-2.svg) *Figure 2: Grounding an agent in the right documents cuts its error rate nearly fourfold. Source: [FinanceBench, arXiv:2311.11944](https://arxiv.org/abs/2311.11944).* A trusted data layer's whole job is to make the second row the default. When filings arrive normalized and every value carries its citation, the agent spends its reasoning on the question rather than on parsing, deduplicating, and hoping it grabbed the right table. We've written about the two ways this goes wrong when it isn't the default — [what compression does to judgment](/blog/when-ai-compresses-your-filings) and [why a smarter model can't fix worse data](/blog/why-a-smarter-model-cant-fix-worse-data). ## Structured data costs fewer tokens and less review effort Accuracy is the headline; cost and reviewability are the reasons this pays for itself. Raw filings are enormous and getting bigger — the median 10-K's text roughly doubled between 1996 and 2013 — so an agent that loads a whole document to extract one number burns tens of thousands of tokens and, worse, degrades as the context fills. Structured retrieval inverts that: pull *Item 1A*, or a single normalized line item, and the token bill drops by orders of magnitude while the answer sharpens. | Cost driver | Measurement | | --- | --- | | Length of a single 10-K (median body text) | grew from **~23,000 words (1996) to ~50,000 words (2013)** — well over 60,000 tokens | | Long-context penalty | **all 18 frontier models degrade** as input grows; a ~300-token focused context beat ~113,000-token prompts | | Filings later corrected (2025) | **708 amended 10-Ks, 1,239 amended 13F-HRs, 1,453 amended 8-Ks** | *Sources: [The evolution of 10-K textual disclosure, Journal of Accounting and Economics](https://www.sciencedirect.com/science/article/abs/pii/S0165410117300484); [Context Rot, Chroma Research](https://www.trychroma.com/research/context-rot); [SEC DERA, Number of EDGAR Filings by Form Type](https://www.sec.gov/data-research/sec-markets-data/number-edgar-filings-form-type).* The second saving is human. When every number a data layer returns keeps a citation to the source document, a reviewer verifies a claim in one click instead of re-deriving where it came from — and the more detail the layer carries (the exact filing, the exact section, the as-amended value), the less review effort each answer costs. That last row matters more than it looks: filers restate, and an agent that scraped naively can confidently cite a figure the company has since corrected. A trusted data layer that resolves filings as-amended keeps that class of error out of the answer entirely. ## The web your agent reads is now written by other agents Here is the part that changed recently. When an agent leaves your connected data and searches the open web, what it finds is increasingly *other machines' output*. AI-generated articles [overtook human-written ones in November 2024](https://graphite.io/five-percent/more-articles-are-now-created-by-ai-than-humans), and nearly three-quarters of new web pages now contain some AI-generated content. That would be fine if the machine-written layer were reliable — but its citations are exactly where it's weakest. | What the numbers say about the open web | Figure | | --- | --- | | New articles that are AI-generated | **~half** — AI overtook human writing in Nov 2024 | | New web pages containing AI-generated content | **74.2%** (April 2025) | | Generative-search sentences fully supported by their citations | **51.5%** | | ChatGPT bibliographic citations that were fabricated | **55%** (GPT-3.5) / **18%** (GPT-4) | | Hallucination rate of leading legal-research AI tools | **17–34%** | *Sources: [Graphite](https://graphite.io/five-percent/more-articles-are-now-created-by-ai-than-humans) (through May 2025); [Ahrefs](https://ahrefs.com/blog/what-percentage-of-new-content-is-ai-generated/) (900k pages, April 2025); [Evaluating Verifiability in Generative Search Engines, arXiv:2304.09848](https://arxiv.org/abs/2304.09848); [Walters & Wilder, Scientific Reports 13:14045](https://pmc.ncbi.nlm.nih.gov/articles/PMC10484980/); [Stanford RegLab & HAI](https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries).* Put those together and the risk is concrete: an agent researching a company can ingest an AI-written summary whose numbers are unsourced or quietly wrong, and then present them to you with the same confident tone as a real filing. You can't stop agents from reading the web, and you shouldn't — commentary and context live there. What you can do is change the order. A trusted data layer guarantees the agent sees the primary filing and the company's own disclosures *first*, as structured, cited fact — so anything it later reads on the open web is weighed against ground truth instead of standing in for it. ## Where FocusAlpha fits FocusAlpha is that trusted data layer. It turns SEC filings (10-K, 10-Q, 8-K, 13F), [earnings-call transcripts](/docs/api/documents), and other trusted company communications into structured, normalized data where every value keeps its citation back to the source document — reachable through a [SEC filings API](/docs/api/filings) or an [MCP server](/docs) built for agents. The demo at the top of this post is the whole thesis in one screen: connect the agent to trusted, cited data, and its answers get more accurate, cheaper to run, and faster to trust. The web your agents read will keep filling with machine-written text. The defense isn't to read less — it's to make sure the first thing every agent reads is the real disclosure, with the citation attached. ## FAQ ### What is a trusted data layer for AI agents? A trusted data layer is a structured, cited source of primary information that an AI agent queries before — or instead of — the open web. Rather than scraping documents and hoping the numbers are right, the agent retrieves normalized values that each carry a citation back to the original source (for company research, that means SEC filings, earnings-call transcripts, and disclosures). It exists because [grounding a model in the correct documents cuts its error rate on financial questions from 81% to 21%](https://arxiv.org/abs/2311.11944). ### Does structured data actually reduce token costs? Yes, substantially. A single 10-K's median text [grew to roughly 50,000 words by 2013](https://www.sciencedirect.com/science/article/abs/pii/S0165410117300484) — well over 60,000 tokens — so loading a whole document to answer one question is expensive and, per [Chroma's Context Rot research](https://www.trychroma.com/research/context-rot), actively degrades the answer as context fills. Retrieving a single cited section or normalized line item instead cuts the token bill by orders of magnitude while improving accuracy. ### How much does grounding improve accuracy on financial questions? On the [FinanceBench benchmark](https://arxiv.org/abs/2311.11944), GPT-4-Turbo answered 81% of financial questions incorrectly when it had to retrieve from raw filings itself, but only 21% incorrectly when the correct documents were already in its context — a nearly fourfold reduction in error rate from the same model, purely by improving the data it reads. ### How much web content is AI-generated, and why does it matter for research? AI-generated articles [overtook human-written ones in November 2024](https://graphite.io/five-percent/more-articles-are-now-created-by-ai-than-humans), and [74.2% of newly created web pages contain some AI-generated content](https://ahrefs.com/blog/what-percentage-of-new-content-is-ai-generated/) as of April 2025. It matters because an agent doing open-web research increasingly ingests other machines' output — and machine-written text is least reliable exactly where it matters most: its citations. ### Are AI-generated citations reliable? Often not. Across generative search engines, [only 51.5% of generated sentences are fully supported by their citations](https://arxiv.org/abs/2304.09848). In peer-reviewed testing, [55% of GPT-3.5's bibliographic citations were fabricated (18% for GPT-4)](https://pmc.ncbi.nlm.nih.gov/articles/PMC10484980/), and leading [legal-research AI tools hallucinate 17–34% of the time](https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries). This is why a trusted data layer, where every value ships with a verifiable source, is a prerequisite for research agents. ### What is the best SEC filings API for AI agents? Evaluate a [SEC filings API](/docs/api/filings) on whether it grounds agents in trusted data: complete coverage, entity normalization, section-level retrieval, and — most importantly — a citation on every value so answers are auditable. FocusAlpha is built to that standard, with filings, transcripts, and disclosures delivered as structured data through an [API or MCP](/docs). ### What is FocusAlpha? FocusAlpha is a [SEC filings API](/docs/api/filings) and agent-ready financial data layer: it turns SEC filings (10-K, 10-Q, 8-K, 13F), earnings-call transcripts, and other trusted company communications into structured, normalized data where every value keeps its citation back to the source document. AI agents connect via [API or MCP](/docs) to research public companies from complete, trusted information. ## Sources - [FinanceBench: A New Benchmark for Financial Question Answering — arXiv:2311.11944](https://arxiv.org/abs/2311.11944) - [Patronus AI — FinanceBench announcement (long-context configuration)](https://www.patronus.ai/announcements/patronus-ai-launches-financebench-the-industrys-first-benchmark-for-llm-performance-on-financial-questions) - [When Summaries Distort Decisions — arXiv:2606.29251](https://arxiv.org/abs/2606.29251) - [Context Rot — Chroma Research](https://www.trychroma.com/research/context-rot) - [The evolution of 10-K textual disclosure — Journal of Accounting and Economics](https://www.sciencedirect.com/science/article/abs/pii/S0165410117300484) - [SEC DERA — Number of EDGAR Filings by Form Type (2025 amendment counts)](https://www.sec.gov/data-research/sec-markets-data/number-edgar-filings-form-type) - [More Articles Are Now Created by AI Than Humans — Graphite](https://graphite.io/five-percent/more-articles-are-now-created-by-ai-than-humans) - [What Percentage of New Content Is AI-Generated? — Ahrefs](https://ahrefs.com/blog/what-percentage-of-new-content-is-ai-generated/) - [Evaluating Verifiability in Generative Search Engines — arXiv:2304.09848](https://arxiv.org/abs/2304.09848) - [Fabrication and errors in the bibliographic citations generated by ChatGPT — Walters & Wilder, Scientific Reports 13:14045](https://pmc.ncbi.nlm.nih.gov/articles/PMC10484980/) - [AI on Trial: Legal Models Hallucinate in 1 out of 6 Queries — Stanford RegLab & HAI](https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries)