Last updated on

AI & Automation

Systems, Wealth & Technology › AI & Automation

This page collects what federal agencies and published studies report about AI tools: how widely businesses use them, what one large study measured about productivity, how often purpose-built tools produced wrong answers in a benchmark, and how the National Institute of Standards and Technology frames AI risk. Each finding is stated with its source, its setting and its date.

What a language model can and cannot look up

The researchers who introduced retrieval-augmented generation described the limit of a model working from its training alone: a pre-trained model’s “ability to access and precisely manipulate knowledge is still limited,” and their approach was to build “models which combine pre-trained parametric and non-parametric memory for language generation,” that is, the model plus a searchable store of documents it can retrieve from (Lewis et al., 2020). Whether a particular AI product searches, retrieves documents or uses other tools depends on how that product is built.

How many businesses use AI

The U.S. Census Bureau’s Business Trends and Outlook Survey found that from December 2025 to May 2026 “overall AI usage hovered between 17% and 20%” of businesses; “37% of firms with at least 250 employees reported using AI in their business operations,” and “Less than 20% of firms with four or fewer employees reported using AI” (Census Bureau, May 26, 2026). The survey question changed in November 2025, from AI use “in producing goods or services” to AI use “in any business function,” so figures from before and after that change are not directly comparable.

One measured productivity effect

In a study of 5,179 customer support agents, access to a generative AI assistant “increases productivity, as measured by issues resolved per hour, by 14% on average, including a 34% improvement for novice and low-skilled workers but with minimal impact on experienced and highly skilled workers.” The authors “provide suggestive evidence that the AI model disseminates the best practices of more able workers” (Brynjolfsson, Li and Raymond, NBER Working Paper 31161, April 2023, revised November 2023). The finding comes from one company’s support operation.

Documented error rates

Researchers at Stanford’s RegLab and Institute for Human-Centered AI tested AI legal research tools built to ground their answers in real sources. Lexis+ AI and Ask Practical Law AI “produced incorrect information more than 17% of the time,” and Westlaw’s AI-Assisted Research “hallucinated more than 34% of the time.” The study counts as an error a “misgrounded” answer, where “the AI tool describes the law correctly, but cites a source which does not in fact support its claims.” In the same authors’ earlier study, general-purpose chatbots “hallucinated between 58% and 82% of the time on legal queries” (Stanford HAI, May 23, 2024, updated May 30, 2024). These rates describe the tools and versions tested in 2024 on legal questions.

How NIST frames AI risk

NIST’s AI Risk Management Framework (AI RMF 1.0, 2023) states that “AI systems are inherently socio-technical in nature, meaning they are influenced by societal dynamics and human behavior.” It organizes risk management into four functions, “govern, map, measure, and manage,” lists “valid and reliable” among the characteristics of trustworthy AI, and states that “The Framework is intended to be voluntary, rights-preserving, non-sector-specific, and use-case agnostic” (NIST AI Resource Center, AI RMF 1.0).

How this connects to the rest of the system

Online Income covers the ways money is earned online and the federal data on how long new businesses survive. Investing (Long-Term Wealth) covers compounding, fees and trading behavior. Personal Capability & Everyday Skills covers the FTC’s fraud figures and its four signs of a scam. The rest of this pillar sits at the Systems, Wealth & Technology hub.

GoldBlock Workshop is educational. We explain how things work and what the evidence shows. This is information, not professional, legal, financial, or medical advice. Use your own judgment, and consult a qualified professional for your specific situation.

Sources and further reading

  • Lewis, Patrick, et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” (arXiv 2005.11401, 2020; abstract read 2026-10-19, entry 85). Pre-trained models’ “ability to access and precisely manipulate knowledge is still limited”; proposes “models which combine pre-trained parametric and non-parametric memory for language generation.”
  • U.S. Census Bureau, “Large Firms With at Least 20 Employees Biggest AI Users” (Grundy, Breaux and Khatiwoda, May 26, 2026; read 2026-10-19). AI use between 17 and 20 percent of businesses, December 2025 to May 2026; 37 percent of firms with 250 or more employees; under 20 percent of firms with four or fewer; question wording changed November 2025.
  • Brynjolfsson, Erik, Danielle Li and Lindsey R. Raymond, Generative AI at Work, NBER Working Paper 31161 (April 2023, revised November 2023; abstract read 2026-10-19). 5,179 agents; 14 percent average; 34 percent for novice and low-skilled workers; minimal impact for experienced and highly skilled workers.
  • Stanford HAI, “AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries” (Magesh, Surani, Dahl, Suzgun, Manning and Ho; May 23, 2024, updated May 30, 2024; read 2026-10-19). Rates for Lexis+ AI, Ask Practical Law AI and Westlaw AI-Assisted Research; definition of misgrounded; earlier chatbot rates.
  • National Institute of Standards and Technology, AI Risk Management Framework 1.0 (NIST AI Resource Center; read 2026-10-19). Socio-technical statement, four functions, “valid and reliable,” voluntary. Replaces the link to the nist.gov overview page, which could not be opened in this check.