The Race — Building Democracy in the Age of AI. People race toward a shared finish line.
The Race
← Trends

Established trend · Active

The ability to implement AI remains uneven across firms and workers

Latest assessment: · Evidence updated

Standing score
Unscored
Direction of power
Undetermined
Window
Unclear
Confidence
Moderate

What we think is happening

Smaller firms and workers with fewer training opportunities report shallower access to usable AI than larger or better-supported peers. The likely power mechanism is complementary resources: staff time, skills, suitable workflows and error supervision determine who turns tool access into capability.

Scope: Different selected surveys and qualitative baselines; findings not pooled into a single adoption rate.

Potential direction of power

Repeated size and learning gaps show that access to AI does not give every organization the same effective capability. The studies do not establish a lasting causal transfer of market share or wages, and some organizations choose not to adopt. Adoption percentages alone cannot resolve the power question.

Supporting cases

Maryland Business AI Benchmark

Maryland published an official summary of a June–July survey of nearly 300 business leaders on October 1. It reports 91% using AI,58% basic use and 92% of regular users reporting gains, mostly slight. Sixty-four percent expect existing staff to do more;4% expect headcount reduction.

The full report was blocked, leaving sampling and weighting unknown. These are attributed respondent findings and expectations, not measured productivity or job loss.

Read the case and sources →

UnidosUS affiliates organize AI adoption around staff needs and shared learning

Johanna Lara and Claudia Ruiz’s October 5 institutional account describes a June 24 convening of about 100 affiliate staff. UnidosUS’s network comprises roughly 300 community organizations; that is not the number trained at the event.

Affiliates described using AI for notes, intake summaries and reports. At the Latino Economic Development Center, staff began with a list of problems they wanted solved, then built an adoption roadmap rather than selecting a product first. Concerns included privacy, cybersecurity and job security; skeptical staff participated in shaping rollout. The account is participant/organizer evidence. It does not provide a controlled productivity study, verified hours saved across the network, a disclosed programme budget or proof that staff gained formal bargaining rights.

Shared implementation knowledge and community-defined problems can help smaller organizations avoid adopting whatever a supplier happens to offer. Staff participation can improve retained capacity, but differs from an enforceable workplace AI clause.

Read the case and sources →

A Dallas experiment finds AI increases self-filed property-tax appeals—and may widen participation disparities

Justin Holz, Ricardo Perez-Truglia, Andrew Simon and Alejandro Zentner’s August 2026 working paper, NBER 35632, reports a randomized field experiment involving 645 Dallas County households. Both groups received a website with personalized appeal information and supporting evidence; half also received an AI chatbot. 78% of those offered it initiated a conversation. Filing an appeal without a human agent rose from 41.4% to 50.5%, a 9.1-percentage-point increase.

The authors report smaller gains among less-advantaged households and call the evidence of wider disparities suggestive. The authors’ abstract, also indexed at SSRN ↗, supports those findings. The available abstract does not establish recruitment details, subgroup uncertainty or appeal outcomes. These figures establish increased self-filing within this experiment, not successful appeals, reduced tax liabilities, county revenue losses or a population-wide effect.

This is directly about AI-mediated access to a public fiscal/legal process. It illustrates how equally available tools can produce unequal effective capability. Its relevance does not depend on claiming AI has already eroded the property-tax base.

Other fiscal research examines possible tax architecture under AI-driven economic change. This study is a different kind of evidence: measured behavior inside an existing tax-administration process. It is not evidence for the shift toward capturing AI-generated surplus.

Read the case and sources →

Workday report describes a shift toward applied implementation skills

Workday’s October 5 report analyzes recruiting activity from over 550 customer employers and a separate September survey of 6,000 workers. Basic AI-skill mentions fell 25% after a January peak while applied automation/engineering demand rose 51% from September 2025 to July 2026. Many respondents know which skills they need, but fewer say employers help them obtain those skills. Recruiting-system data describe participating employers, and worker answers are self-reports; neither represents a measured economy-wide replacement rate. Primary report ↗.

The evidence separates access and training from effective implementation. Participating-employer data and business-contact anecdotes can identify uneven capacity, but do not establish causal job replacement across the economy.

Read the case and sources →

St Louis Fed business contacts describe limited AI use and implementation barriers

Separately, a St Louis Fed article published July 16 ↗, resurfaced in October distribution, reports May business-contact survey results. Thirty-four percent regularly used AI in a small portion of operations, with 38% of nonusers citing skills/data/technology barriers and 34% seeing no use for AI. Examples include claimed time savings and growth without hiring; these anecdotes are not causal estimates. Almost half expected no headcount change. July publication and May evidence must not be relabeled an October study.

The evidence separates access and training from effective implementation. Participating-employer data and business-contact anecdotes can identify uneven capacity, but do not establish causal job replacement across the economy.

Read the case and sources →

Kansas State creates an institutional process for reviewing AI tools

Kansas State University announced October 7 that it had established an AI stewardship committee and adopted an interim framework for reviewing AI uses. Faculty and staff can submit tools through an intake process, allowing the university to assess security, privacy, records and other institutional implications.

The framework follows an April report. Some access is already available, including limited Claude Teams use without institutional data; fuller policies and a registry remain work for 2026–27. The announcement does not establish a completed university-wide AI governance system.

This is more than purchasing access to a chatbot. It begins to retain the knowledge needed to decide which uses are acceptable and how existing institutional responsibilities apply.

Distinct connection: Retained institutional implementation/intake knowledge and practical public capacity. Overlapping evidence is not independent confirmation.

Read the case and sources →

Window of contestability

Unclear for smaller firms and less-supported workers: access differences do not yet establish durable exclusion, and some nonadoption is a rational choice.

What would change this assessment

Independent long-term outcomes, accessible learning opportunities and retained skills in smaller institutions would help distinguish durable disadvantage from a temporary adoption gap.

Related cases

← All 24 trends