Government AI Governance & Research
GIAG, the Government IT/AI Governance Initiative, is independent research on whether government AI controls work. It builds instruments that measure them, using public-record case analysis and practitioner conversations, free of vendor marketing.
In 2012 GAO found that U.S. Customs and Border Protection updated the cargo-targeting rules in its Automated Targeting System without first assessing whether the update was more effective.1 In June 2026 the Treasury Inspector General for Tax Administration found that the IRS, which began piloting chat services in 2017, had not taken sufficient steps to evaluate their effectiveness.2 Both agencies agreed with the recommendations. GIAG builds the instruments that supply the missing evaluation evidence.
Phase 2 is recruiting. Phase 1 tested the instruments on public cases. Phase 2 interviews with government CIOs, chief AI officers, procurement and audit staff are open now. Target: 15 conversations completed by February 1, 2027.
Cases discussedCBP Automated Targeting System, GAO 20121 · IRS chat applications, TIGTA 20262 · SBA AI implementation, SBA OIG 20263 · Army contract award litigation, 2026 (allegations)4 · All sources
Why It Matters
Federal mandates, OMB guidance, and NIST frameworks have established the what of AI governance. ThinkCapital's research program is focused on the harder questions: the how, the how much, and the how do we know. We work at the intersection of measurement science, organizational behavior, and government IT, producing frameworks practitioners can use.
The three audit cases share one pattern: a requirement or approval on paper, and an evaluation or implementation step that had not been taken when the review was done. GIAG builds the instruments that supply the confirmation.
Five questions that separate governance from paperwork:
The GIAG Governance Pipeline
A failure at one stage does not require a failure at another, so GIAG measures each stage separately. Four stages have instruments today. Two are open research.
Does the system still do only what it was authorized to do, and who would notice if it did not?
Forming researchStream Three: Authorization & Scope Integrity →
Could the testing claim have failed?
InstrumentTesting Rigor: four conditions
Solicitation & Due-Diligence Review →
Did review change or stop anything, and did governance scale up when the pilot moved to production?
Forming researchStream Four: Deployment Review →
Did the human reviewer have the authority, time, and information to exercise judgment?
InstrumentHuman Oversight Quality Index
How long did an unauthorized action run before someone with authority knew?
InstrumentDetection Clock (Working Paper 4)
Can a single decision survive scrutiny after the fact?
InstrumentGovernance Evidence test (Working Paper 5)
Government AI in Practice
The newsletter applies GIAG's instruments to current federal and state events and ends each issue with a diagnostic you can run on a system in your portfolio.
Pick one AI system in production. Find the document that first defined its authorized purpose, then list what the system does today. Does anyone own the comparison between the two, and when did it last run?
ThinkCapital Research Initiatives
Two streams are active and four are forming. Two of the forming streams cover the pipeline stages that have no instrument yet: authorization and scope, and the pilot-to-production gate. Visit the research page to read full descriptions, track progress, or take part in an interview or benchmark. Phase 2 target: 15 practitioner conversations completed by February 1, 2027.
Examining how the NIST AI Risk Management Framework is being adopted, or not, across federal and state agencies, and which factors practitioners associate with successful operationalization.
What does meaningful human oversight of AI look like in operational government settings? The Human Oversight Quality Index is published; this stream collects practitioner data to refine it and its maturity criteria.
When an agency authorizes a system for one purpose, what shows it stayed within that purpose? We examine how scope grows after approval, who owns the comparison between authorized and actual use, and what happens when the authorizing role sits vacant.
Where does pre-deployment review change or stop a system, and where do pilots reach production with governance sized for a pilot? We plan to test the Five Intervention Criteria and risk tiering against agency cases.
Applying Schelling-Granovetter threshold models to explain why identical AI initiatives succeed in some agencies and stall in others, and what moves them past the tipping point.
Developing empirical methods for measuring whether AI tools are delivering measurable productivity improvements in government mission contexts, beyond technology metrics alone.
Advisory Services
The anchor engagement is a six-week AI Governance Assessment of an agency's policies, oversight structures, risk management practices, and measurement capabilities, benchmarked against the NIST AI RMF. It delivers a written maturity profile, a prioritized improvement roadmap, and an executive briefing. The work uses the same instruments and measurement discipline as the research. Advisory work is separate from the research program, and no advisory client funds or reviews research findings. See Research Integrity.
"Agencies that cannot measure their controls cannot show those controls work."
ThinkCapital Research Perspective
Get Involved
Whether you want to follow our research, take part in an interview or benchmark, or ask a question, the Engage page is where conversations begin.
Go to the Engage Page →