Published evidence
Case Studies From the Field
Documented examples of organizations changing how work gets done, with sources you can inspect.
These are external research examples, not AgentC Foundry client engagements, testimonials, or endorsements. Reported findings belong to the organizations and researchers cited below. Our interpretations are labeled separately; results are not guarantees for another business.
Nonprofit operations
Degrees of Change: connecting applications and placement decisions
Degrees of Change faced manual application reviews and disconnected program data. Microsoft's case study describes connected applications and portals for applicants, volunteer assessors, employers, and staff. AI Builder extracted information from resumes and proposed internship matches for staff review. The organization reported being able to process placements faster and at greater volume.
What the evidence supports: A connected application and decision-support workflow. This is a vendor-published customer account, not an independent impact evaluation. It does not establish audited financial savings or measured bias reduction.
AgentC interpretation: Connect intake, review, and placement decisions around shared information, while keeping staff responsible for the decisions.
Read Microsoft's Degrees of Change case study. Source last updated April 8, 2026; reviewed September 27, 2026.
Customer support
AI assistance inside an existing support workflow
At an unnamed enterprise-software company, support staff diagnosed technical problems during live customer chats. Researchers studied the introduction of an AI assistant that suggested responses and relevant documentation. Employees remained responsible for the conversation and could ignore its recommendations. The April 2023 working paper reported higher issue-resolution productivity, with larger gains among newer and lower-performing staff.
What the evidence supports: Benefits varied by worker experience within one company's rollout. This working-paper version was not peer-reviewed and disclosed potentially relevant author relationships. It is not proof that every support operation will benefit equally.
AgentC interpretation: Put assistance inside the support workflow, preserve employee judgment, and measure resolved problems rather than generated replies.
Read Generative AI at Work (PDF, April 2023 version), by Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, NBER Working Paper 31161. Reviewed September 27, 2026.
Governance and measurement
HMRC: evaluating the work, not just access to AI
HMRC tested Microsoft Copilot for everyday Office tasks from September to December 2024. Its evaluation combined surveys, usage data, focus groups, and a task-based exercise. Participants reported time savings and benefits from meeting summaries and document searches. HMRC adjusted savings estimates for non-use and survey-response bias. During the trial, use was restricted to material below the Official Sensitive classification.
What the evidence supports: A department's evaluation of its own deployment, not an independent external audit. Savings relied substantially on self-reporting and were presented as estimates rather than precise measurements of impact.
AgentC interpretation: Set information-handling boundaries before rollout. Measure adoption, work quality, and realized benefits separately; buying access does not demonstrate value.
Read HMRC's Copilot evaluation. Published July 9, 2026; trial conducted in 2024. Reviewed September 27, 2026.