The Canada Revenue Agency has run four generative-AI pilots in human resources since January 2025, testing tools for workforce adjustment, labour relations, performance management, and work-description writing. The experiments cost a combined $4,900, but unions say they raise serious questions about how AI could influence decisions that affect public servants' careers.
"These are not low-stakes administrative functions. Workforce adjustment, performance management and labour relations can have very real consequences for public servants and their careers," said Sean O'Reilly, president of the Professional Institute of the Public Service of Canada (PIPSC).
The CRA, which employed about 48,774 people as of March 2026, said the tests used sample, hypothetical, anonymized, or otherwise non-live scenarios and did not support or inform actual HR actions involving real employees.
What the CRA tested
The agency tested GENNI, an internal generative-AI system built on a Microsoft-based platform, with about 20 labour-relations advisers from October 2025 to March 31, 2026. Labour-relations advisers handle workplace issues between employees and management, including interpreting collective agreements. GENNI searched and retrieved approved reference material such as policies, directives, procedures, and collective agreements.
CRA spokesperson Charles Drouin said the goal was to improve access to consistent and up-to-date information. But after testing, Labour Relations decided the tool did not sufficiently meet operational requirements and chose not to proceed with implementation.
A second GENNI pilot for performance-management advisers began in early 2026 and remains in the experimentation stage. Seven advisers are testing the tool for research and retrieval from approved reference materials. The CRA said it is evaluating accuracy, reliability, and user experience, and has not reached conclusions about efficiencies.
For workforce adjustment, the CRA tested NotebookLM with 26 advisers starting in August 2025. Workforce adjustment is the process used when employees' services may no longer be required because of changes such as a lack of work or the discontinuation of a function. In March, the CRA launched a new workforce-adjustment exercise affecting 479 employees across four branches, including 284 Union of Taxation Employees (UTE) members.
The CRA said the NotebookLM pilot was not used in operational workforce-adjustment processes, Comprehensive Expenditure Review activities, or active staffing-reduction cases. The documents the bot drew on included collective agreements covering Public Service Alliance of Canada and PIPSC employees, internal training material, FAQs, and intranet content. Advisers tested it on questions about workforce-adjustment rules, including how long employees whose positions are eliminated receive priority for other federal jobs, and questions about pay and timelines.
The CRA said the workforce-adjustment pilot is complete, but did not provide an outcome or say whether it plans further testing.
The work-description writing pilot began in January 2025 and involved six employees using a secured business version of ChatGPT. Work descriptions set out the duties attached to a position and are used in federal job classification. The CRA said ChatGPT helped draft descriptions aligned with classification standards, but did not determine or recommend the classification, group, or level of a position. Those decisions remained with qualified employees.
"The work description writing pilot demonstrated efficiencies in the drafting process and contributed to improvements in the quality and clarity of work descriptions," Drouin said.
The CRA said GENNI did not generate recommendations, assessments, or conclusions about individual employees in the labour-relations and performance-management pilots, and that AI-generated content remained subject to human review. The labour-relations GENNI pilot cost $1,950, the performance-management GENNI pilot cost $2,850, the ChatGPT work-description pilot cost $100, and the NotebookLM pilot had no reported licensing cost.
Unions raise concerns about consultation and privacy
UTE national president Adam Jackson said he was aware of GENNI generally, but the NotebookLM and ChatGPT pilots were the first he had heard of them. The CRA said bargaining agents were not specifically consulted before the GENNI pilot.
"The pilot was limited to testing a research and information-retrieval tool used by HR professionals and did not involve employee-specific information, HR decision-making, or changes to employees' terms and conditions of employment," Drouin said. "As a result, bargaining agents were not specifically consulted prior to the pilot."
O'Reilly said PIPSC's Audit, Financial, and Scientific group was not aware of the pilots and was not consulted before they began. He said previous discussions between the union and the CRA about AI focused on employees using the technology in their own work, not the employer testing it within HR and labour-relations processes.
The CRA said the initiatives were limited internal tests that did not involve employee-specific information, HR decision-making, or changes to terms and conditions of employment, and therefore did not require union consultation.
"At a time when CRA employees are already facing uncertainty about workforce reductions, learning that generative AI has been piloted in a workforce-adjustment context without prior consultation is particularly concerning," O'Reilly said.
Jackson raised privacy concerns. "This is performance management, this is HR, this is WFA. These are life-changing events for our members, and to have a machine aid in that is not only cold but, I would argue, risky because of the risk of privacy leaks."
Joanna Redden, a professor at Western University who studies data and AI in the public sector, said describing the tools as research and information-retrieval aids does not eliminate concerns. Generative-AI systems can produce incorrect information, she said, which creates additional work for employees responsible for checking outputs. She also said the CRA's workforce reductions make questions about oversight and organizational capacity particularly relevant.
"There is a need, particularly when we are talking about essential public services, for there to be transparency, accountability, reliability, and care," Redden said. She said assessments of public-sector AI should examine effects on working practices, discrimination, institutional capacity, and rights, and that independent reviews should be made public.
Why this matters for HR professionals
For HR teams watching this story, the CRA's experience offers a concrete lesson: piloting AI in HR functions without union consultation creates friction even when the tools are limited to research and retrieval. The agency's own outcome - one tool rejected, one still in testing, and no clear answers on the others - shows that cost is not the main risk. The bigger risks are trust, transparency, and the perception that machines are touching decisions that shape people's careers.
HR professionals considering similar tools should document what the AI can and cannot do, clarify who reviews its output, and engage with employee representatives before testing begins. The AI for Human Resources coverage and the AI Learning Path for HR Managers offer practical grounding for teams weighing these questions.
Your membership also unlocks:
