Government organizations are under pressure to meet rising service demand while managing tight budgets, workforce shortages, and disconnected legacy systems.
Generative AI has helped employees find and produce information faster, but it has not necessarily changed the workflows behind government services. Agentic AI offers a potential next step: systems that can plan, reason and act across workflows under human oversight.
This was the focus of Public Sector Network’s webinar, Reimagining Government Workflows with Agentic AI, hosted in partnership with BMC Helix.
The discussion featured Antoine Elias, Director of Digital Architecture and Innovation and Lead of the AI Centre of Expertise at the Treasury Board of Canada Secretariat; Jean-Francois Goubet, Director of the Analytics and Program Insights Division at the Canada Revenue Agency; Stephanie Katz, Director of Contact Centre Service Design, Implementation and Centre of Expertise at Shared Services Canada; and Katie Tierney, Area Vice President at BMC Helix.
The central message was clear: success will not come from giving AI maximum independence. It will come from applying the right level of autonomy to clearly defined problems.
Governance Must Match the Level of Autonomy
As AI moves from providing information to completing tasks, the risk changes.
Stephanie Katz suggested thinking about AI across three levels: informing, recommending, and acting autonomously. Each carries a different risk profile and requires different controls.
“Stop asking what the technology is capable of,” Katz said. “What is it authorized to do, on whose authority, and what happens the moment something goes wrong?”
Before deploying an agent, organizations must determine what information it can access, which decisions it can support, and when a person must intervene. Actions should also be traceable and reversible, particularly when they affect citizens or sensitive information.
Start With Clear Operational Value
Contact centres provide a practical starting point. AI can surface relevant information during calls, draft summaries, suggest disposition codes, and reduce administrative work. Employees retain responsibility for decisions while receiving better support.
These tools can also reveal why citizens are contacting government, where they encounter difficulty and which problems begin earlier in the service journey.
The value is therefore not limited to handling calls faster. AI-generated insight can help organizations improve services and prevent recurring problems.
Redesign Workflows Before Automating Them
Agentic AI cannot simply be added to every existing process. Many government workflows rely on spreadsheets, manual checks, and institutional knowledge accumulated over several years.
Antoine Elias explained that organizations should map these workflows, divide them into smaller steps, and identify where AI can provide value.
“The goal is to give it appropriate autonomy,” Elias said. “The best AI is a verifiable AI.”
Complex workflows need checkpoints where outputs can be tested before moving to the next step. Human review should happen where judgment matters most, with visibility into the information the AI used and the assumptions it made.
Human oversight should not be a final approval button. It must be an informed review of a transparent process.
Connect AI to a Business Case
The Canada Revenue Agency has used machine learning and automation in risk assessment for more than a decade. Agentic AI could extend this work by helping employees analyze complex company networks, financial information, and other large datasets.
Tasks that once required weeks of manual analysis could be completed more efficiently, allowing employees to focus on interpretation and decision-making.
For Jean-Francois Goubet, however, adoption must be tied to measurable value.
“We are not looking at being fancy,” Goubet said. “It has to have a business case.”
A strong use case might reduce repetitive work, improve consistency, or help an employee manage more files without compromising oversight. The objective is to augment employees, not remove them from consequential decisions.
Bring the Right Teams Together
Scaling agentic AI requires early involvement from business, IT, cybersecurity, privacy, infrastructure, and data teams.
Security and infrastructure requirements should not be treated as obstacles added after a pilot. They are essential parts of designing a system that can operate responsibly.
Organizations also need reusable standards for security, privacy, accessibility, identity, auditability, and performance. Shared approaches can prevent departments from repeatedly solving the same foundational problems.
Start Small and Learn by Doing
For organizations beginning their agentic AI journey, the panel recommended three practical steps:
- Use approved AI tools and build hands-on experience.
- Identifya real employee or citizen pain point.
- Begin with a small, bounded use case that can be measured and scaled.
Agentic AI offers governments an opportunity to rethink how work moves across teams and systems. Its success, however, will depend less on how much autonomy the technology receives and more on whether organizations can apply it to meaningful problems, verify its work and preserve human accountability.
To watch the full virtual session, visit the Public Sector Network on-demand page.
Help your peers
Share what you've learned with fellow public servants