PSC’s travel-approval process was slowing down both employees and finance experts. Finance Officers were acting like a manual search engine for travel policies, constantly fielding routine questions. This led to:

• Delays in getting answers to complex travel policy questions sometimes days or even weeks.
• More last-minute bookings, which increased travel costs.
• Frequent interruptions to high-value finance work, with no budget to add more staff.

Using Microsoft 365 Copilot Chat and Copilot Studio, PSC built an in-house AI agent that:

• Pulls guidance from SharePoint, internal policies, and SOPs.
• Provides instant, bilingual (English and French) answers to travel policy questions, 24/7.
• Acts as a single, authoritative source of truth, improving the accuracy of travel claims before they reach finance.

The result is a reimagined travel-approval experience: employees self-serve clear, compliant answers in seconds, while finance teams stay focused on strategic, value-added work.

PSC tracked clear, quantifiable outcomes from its AI travel agent rollout:

• 90% reduction in emails to the travel mailbox in fiscal year 2025–26.
• Complex inquiry response times dropped from 1–5 days to about 10–30 seconds.
• Earlier approvals and bookings now save hundreds to thousands of dollars per trip, as employees can book well in advance instead of at the last minute.
• The department avoided the cost of an additional full-time role, estimated at about CAD 120,000 per year in salary cost avoidance.

Beyond the numbers, PSC reports:

• Higher accuracy of travel claims thanks to a single, consistent interpretation of policy.
• Improved morale, as finance experts spend more time on advisory work and high-risk files instead of repetitive lookups.

All of this was achieved using existing Microsoft licenses, without external consultants, and embedded directly into tools staff already use every day.

PSC treated the AI agent as a governed, repeatable capability rather than a one-off experiment. They focused on three key areas:

1. Accuracy and human oversight

• The team ran a three-month pilot inside accounting operations, where financial officers and specialists validated the AI’s answers.
• Subject matter experts shifted from doing the research themselves to reviewing and refining AI-generated responses.
• Live links to the PSC knowledge base keep guidance dynamically up to date as policies change.

2. Security and guardrails

• The agent operates inside a defined “digital fence,” with clear visibility into data residency and security.
• Behavioral boundaries are enforced so the agent understands the limits of its advisory role.
• For complex or out-of-scope questions, the agent routes users to human experts instead of guessing, reducing hallucination risk.

3. A modular playbook for scaling

To extend this approach to other areas (such as HR and procurement), PSC created a three-pillar playbook:

1. Identify the authoritative knowledge layer – digitize and structure policies so the agent always draws from a single source of truth.
2. Apply domain guardrails – define what the agent can and cannot advise on in that domain.
3. Close the loop with human oversight – involve subject matter experts in validation before broad rollout.

This model has attracted interest from 40+ departments and agencies, positioning PSC’s solution as a practical template for safely scaling agentic AI across government.