Canadian public sector HR advisers need to learn how an AI agent reaches a verdict, how to review its drafts in both official languages, and how to correct it so errors stop recurring. Staffing, accommodation and discipline decisions stay with people. Public-sector privacy law and federal policy on automated decisions, not PIPEDA alone, set the frame.
How do HR advisers in Canadian government spend their time?
A typical adviser in a federal department or a provincial ministry serves managers more than employees. They explain collective agreement provisions on overtime, acting pay and leave, advise on staffing options and the merit criteria behind them, support labour relations files, and handle requests for accommodation. Many questions arrive twice, in English and French, and the answer must say the same thing in both.
Pay and classification generate their own queue. An employee disputes an acting assignment, a manager wants to know whether a position can be reclassified, a union representative asks for the basis of a decision. Each answer has to cite the agreement or the policy, and each is a record the department may later need to produce.
What changes when agents arrive?
Agents shorten the search and the first draft. Each of these runs at a depth earned on the department's own past files:
- Collective agreement lookups (depth 1). The agent finds the governing clause across the relevant agreement and any interpretation bulletins, and quotes it with the source.
- Bilingual reply drafting (depth 2). It drafts the manager's answer in both official languages. An adviser checks that the two versions match and sends.
- Staffing file preparation (depth 2). From the statement of merit criteria and assessment notes, the agent drafts the rationale for a staffing decision. The hiring manager signs.
- Accommodation request intake (depth 1). It organises what the employee has submitted and lists what is missing, without assessing the medical content.
Anything that would alter pay, classification or an appointment is outside the agent's role scope and is refused or held for a named adviser.
What must advisers learn, and what stays with them?
Before reviewing anything, advisers should understand the decision chain. Each action meets five bounds at once, and the lowest one wins: the agent's role scope, its certificate, its sources, the department's limits and the platform guardrails. A verdict names the bound that decided. Our guide to what HR advisers should learn before working with AI agents treats this as the first lesson.
Next comes reviewing a bilingual draft. Advisers should be trained to check the French and English versions against each other and against the source clause, because an agent can be right in one language and loose in the other. The review routine for HR advisers adapts well to this two-version check.
Third, consistency across advisers. Two advisers who approve the same kind of draft differently will teach the agent nothing useful. Regular calibration sessions for HR advisers, where the team reviews the same anonymised cases and compares, keep the standard shared.
What stays with the adviser: the duty to accommodate, which turns on individual facts and human rights obligations; advice on discipline and grievances; any judgement about merit; and every conversation where an employee is upset or at risk. These need a person who can be questioned and who answers for the outcome.
Which Canadian rules shape oversight and literacy?
Canada has no federal AI act in force. PIPEDA is the federal private-sector privacy law; government institutions are covered by separate public-sector privacy legislation, federal or provincial, which limits collection and use of employees' personal information and gives them rights of access. Advisers must know which law governs their employer, because that determines what the agent may read.
Federal policy on automated decision-making expects departments to assess the impact of a system before using it for administrative decisions and to keep a person involved in proportion to that impact. Public bodies also carry transparency and fairness duties that private employers do not, so a staffing rationale drafted by an agent must still be one a manager can explain and defend.
OSFI and the other financial regulators do not reach this work. Voluntary standards such as ISO/IEC 42001 and the NIST AI Risk Management Framework help structure roles and competence for departments that want a reference.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
A 90-day training plan for an HR team
Weeks 1 to 4. The whole HR team learns the basics: depth, the minimum chain, the four verdicts. In Pryme Intelligence Academy this is the free Foundations track, about four hours. The team chooses collective agreement lookups at depth 1 as the first use, because the output can be checked line by line.
Weeks 5 to 8. The advisers who will review take operator training and run the agent alongside their own work on live manager queries. They record every disagreement, the bound named in each held verdict, and the corrections they bank against sources.
Weeks 9 to 13. The team holds its first calibration session, settles the bilingual review rule, and decides whether bilingual reply drafting at depth 2 is ready to be proven on the department's own cases. The HR director signs a list of who may approve held drafts and confirms what the impact assessment says about human involvement.
Questions people ask
Does PIPEDA apply to a federal department's HR files?
Generally no. PIPEDA covers private-sector organisations; government institutions fall under public-sector privacy law. Advisers should confirm which statute applies to their employer.
Can the agent answer employees directly?
It can at depth 3 for routine, low-risk information if the department allows it, such as where to find a form. Answers that interpret an entitlement are better held for an adviser.
Who should own the agent inside HR?
A named senior person who signs off the bounds and the approver list. Our piece on why HR advisers are the real control explains why review sits with advisers, and the guide to AI agent skills by role covers the owner's skills.