AI HR tools

Drafting. Job descriptions, policy drafts, performance templates, onboarding guides and manager scripts. The output is a starting point that a senior person edits in minutes instead of writing from a blank page, which is where a large share of HR hours actually go.

Answering the repeat questions. An assistant trained on your own handbook can answer holiday entitlement, expense rules and leave policy instantly, and hand the harder questions to a person with the context already gathered. The value is not the bot, it is the queue it shortens.

Reading volume. Summarising engagement survey comments, screening application volume, extracting terms from contracts, flagging anomalies in payroll before they run. These are tasks where nobody needs the model's opinion, only its throughput.

Compliance monitoring. Tracking expiry dates, missing acknowledgements, upcoming legislative changes in your jurisdictions, and telling you what needs attention this month rather than at audit time.

Any decision about an individual's employment: hiring, promotion, performance rating, discipline, termination, accommodation. A model can prepare the information, but the decision needs a named human who can explain it.

Protected grounds. Screening and ranking tools are exactly where bias claims get made, and in Canada the burden of showing a defensible process sits with the employer. If you use AI in selection, you need to be able to describe how it works and show that a human reviewed the outcome.

Confidential and personal information. Employee relations notes, medical accommodation details and investigation records do not belong in a general-purpose tool. Before anything is uploaded anywhere, you need to know where it is processed, how long it is kept, and whether it is used for training.

Anything you cannot audit. If the tool cannot show you what it did and why, you own the consequence without the evidence.

Give it your own material. Take three real documents, a real policy question and a real job description, and see what comes back. Vendor demos are curated; your handbook is not.

Ask where the data lives and who can see it. Data residency, retention period, subprocessors, whether prompts and outputs are used for model training, and how you get everything back out if you leave.

Ask what the human review step is. A serious tool shows its work and makes the approval explicit. A vague answer here is the answer.

Check the Canadian ground truth. Provincial employment standards, PIPEDA, human rights codes, and bilingual requirements in Quebec. AI trained on American content will confidently cite the wrong notice period.

Price the exit. What it takes to migrate off, whether the records export in a usable format, and how long the vendor keeps them afterwards.

An HRIS as the authoritative employee record, an ATS for hiring, a payroll platform that reads from the HRIS rather than being retyped into, and one AI layer on top for drafting and answering policy questions from your own documents.

The order matters more than the brands. The AI layer is only as good as the documents it can read, so the first job is almost always consolidating policies into one maintained source instead of six PDFs written in different years.

Where a client already owns three of the four, we keep them and fix the integration. Replacing a working payroll system to gain an AI feature is almost never the right trade.

A short AI use policy: which tools are approved, what may never be entered into them, and who approves a new one. One page is enough, and its absence is what turns a helpful tool into an incident.

A human in the loop, in writing, for every process that affects an employee's rights. Name the role responsible, not the department.

Disclosure where it applies. Candidates and employees should know when AI is being used in a process that concerns them.

A review date. Models and vendors change quarterly. Whatever you approved in January should be re-checked before it quietly becomes something else.