Technology · September 26, 2026 · 8 min read
Where AI genuinely helps HR in a small company, and where it quietly creates risk
HR is one of the few functions where AI can produce a document that looks finished while being unfit for the jurisdiction you operate in. Here is the honest split between what to automate and what to keep human.
The case for AI in a small HR function
A company of sixty people has roughly the same policy surface area as a company of six hundred. The difference is that at sixty, the HR work is done by a founder, an office manager, or a part-time generalist who also runs payroll.
That is the real opportunity. AI is least interesting for the decisions and most interesting for the drafting, summarising and monitoring that surround them: policy drafts, job descriptions, meeting notes into action items, question answering against your own handbook.
Used that way it compresses administrative load by a meaningful margin. Used as a decision maker it produces confident, well-formatted answers that may be wrong for your province, your contracts, or your own precedent.
Safe territory: drafting and summarising
First drafts of job descriptions, interview guides, policy outlines, onboarding checklists and internal announcements. All of these benefit from a structured prompt and a review pass, and none of them create employee entitlements on their own.
Summarising is the other reliable win: turning a long employee survey into themes, turning meeting notes into decisions with owners, turning an exit interview into a pattern. The output is only useful if a human decides what to do with it, which is exactly the right division of labour.
- ·Drafting: job descriptions, interview guides, policy outlines, internal comms
- ·Summarising: engagement surveys, exit interviews, manager one-on-ones, documentation
- ·Answering: internal questions grounded in your own handbook and policies
- ·Scheduling and coordination: interview logistics, review cycle reminders
The high-risk zone
Anything that affects an individual's employment outcome. Performance ratings, termination reasoning, promotion decisions, compensation recommendations, accommodation assessments and candidate screening all carry human rights exposure.
The specific danger is not that AI is biased in the abstract. It is that these decisions need a documented, defensible rationale tied to the person's actual role and record, and a generated rationale is not a record of anything that happened.
Employment law also differs by jurisdiction, and generated policy text regularly blends expectations from different provinces and states. A confident paragraph is not a compliant one.
The data rules to set before anyone uses a tool
Decide what may never be entered into a public model: health information, accommodation details, disciplinary records, salary data by individual, and anything from a personnel file. Write it down and tell people once, clearly.
Decide where the data goes. Enterprise tools with data processing terms are a different risk category from consumer chat interfaces. If you operate in a regulated industry or handle EU or UK data, this is not a detail.
Decide who reviews. A named person, not a committee, approves anything that becomes a policy or a decision record.
What an AI policy should actually say
Most AI policies are aspirational and unusable. A working one covers four things: which tools are approved, what data is prohibited, which decisions require human sign-off, and how employees should disclose AI use in work product that affects people.
It should also be short. A two-page policy that employees read beats a twelve-page framework that nobody opens.
- ·Approved tools, by name, with the business account that must be used
- ·Prohibited inputs, with examples rather than principles
- ·Decisions that require human review and a named reviewer
- ·Disclosure expectations for candidate and employee facing output
- ·A review date, because the tool landscape moves faster than annual cycles
The configuration most small companies should aim for
An AI assisted HR stack has three layers. A payroll and HRIS system that holds accurate records. A knowledge layer, usually your handbook and policies, that questions can be answered against. And a senior human, fractional or full-time, who owns the decisions and reviews anything that touches an individual.
The mistake is buying the middle layer and assuming it replaces the third. It does not. It makes the senior person faster, which is precisely why the arrangement is worth the spend at fifty employees rather than five hundred.
Questions about this
Can AI write our employee handbook?
It can produce a plausible draft, and a plausible draft is not a compliant document. Use it to structure sections, then have someone who knows your jurisdiction, headcount and contracts review every clause before it is issued.
Is it legal to use AI for candidate screening in Canada?
It is legal but regulated, and increasingly requires disclosure and accommodation consideration. Most small companies are better served by structured human screening against written criteria than by an automated filter they cannot explain.
What employee data should never go into an AI tool?
Health and accommodation information, disciplinary records, individual compensation, and any identifying details tied to a performance or termination decision. Treat personnel files as off limits unless the tool is approved for that data specifically.
Will AI reduce the need for an HR hire?
It reduces the administrative share of the role, which is often the reason a part-time generalist was needed. It does not remove the need for judgement on performance, restructuring and compliance, and those are the expensive parts.