AI for HR: Where It Helps, and Where the Law Draws a Line

AI is useful in HR when the job is to find a rule and cite it. It's risky when the job is to judge a person.

Here's the useful side. A manager asks when final pay is due in her province. The tool tells her. It quotes the statute and your stricter policy side by side. And it adds a note that this is information, not legal advice.

It should score its own answer. A weak one goes to a person on the team, with the whole chat. And the tool should never see an employee's name it doesn't need.

Now the risky side. Point AI at hiring, promotion or firing. You've moved from a tool that saves time to one the law regulates. Audits and notices come with it.

What can AI do for an HR team?

In order of how much time it gives back:

1. Answer the questions that keep coming back. Leave, final pay, notice periods and overtime rules. Or which rule applies to a part-timer in another state. Each answer is cited, with a disclaimer. And it's tied to the right place.
2. Draft from approved templates. An exit checklist, a leave letter or a new-hire summary. It drafts from your own templates. Then a person reviews and signs.
3. Sort employee questions. It sorts what comes in by topic and how urgent it is. It answers what's routine and sends the rest to the right person.
4. Show what's being asked. You see which questions repeat, and which ones it couldn't answer. You see where your rules have gaps.

Now draw a line. On one side are those four. On the other is any tool that acts on a specific person. It scores, ranks or recommends what to do. The first group helps your team answer. The second group decides. The law treats it differently.

Can AI answer employment law questions?

It can answer questions on employment standards from the standards themselves. That's a narrower and safer thing. Employment standards are the base rules each state or province sets. They cover pay, hours, leave and notice. The design has three parts.

Route by jurisdiction. A jurisdiction is the state or province whose law applies. Standards differ from place to place. So the same question has a different answer in each. The tool should find the place first. Then it searches only that place's standards. It adds your own policy. A tidy answer for the wrong province is worse than no answer.

Cite the law and the policy side by side. Your policy can be stricter than the standard. Say a manager asks about vacation pay when someone leaves. She needs both, each labeled, each with the passage it rests on. (The general method is in our guide to RAG with citations.)

Add the disclaimer in code. Each answer says it is information, not legal advice. That note is a fixed block. The model doesn't have to remember it.

We built an internal assistant this way. It serves an HR team of 2 to 5 people. They cover several provinces in Canada. It routes each question to that province's standards and the firm's own rules. It answers with citations and a legal disclaimer. And it is deployed. Admins tag each file they upload by province and topic. That's what makes the routing work.

How does an HR assistant know when not to answer?

It scores each answer before it goes out. The tool above uses three parts:

Take a two-part question from a site lead. When does an employee's final pay have to go out? And does vacation pay go with it? Here is what the weights do to three possible answers. The inputs are made up, scored from 0 to 1:

The last row should worry anyone buying an HR tool. It's well written and sure of itself. But it rests on passages that don't quite fit. That's the trap. With 80% of the weight on retrieval, it can't talk its way up. The most the other two parts can ever add is 0.20.

Then there's the line itself. Below it, the tool sends no answer. Instead, one click makes a ticket. A senior person sees the full chat. We call it The Ask-A-Person Line.

Where the line sits depends on your questions. So you set it with a test set. That's 15 to 25 real questions. Each one comes with the sources a right answer has to cite. It's just what the tool above ships with.

Do wrong answers keep scoring above the line? Then the line is too low. Do right answers keep landing below it? Then your senior people have become a help desk for the tool. Move the line until neither happens.

Run the set again when a province or state changes a rule. Do the same when you change a policy. (The wider pattern is in human in the loop AI.)

Is it safe to put employee data into an HR assistant?

It can be, if the data never gets in. Most HR questions don't need the employee's name. You can answer them without it. So the safest design strips names and private details first. That happens before a word is indexed, logged or sent to a model.

The tool above strips them on the way in. Its rules on keeping data are written to PIPEDA and GDPR. That's a design standard, not a certification. Access is by role. A junior staffer and an admin see different things.

Know which privacy law you're under. In Canada, PIPEDA is the federal privacy law for private firms. It applies when they handle personal information in commercial activity. It also covers staff data at federally regulated firms. In the EU and the UK, the data law called GDPR covers employee data. In the US, it depends on the state.

Ask any vendor one plain question. If someone types a name, where does the name go? Then ask to see the logs.

What does an HR assistant cost, and what is it worth?

Under the tool above sit the model, the search index and hosting. There's also a link to your ticket system. That's estimated at $115 to $465 a month. Caching trims the model part of the bill. It means saving answers to repeat questions and reusing them.

For a small HR firm, we put an assistant like this at $60K to $140K a year. That's our estimate. It's a modeled range from our build library, not a client's result. The core of it is senior time. Our estimate assumes it takes 60 to 80% of repeat questions off senior staff.

The math is simple. Say a senior consultant bills $150 to $250 an hour. She spends just 5 hours a week on routine questions. That's $39K to $65K a year of billable time. (It's 5 hours x 52 weeks x the rate.) The rest is faster turnaround for clients.

Run an in-house HR team? Count the same hours at what your senior people cost.

What goes wrong

• One province's answer to a question that spans two. The tool has to know which province or state it's answering for. If it can't tell, it should ask.
• Policy missing from the answer. Say the firm's rule is stricter than the law. The tool cites only the law. The answer is right on paper and wrong in practice.
• Stale sources. A province changes a rule, and the index doesn't. The test set catches it, if someone runs it again.
• Names in the logs. Say questions are logged while the names are still in. Then you have a log full of names.
• No one under the line. An escalation needs a person and a deadline. If not, "ask a person" just moves the guess somewhere quieter.
• Treated as a lawyer. Firings and disputes go to a lawyer. So does any case heading to a tribunal or a court. That's true with or without AI.

What to ask a vendor

1. How does the tool decide which province or state a question is about?
2. Does every answer cite both the standard and our policy, with the passage?
3. How is the confidence score worked out? Can we see it on every answer?
4. Who writes the test set? (It should be you, from your own inbox. A vendor writing it is marking its own homework.)
5. Where do names go if someone types one?
6. What's kept, for how long, and on whose accounts?
7. What happens to a question below the line, and who gets it?

When not to build an HR assistant

You work in one state or province, with a few dozen staff. A shared doc of answers, kept current, does the job. It costs an afternoon.

You get fewer than about ten policy questions a week. The build won't pay back.

You have no written policies. The tool can only cite what exists. Write the policies first.

You want legal advice. A tool like this cites standards and policy. Each answer carries a disclaimer. It isn't counsel. It shouldn't be sold as counsel.

Where is AI in HR regulated?

Mostly where it decides about people. Two examples show the shape.

New York City, Local Law 144. It covers what it calls automated employment decision tools. That's software that helps decide who gets hired or promoted. An employer or employment agency may not use one unless three things are true:

• the tool has had a bias audit within one year before use
• a summary of the results is public
• candidates or employees get notice, 10 business days ahead

The city began to enforce it on July 5, 2023.

The EU AI Act. This is the EU's law on AI. Its Annex III is a list of high-risk uses, which face extra rules. Two items on it matter here.

• AI meant for recruitment or selection. That means aiming job ads, sorting applications and rating candidates.
• AI used to make decisions that affect the terms of work. That includes promotion and dismissal. It also covers handing out tasks and tracking how well people work.

Take a tool that tells your HR team what the standard says. It isn't on either list. A tool that screens resumes or flags employees for review is. The difference isn't the tech. It's whether the output decides something about a person. Heading toward hiring tools? Read our guide to AI for recruiting first.