AI is already making decisions about your people

Not in a dramatic way. In the ordinary way.

A scheduling tool that decides who gets the Saturday shift. An applicant tracking system that ranks résumés before a human sees them. A performance dashboard that flags who’s falling behind. A note-taker summarizing a one-to-one.

Most companies didn’t buy any of that as “AI.” They bought scheduling software. The feature arrived in a product update, switched on by default, and nobody circulated a memo.

Why that matters

The legal exposure attaches to the decision, not to the vendor.

If a tool produces a pattern that disadvantages a protected group, the employer owns the outcome. Anti-discrimination law is about effects, and it has never cared much how the effect was produced. “The software did it” has not been a defense in any other context and won’t be in this one.

The regulatory picture is also moving, unevenly, by jurisdiction. Some states and cities now have specific requirements around automated employment decision tools — notice, bias auditing, or both. More are drafting.

The first problem is that nobody knows

Most audits start with a surprise. The applicant tracking system was scoring candidates on criteria nobody configured. The scheduling platform was optimizing labor cost in a way that consistently gave the same people the worst shifts. Somebody’s team had been running performance notes through a chatbot.

You can’t govern what you haven’t inventoried. So that’s where we start: every place software is making, shaping or ranking a decision about a person — hiring, scheduling, pay, performance, discipline.

Then: what is it actually deciding?

There’s a difference between a tool that surfaces information and a tool that makes a call. Sorting applicants by years of experience is a filter you chose. A relevance score derived from patterns in your past hiring is something else — it has learned from your history, including the parts of your history you wouldn’t defend.

We work through each one: what it’s deciding, what it learned from, who reviews it, and whether a human is genuinely in the loop or nominally in the loop.

What we do about it

Some uses are fine and we’ll say so. Drafting a job description, summarizing notes you then check, answering a policy question — low stakes, easily supervised.

Some need a human decision-maker and documentation showing there was one. Any screening, ranking or scoring that affects who gets hired, promoted, scheduled or disciplined.

And some should stop. We’ll tell you which.

Then the practical layer: what managers are allowed to use, what they’re not, what gets recorded, and what candidates and employees are told.

Why we’re comfortable advising on this

We use AI across our own operation. That’s the reason, and it’s not a small one — we’ve seen where it saves real time and where it produces something confident and wrong.

What’s included

• Inventory of where AI touches employment decisions, including features inside tools you already own
• Assessment of what each tool decides versus what you assumed
• Human-in-the-loop requirements: which decisions need a person, and what “reviewed” has to mean to count
• Documentation, because defensibility depends on showing your process
• Candidate and employee notice practices
• Manager guidance: what to use it for, what not to, what to record
• A review cadence, because your stack changes under you

Where we stop

This is a fast-moving regulatory area that varies by jurisdiction. We’re not your attorney, and where a specific tool raises a legal question we’ll say so rather than guess.

We’re also not here to talk you out of using AI. We use it. The point is knowing which decisions you can hand over and which you can’t.