Do we need a bias audit if we only use AI for sourcing, not final hiring decisions?

Yes, if the tool ranks, filters, or scores candidates in a way that affects who a human sees. NYC Local Law 144 covers any automated tool that substantially assists an employment decision. Sourcing tools that surface some candidates and bury others count. The safer move is to audit any tool that touches candidate selection, even if a recruiter makes the final call.

We use a third-party AI vendor. Isn't compliance their job?

Their audit covers their model. It does not cover how you use it. Regulators and enterprise buyers want to know your workflow, your configuration, and your oversight. Ask your vendor for their bias-audit results and their intended-use documentation. Then document how you apply the tool on your side. You own the decision, so you own part of the proof.

How much does a bias audit actually cost?

For a single tool, an independent audit typically runs a few thousand to low five figures, depending on data volume and how many tools you run. That sounds like a lot until you lose one enterprise RFP over a missing document. The audit is cheaper than the deal you did not win. Start with the tools that touch the most candidates.

Prove Your Staffing AI Is Fair: Governance Guide

Staffing firms are losing RFPs because they can't document how their AI makes decisions. Between NYC Local Law 144 and the EU AI Act, proof of AI governance is now a buying criterion. Here is what to build this week.

Enterprise buyers are asking staffing firms a question most cannot answer: prove your AI is fair. Not describe it. Prove it. With a document.

If you run AI for sourcing, screening, or matching, and you have no bias-audit trail, you are already losing deals. Not next year. Now. The fine risk was the old story. The new story is the RFP you never even placed in.

Three things changed at once. New York City's Local Law 144 requires a bias audit for automated employment decision tools, and enforcement is getting sharper. The EU AI Act classifies recruiting and candidate selection as high-risk, with obligations landing in August 2026. And enterprise procurement teams now put AI governance on the vendor scorecard, right next to insurance and data security.

You do not get to opt out of this by saying a human makes the final call. If your AI ranks, filters, or scores a candidate before that human sees them, it is part of the decision. Regulators know it. Buyers know it. The only question is whether you can show your work.

What Buyers Actually Ask For

The RFP language is getting specific. I have seen procurement questionnaires from healthcare systems, banks, and manufacturers that ask staffing firms to attach a bias-audit summary, name their tools, and describe human oversight. One VP told me they lost a seven-figure MSP renewal because a competitor handed over a clean governance packet and they handed over a shrug.

Here is what a buyer wants to see, in plain terms:

Which AI tools you use and what each one does in your process.
A bias-audit result for any tool that scores or ranks candidates.
Where a human reviews the output and can override it.
How you tell candidates an automated tool is involved.
What you do when the audit finds a problem.

None of this requires a data science team. It requires you to write down what you already do, find the gaps, and fix them before someone else finds them for you.

The Four-Part Trail You Can Build in 30 Days

You do not need a governance department. You need a trail. Four parts, and you can assign each one to a person this month.

1. An inventory of every AI tool touching candidates

List each tool. Note what it does, what data it uses, and where it sits in your workflow. Include the ATS features you forgot are AI. Include the sourcing plugin your recruiters bought without asking. Most firms find two or three tools they did not know counted. The inventory takes an afternoon and it is the foundation for everything else.

2. A bias audit for anything that scores or ranks

For tools that filter or rank candidates, get an independent bias audit. Local Law 144 requires it if you hire in NYC, and the format is a good template even if you do not. The audit checks whether the tool produces different outcomes across race and gender categories. Your vendor may already have one. Ask for it. If they cannot produce it, that tells you something about the vendor.

3. A human-in-the-loop record

Write down where a person reviews AI output and can change it. Not a vague promise. A real step. "Recruiter reviews the top 20 ranked candidates and can add anyone the tool scored lower." That sentence is worth more in an RFP than a paragraph of policy language. It shows the AI assists and the human decides.

4. Candidate notice and a fix process

Tell candidates when an automated tool is part of the process. A short line on the application works. Then write down what happens when an audit flags a problem. Who gets told. Who pauses the tool. Who signs off on the fix. Buyers want to know you have a plan for the bad day, not just the good one.

Why This Wins RFPs, Not Just Avoids Fines

Fines get the headlines. Deals get the money. The firms that treat governance as a sales asset are pulling ahead of the ones treating it as a legal chore.

Think about the buyer's side. A hospital or a bank has its own compliance obligations. When they hire you, your AI becomes their exposure. A firm that hands over a clean governance packet removes that worry. You just made the buyer's job easier, and you gave them a reason to pick you over the competitor who cannot answer the question.

This is the Build. Change. Adopt. pattern in action. You build the trail. You change how your recruiters work so the human step is real and recorded. You get the team to adopt it so the packet stays current, not stale after the first audit. The firms that skip the adopt step end up with a governance binder nobody has updated in a year. That binder loses deals too, once a buyer asks a follow-up question.

There is a speed advantage here as well. Right now, most staffing firms have nothing. If you build your trail this quarter, you show up to the next RFP with a document your competitor does not have. That gap closes fast. In 18 months this will be table stakes. For the next few, it is an edge.

Do This Week

Pull your team together for one hour and build the tool inventory. List every piece of software that touches a candidate and mark which ones use AI to score, rank, or filter. Do not solve anything yet. Just get the honest list on one page.

That list tells you which tools need a bias audit and where your gaps are. It is the first document in your governance trail, and it is the one thing standing between you and a real answer the next time a client asks whether your AI is fair.

Your AI Is Live. Can You Prove It's Fair?