AI for Recruiting and Job Platforms: Where It Pays and Where the Law Starts

Most of the money in recruiting AI comes before a hire is made.

For a job platform, it's in content. Good content makes a weekly job alert worth opening. It's also in the employer profiles that make a listing worth a click.

For a staffing firm or an in-house team, it's in research, outreach and booking. That is the work from the day a role opens. It runs up to a first talk with a candidate. These uses have high volume and light legal risk.

Further down the funnel, software ranks, scores or turns down people. The value there is real. But the law is dense. New York City has bias audits. Illinois has an AI amendment to its Human Rights Act. There is the EU AI Act and the US four-fifths rule. And there is a federal case against a screening vendor.

So build at the top of the funnel first. Then go further down with care. Keep a person on each decision and an audit on each tool.

Upstream First: the hiring funnel, and where the law sits

We call this order Upstream First. Read the funnel at the top of this page. Go from the first stage to the last. Two things rise together. One is how close the AI is to a choice about a person. The other is how much law applies.

Value doesn't rise with them. For platforms, the biggest payoffs sit at the start.

Where AI pays in recruiting, ranked

1. Job alert content made for each reader (platforms). Articles, Q&As and email intros are matched to each reader. The match is on field, role and level. The weekly alert then reads like it was written for them. Take a job site with a large list split into segments. For that site, we estimate the worth at $320,000 to $2.4 million or more a year. It lifts results on each send. And it takes the place of two to five content roles.
2. Employer research and profiles (platforms and agencies). The AI builds a full, even profile. It starts from an ATS link or a job feed. An ATS is an applicant tracking system. Employers use it to post jobs. They also use it to track who applies. Take a job site that adds 50 or more employers a month. We estimate this is worth $100,000 to $370,000 a year to it. That comes from 100 to 200 or more hours of research a month.
3. Outreach and re-engagement (recruiters). The AI drafts first notes to a job seeker. It works from their public profile and the role. It also wins back people already in your database. That second job is most often worth more. The method is in how to reactivate old leads.
4. Scheduling and candidate questions. It books interviews across calendars. It answers "what's the pay range?" and "is it remote?" from the job record.
5. Matching and picks. It suggests roles to job seekers, and job seekers to recruiters. This helps. It is also where the law starts to care how the model ranks people.
6. Screening and ranking. It scores resumes, ranks applicants and studies interviews. These uses carry the heaviest rules. They are the most likely to be challenged.

The content engine: personalization at platform scale

A plain job alert is a list of links. A good one opens with something the reader wants to read. Take a nurse just out of school. She might want to know how to negotiate a first nursing contract. Take a sales rep five years in. He might want to know how to go from field sales to running a team.

We built an engine like this for a job platform. The platform has tens of millions of job seekers. It finds topics across six kinds of public source. Those are videos, blogs, news, podcasts, talks and books.

It writes new posts for each segment. Each fits a field, role and level. Each one gets 5 to 10 follow-up Q&As, plus five email intros. Then it publishes them. Each weekly alert pulls the article for its reader's segment.

Each piece moves through a set of stages: found, enriched, tagged, draft, published. And each one credits its sources.

Two outside rules decide whether this kind of tool helps or hurts.

• Search. Google's spam policies name a problem called scaled content abuse. It means pages made in bulk to game search. In Google's words, it is when "many pages are generated for the primary purpose of manipulating search rankings." The policy applies "no matter how it's created." Now take pages written for real readers in a segment. They cite real sources. They are the opposite of that. Pages written for keywords are not.
• The inbox. Gmail has rules for those who send more than 5,000 messages a day. They must offer a one-click unsubscribe. They must prove the mail is really theirs with SPF, DKIM and DMARC. Those are three email checks. And they must keep reported spam below 0.30%. Content that doesn't fit its reader drives the spam rate up. Staying under that line is the price of sending at volume.

Employer profiles from a link

Adding an employer to a job site takes research. You need the firm's own site and what it says about itself. You need what its staff say about it. And you need what its listings say about pay and perks. By hand, that is hours per employer. And the profiles come out uneven.

We built an engine for this too, for a job platform. It starts from an ATS careers link or an XML job feed. That is a file that lists the employer's open jobs. It finds the employer's own site and its key pages. It gathers public ratings from seven review sites. It pulls culture and perk signals from the job listings.

Then it writes three pieces that are safe for the brand. They are an overview, why join, and the culture. It adds five intros of 300 characters each. All told, it fills 50 or more fields per employer.

It was built to save 2 to 4 hours of research per employer. Anything it couldn't confirm goes to a review queue. It does not go into the profile.

Screening and ranking: where the law starts

The rules below apply to tools that decide who moves on. They cover tools that do much of the deciding, too. Read them as the reason screening comes last. They are not a reason to avoid it.

This isn't legal advice. Your employment lawyer reads the law. The wider picture for HR teams is in AI for HR.

The vendor isn't safe from this either. Take a case against Workday. In May 2025, a federal court in California ruled on it. It let a nationwide age bias case go ahead as a group. The ruling came at an early stage. The group is people aged 40 and over. They were denied job picks through Workday's platform. Workday says the case is without merit.

The point for a buyer is plain. The firm that makes a screening tool can be sued over it. So can the employer that uses it.

The four-fifths math is simple. You can run it on your own funnel this afternoon.

Say a screening tool moves 120 of 300 people forward in one group. That is 40%. It moves 60 of 200 forward in another. That is 30%. Divide the lower rate by the higher: 30 divided by 40 is 0.75.

That is below 0.80. So the guidelines would generally treat it as evidence of adverse impact. A New York City bias audit would report the ratio.

The fix is never to hide the number. It is to find which input drives the gap. And keep a person deciding each case while you do.

What it costs to run

We scoped the content engine at $190 to $900 a month. That covers tools and model costs. The employer profile engine came to $290 to $1,100. Most of each bill goes to the search and data services that feed it.

Now take our live systems at 1,000+ daily AI interactions. Their all-in model costs typically run $200 to $1,200 a month. That is on the client's own accounts. Details are in what it costs to run an AI system. To weigh cost against value, see how to calculate AI ROI.

Here is what takes the work. First, the list of segments: field, role and level. Agree on it with the people who own the email program. Second, the links to the CMS and the email platform. A CMS is the software that publishes your pages. Third, an editor's review of a sample of each week's output.

What goes wrong with AI in recruiting

1. Content written for keywords, not readers. Thin pages at scale are what the spam policy describes. Write for a segment.
2. An employer fact no one checked. A perk the employer doesn't offer. A rating from a different firm with a like name. Each field needs its source and a confidence score. The low ones need a person.
3. Alerts that annoy. Content that fits no one pushes spam complaints up. Gmail tells bulk senders to stay below 0.30%. The alerts that carry your revenue depend on that number.
4. A screening tool with no audit. It was bought on a demo and used for a year. No one ever measured it by group.
5. Zip codes in the model. Illinois names them as a stand-in for protected classes. Location data needs a lawyer's look before it needs a data scientist's.
6. Automation that decides. Some tools turn down applicants with no person seeing them. Regulators and plaintiffs look for those first.

What to ask a vendor

• For content: what sources does each piece cite? How is it matched to a segment? Show us three pieces for three segments.
• What happens to our spam rate and unsubscribes? How will we see it week by week?
• For profiles: where does each field come from? What happens to a field it couldn't confirm?
• For any tool that ranks or screens people: has it had a bias audit? By whom, and when? Can we see the summary?
• Which inputs does the model use? Does any of them stand in for a protected class? Include location.
• Where does a person decide? Can the tool turn someone down with no person involved?
• Whose accounts, keys and data does it run on? What do we keep if we leave?

When not to use AI in recruiting

A small list. Under about 5,000 names on the list, a tool like this adds work. It won't move the numbers. A well-picked newsletter will.

A trickle of employers. Say you add fewer than about 10 new employers a month. Then a template plus a researcher costs less.

No way to publish. You need a CMS and an email platform. Other software must be able to plug into both. Without them, the work has nowhere to go.

You want every word written by a person. Then this is a job for a writing team. AI should stay in the research.