AI Workflows · HR content build · Updated July 2026

AI Job Descriptions and Interview Scorecards

You searched for a job description generator. What you actually need is a bias-checked build you can defend in an audit. Here is the workflow: the prompts, a red-flag word list, and an adverse-impact self-check you run before anyone gets hired.

Why "just use a JD generator" is a legal trap

Type "AI job description generator" into any search bar and you get a hundred free tools that spit out a polished posting in five seconds. The output looks professional. That is the trap. A generator optimizes for reading well, not for hiring fairly, and it learned to write from millions of real postings that were themselves full of coded language, inflated requirements, and quiet exclusion.

Here is the part the tool never tells you. When your posting or your interview process screens out a protected group at a higher rate, the fact that a machine wrote the words is no defense. Regulators and courts look at the effect on real people, not the software you used. Connecticut made this explicit in its CART Act, which our briefing on Connecticut's CART Act (PA 26-15) breaks down: deploying an automated hiring tool does not shield an employer from a discrimination claim. New York City goes further for the tools themselves. Under NYC Local Law 144, an employer using an automated employment decision tool must commission an independent bias audit each year, publish a summary of the results, and notify candidates that the tool is in use.

So the honest reframe is this. AI can absolutely help you build the hiring content. It cannot take the liability off your desk. The workflow below keeps the speed and removes the trap.

The machine wrote the words. You still own the outcome. Build like you will have to prove it.

The compliant AI hiring-content workflow

Two documents. One clean process. The job description sets who you are looking for. The structured scorecard sets how every interviewer will judge them, using the same criteria and the same rating scale. Build both with AI, bias-check both by hand, and you have a hiring process that is faster to produce and easier to defend than the one you have now.

Drafting a bias-checked job description (prompt + red-flag word list)

Start with the role, not the tool. Collect the reporting line, the three or four outcomes this hire must deliver in year one, the genuine must-have skills, and your standard sections for level, location, and pay transparency. A generator invents these when you leave them blank, and invented requirements are exactly what creeps in and screens out good people. Give the model the real brief instead.

Example prompt: "You are helping an HR team draft a job description. Role: Operations Manager, reporting to the VP of Operations, hybrid in Denver. Year-one outcomes: stand up a vendor scorecard process, cut fulfillment errors, and coach two team leads. Must-have skills: process design, vendor management, and comfort with data. Write a clear, inclusive posting with these sections: about the role, what you will do, what you bring, nice to have, and how we work. Use plain language. Do not invent a salary, a degree requirement, or a benefit I have not given you. After the draft, list any phrase that could discourage qualified candidates and suggest a plainer replacement."

That last line does double duty: it produces the draft and a first-pass bias flag in one call. Then you run your own eyes over it against a short list of the usual offenders.

Ask the model to check the draft against those categories, then judge its suggestions yourself. It is good at catching the ten-years-for-an-entry-role inflation. It is not the final authority on which requirements are real. You are.

Building a structured interview scorecard with AI

A biased hire rarely starts in the posting. It starts in an unstructured interview where five people ask five different questions and score on a feeling. A structured scorecard fixes that. Every panelist evaluates the same criteria, asks from the same question bank, and rates on the same anchored scale. Research on structured interviewing has shown for decades that it predicts job performance better and drifts toward bias less than the freewheeling version. AI is genuinely useful here, because building the scaffold is tedious and the model does it in one pass.

Example prompt: "Build a structured interview scorecard for the Operations Manager role below. Define four to six evaluation criteria tied directly to the year-one outcomes and must-have skills. For each criterion, write two behavioral interview questions and a 1-to-5 rating scale where 1, 3, and 5 each have a concrete, observable anchor describing what that answer looks like. Keep every criterion job-related. Do not include anything about personality fit, culture fit, age, accent, family status, or where someone went to school. [Paste the finished job description here.]"

What you get back is a table of criteria, questions, and anchored ratings. The anchors matter more than anything else on the page, because a 1-to-5 scale with no definitions is just gut feel wearing a number. A good anchor for a "process design" criterion might read: a 5 describes a specific process they built, the problem it solved, and how they measured the result; a 3 describes involvement in a process without ownership or a metric; a 1 speaks only in generalities. Now two interviewers who saw the same answer land in roughly the same place, and you can compare candidates on evidence rather than impressions.

Two edits are yours to make before it ships. Cut any criterion that is not clearly tied to doing the job, since "culture fit" is where bias hides. And confirm the questions are legal, with nothing that fishes for age, disability, family plans, or national origin. The model drafts the structure. You certify that it is job-related and lawful.

The four-fifths rule and an adverse-impact self-check

Here is the measurement that turns "we think our process is fair" into "we checked." The four-fifths rule comes from the federal Uniform Guidelines on Employee Selection Procedures. It works as a rule of thumb: if the selection rate for any protected group is less than eighty percent of the rate for the group with the highest selection rate, that gap is generally treated as evidence of adverse impact worth investigating.

You can run the basic check yourself in a few minutes.

1. For each group, divide the number selected by the number of applicants. That is the group's selection rate.
2. Find the group with the highest rate.
3. Divide every other group's rate by that highest rate.
4. Any result below 0.80, or eighty percent, is a flag to investigate.

A worked example. Say 100 men apply and 50 are advanced, a rate of 50 percent. And 100 women apply and 35 are advanced, a rate of 35 percent. Divide 35 by 50 and you get 0.70. That is seventy percent, which sits under the eighty percent threshold, so this stage flags for adverse impact and you dig into why. Maybe a job requirement is doing the screening. Maybe a scorecard criterion is off. The rule does not prove discrimination on its own, and small applicant counts can swing the ratio, but it tells you exactly where to look before a pattern hardens into a lawsuit. Run it at each stage where people get cut: resume review, interview, offer.

Keeping records for a bias audit

A defensible process leaves a trail, and the trail costs about a minute per hire to keep. For each role, save four things: the final job description with a note that it was AI-drafted and human-reviewed, and by whom. The scorecard template every panelist used. The completed scorecards with their ratings and notes. And your four-fifths check on the selection rates at each stage.

That folder is the difference between a confident answer and a shrug if you are ever questioned, whether by a regulator, a plaintiff's attorney, or your own leadership. If you are using any vendor tool that touches candidate selection, add its audit summary and candidate notices to the same folder, since that is what Local Law 144 and similar rules expect you to be able to produce. For the wider data and privacy posture behind all of this, our companion piece on how HR teams use AI safely covers what belongs in a tool and what never should.

Human sign-off checklist

Before a job description posts or a scorecard goes to a panel, a person confirms all of it:

• Every requirement is real and truly needed for the job. Inflated tenure and unnecessary degrees are out.
• The posting is clean against the red-flag word list, with coded and gendered language removed.
• The pay range is present if your jurisdiction requires it, and accurate.
• Each scorecard criterion is job-related, with no "culture fit" standing in for bias.
• Every rating scale has concrete anchors, so a number means the same thing to every interviewer.
• No interview question fishes for age, disability, family status, or national origin.
• No AI tool scored, ranked, or rejected any actual candidate.
• The record is saved: what was drafted, who reviewed it, and the four-fifths check on your rates.

Part of TLY's AI Workflows → see the function-wide AI in HR playbook for how this fits the rest of the job.

Build the process, not just the posting

Anyone can generate a job description in five seconds. Building a hiring process that is fast, fair, and provable, with bias-checked postings, structured scorecards, an adverse-impact check, and a records trail, is the actual skill. That is what we teach: how to put AI to work across the HR function without ever handing it a decision about a person that it should not make.

Sources: Uniform Guidelines on Employee Selection Procedures (four-fifths rule); NYC Local Law 144 (automated employment decision tools bias audit and notice); Connecticut CART Act (PA 26-15) as covered on TLY's AI regulation desk; TLY hands-on use of AI to draft and bias-check job descriptions and structured interview scorecards (July 2026). Employment rules around AI vary by jurisdiction and change quickly; verify against current guidance and your own counsel before relying on this page.