How a recruiting team writes shared norms for responsible AI use: an acceptable-use policy, human-in-the-loop rules, disclosure standards, and a clear map of what AI may and may not decide.

Diane runs talent acquisition for a 240-person regional hospital network, with a team of eight recruiters filling roughly 60 clinical and administrative roles a quarter. When AI tools arrived, they did not arrive on a schedule. One recruiter started drafting rejection emails with ChatGPT. Another was pasting full resumes, candidate names and all, into a free summarizer. A third quietly built a scoring rubric the AI applied to nursing applicants. None of it was malicious, and some of it was genuinely good work. But Diane realized she could not answer a simple question a hospital compliance officer asked her in a hallway: "When a candidate gets rejected, can you tell me whether a human or a machine made that call?" She could not. That gap is what a team agreement closes. Individual responsibility is necessary, but on a team of eight it is not sufficient. You need shared, written norms that say how this team uses AI, and crucially, what AI is never allowed to decide.

Why Shared Norms Beat Individual Good Judgment

Every recruiter on Diane's team had good judgment. That was not the problem. The problem was that eight people exercising good judgment independently produced eight different standards. One recruiter de-identified candidate data before using AI; another did not know that was expected. One reviewed every AI-flagged rejection by hand; another trusted the summary. When a team scales AI use, the variance between recruiters becomes the liability, not any single person's competence.

A team agreement is not a policy document handed down from legal. It is a shared commitment the team writes together: here is how we work, here is what we will not do, here is who decides. The difference matters because agreements people help write are agreements people follow. The legal and regulatory stakes are real, and they are the reason instincts alone do not protect you. In hiring, automated tools touch decisions that are directly regulated, and a hospital network operating across multiple jurisdictions has to assume the strictest of those rules applies to everyone.

The Legal Floor Every Recruiting Team Must Know

A responsible AI agreement is built on top of the law, not instead of it. Diane anchored her team's agreement to four widely recognized requirements, stated plainly so every recruiter could apply them.

NYC Local Law 144. New York City requires that automated employment decision tools used to screen candidates undergo an independent bias audit before use, that the results be published, and that candidates be notified the tool is being used. The law is built around human oversight: an automated tool may assist, but it cannot be the unsupervised gatekeeper. Even if your roles are not all in New York City, this is the clearest existing template for what regulators expect, so Diane treated it as the team's working standard.

EEOC and disparate impact. The Equal Employment Opportunity Commission has made clear that an employer remains liable for discrimination produced by an AI tool, including tools built by a vendor. A related yardstick auditors apply is the four-fifths rule: if the selection rate for any protected group is less than 80 percent of the rate for the highest-selected group, that is treated as evidence of adverse impact worth investigating. If your AI-assisted screen passes 50 percent of one group but only 30 percent of another, the ratio is 0.30 divided by 0.50, which is 60 percent, below the four-fifths threshold and a signal to stop and audit.

GDPR and data handling. For any candidate in the EU or UK, the General Data Protection Regulation grants a right not to be subject to a decision based solely on automated processing that produces a significant effect, such as rejection from a job. It also requires a lawful basis for processing personal data and data minimization. The practical rule for the team: do not paste identifiable candidate data into a third-party tool that has no data agreement, and keep a human in any decision loop.

Disclosure norms. Across these regimes the common thread is transparency. Candidates and hiring managers should be able to understand, in plain terms, where AI is used in the process. A disclosure norm is not just compliance hygiene; it is the cheapest trust you will ever buy.

The Five Elements of a Team Agreement

Diane structured the team's agreement around five elements. Each one answers a question that, left unanswered, becomes a gap someone fills with a guess.

Shared values. What does responsible AI use mean to this team, in one or two sentences everyone would sign? Diane's team landed on: "We use AI to support recruiting decisions and never to replace human judgment in hiring. We prioritize fairness and candidate respect over speed."

Specific practices. Values translated into concrete, observable behavior. Not "be careful with data" but "we de-identify candidate information before sending it to any third-party AI tool." A practice you cannot observe is a practice you cannot enforce.

A decision framework. The heart of the agreement: a clear map of what AI may do on its own, what it may assist with under human control, and what it may never touch. This is where the human-in-the-loop rule becomes specific instead of aspirational.

Accountability. Who owns what, and what happens when the agreement is not followed. On Diane's team, each recruiter owns the review of any AI output they act on, and systematic issues are surfaced to the whole team rather than handled privately.

Continuous improvement. How and how often the team checks whether the agreement still fits. Tools change, roles change, and regulations change. An agreement with no review date becomes wallpaper within a year.

What Each Element Looks Like Written Down

Abstractions are easy to nod along to, so it helps to see the sentences a team actually commits to. On values, Diane's team wrote four: AI supports decision-making and never replaces human judgment in hiring; accuracy and fairness come before speed; the team defaults to transparency about how AI is used; and candidate privacy and data security are respected as a matter of course. Notice that accuracy sits alongside fairness. A team that only promises to be fair can still ship a summary full of unchecked facts.

On practices, three sentences carried most of the weight. Every AI-assisted screening decision is reviewed by a human before a rejection goes out. Every summary that flags a concern about a candidate includes the evidence and the context behind the flag, so nobody acts on an unsupported worry. And candidate data is de-identified before it goes to a third-party tool.

On the decision framework, the team wrote the boundary in one line before elaborating it: AI is used for research and drafting support, and it is not used for final hiring decisions. They added a rule about growth as well, since most trouble arrives with a new use case rather than an old one: any new application of AI is discussed by the team before anyone implements it.

On accountability, two commitments: each recruiter is responsible for reviewing AI outputs before acting on them, and if someone uses AI in a way that breaks the agreement, the team discusses it together rather than letting it pass or handling it in private. On improvement, the team set a rhythm: monthly, they review one AI-related decision or process and look for something to improve; quarterly, they report errors and analyse them for patterns; every six months, they step back and reflect on what is working and what needs to change; and annually they revisit the agreement itself.

A Sample Agreement You Can Adapt

Diane's team found it easier to react to a draft than to invent one from nothing, so here is the shape of what they adopted. Treat it as a starting text to argue with rather than a form to sign.

Our commitment. We use AI to support recruiting decisions and improve our efficiency. We never use AI as a substitute for human judgment in hiring. We prioritize fairness, accuracy, and candidate respect.

Our practices. Five, written so anyone could tell whether they had been followed.

• Every piece of AI output we use for a hiring decision is reviewed by a human recruiter.
• We de-identify candidate information before sharing it with third-party AI tools.
• When we find errors or bias in AI output, we report them rather than quietly working around them.
• We verify critical facts in AI output against the source materials.
• We default to transparency, so candidates and stakeholders can understand how we use AI.

Our decision-making. AI may work autonomously on research, drafting, summarization, and information extraction. AI supports the recruiter, who decides, on red flag identification, evaluation of concerns, and candidate comparison. AI never makes final hiring decisions, never determines compensation, and never handles legal or compliance matters.

Our accountability. Each recruiter owns their review of AI output. If we discover a systematic issue, we discuss it as a team.

Our improvement. Monthly, we review one AI-assisted decision for quality and fairness. Quarterly, we analyse error patterns and implement improvements. Annually, we review this agreement and update it as needed.

Worked Example: A Tiered Acceptable-Use Policy

The single most useful artifact Diane's team produced was a tiered policy that sorts every recruiting task into one of three lanes. It removes the daily ambiguity of "is it okay to use AI for this?" by answering the question in advance. Here is the agreement her team of eight adopted, sized to their 60-roles-a-quarter workload.

Tier 1, AI may do this on its own (low stakes, no protected decision).

• Drafting job descriptions from an approved template, with a human edit before posting.
• Summarizing public company or market research for a hiring manager briefing.
• Extracting structured fields (years of experience, certifications, location) from a resume the candidate submitted.
• Drafting first-pass interview scheduling and logistics messages.

Tier 2, AI may assist but a human decides (judgment about a candidate).

• Flagging possible red flags in phone-screen notes. The AI suggests; the recruiter confirms with evidence before acting.
• Comparing two candidates against the same scorecard dimensions. The AI organizes; the recruiter ranks.
• Summarizing multi-interviewer feedback. The AI synthesizes; the hiring panel weighs it.
• Drafting a rejection message. The AI writes; a recruiter reads and sends.

Tier 3, AI may never do this (regulated or final decisions).

• Making a final hire, reject, or advance decision without a named human owner of that decision.
• Setting or recommending compensation.
• Auto-rejecting candidates at any stage with no human review of the rejection.
• Handling background-check, accommodation, or other legal and compliance matters.
• Processing identifiable candidate data through any tool without an approved data agreement.

The human-in-the-loop rule made operational: on Diane's team, no candidate is rejected at the screening stage until a second recruiter has reviewed the AI summary that informed it. The reviewing recruiter has a one-hour window to flag concerns; then the two decide together. With roughly 60 roles a quarter and an average of 20 screened applicants per role, that is about 1,200 screens, and the team agreed that the ones ending in rejection are the ones that get the second set of eyes. The rule did not slow hiring to a crawl, because Tier 1 and Tier 2 drafting work still saved time everywhere else. It simply guaranteed that the irreversible decision, telling someone no, always had a human signature on it.

Building a Peer Review Culture

The strongest team agreements have peer review built into them rather than bolted on, and the tone of that review decides whether it survives contact with a busy quarter. The structure is simple: before an important decision, another team member reads the AI output. The process is supportive rather than punitive, and everyone needs to feel that in practice, not just read it in the document. A reviewer who is hunting for someone to blame will get defensive submissions and shallow reviews; a reviewer who is helping a colleague get a decision right will get the real questions.

Give reviewers three questions to ask so the review does not drift into vague approval. Does this look right? Did the AI miss anything? Do you have concerns about it? Diane's team runs this as the rule described above: before rejecting a candidate, the screening recruiter shares the AI summary with another recruiter, who has one hour to review and flag concerns, and then the two decide together.

The payoff shows up in four places. It catches errors before they reach a decision, which is the obvious one. It makes accountability concrete, because a decision two named people reviewed is a decision two named people own. It spreads knowledge across the team, since reviewing someone else's AI output is the fastest way to learn what good and bad output look like. And it builds confidence in using AI at all, because recruiters who know their work gets a second read are more willing to use the tools well rather than avoid them quietly.

Building the Agreement and Setting Disclosure Norms

Diane did not write the agreement alone and hand it down. She drafted a first version, then ran a 45-minute team session built on a simple question for each line: "Can we actually do this every day?" Two practices got softened because they were unrealistic, and one got tightened because a recruiter pointed out a gap in how vendor tools handled data. The agreement people argue over is the agreement people own.

On disclosure, the team set three plain norms. First, candidates are told, in the application flow, that AI tools assist in reviewing applications and that humans make hiring decisions. Second, hiring managers are told which parts of a candidate packet were AI-summarized so they can read the source if a decision is close. Third, any new AI use case is discussed before it goes live, not discovered after. That last norm is what would have caught the recruiter quietly building a scoring rubric: not because the rubric was wrong, but because a tool that influences who advances is exactly the kind of thing NYC Local Law 144 expects to be audited and disclosed.

The Process, Step by Step

What Diane did compresses into six steps you can run with your own team. Start with your values: hold a discussion about what responsible AI use means to this particular team, doing the hiring you actually do, rather than importing someone else's words. Then draft the agreement, writing a first version that captures those values and translates each of them into a practice.

Get feedback next, and ask for it precisely: does this capture what we want, what is missing, and what is too much? The third question matters as much as the second, because an agreement that overreaches gets ignored in exactly the moments it was written for. Refine from there, incorporating the feedback and iterating until the team feels genuine ownership rather than compliance.

Then make it live. Post it where people see it, reference it by name in real decisions, and use it to settle the arguments it was written to settle. Finally, review and improve on a cadence, with monthly check-ins at minimum, asking whether it is working and what needs to change. The last step is the one teams skip, and skipping it is what turns a good agreement into a document nobody has opened since the offsite.

Anti-Patterns That Kill a Good Agreement

An agreement without teeth. The team writes a thoughtful document, posts it, and never references it in an actual decision. Within a month, habits revert. The fix is to use the agreement out loud: when a screening rejection comes up, someone says "Tier 2, who is reviewing?" The document has to live inside real decisions or it is decoration.

Rules too rigid to follow. If the policy forbids AI for tasks recruiters genuinely need it for, they route around it quietly, and now you have shadow AI use with no oversight at all, which is worse than where you started. Diane's tiered structure works precisely because Tier 1 gives people generous, sanctioned room to move fast on low-stakes work.

No improvement loop. An agreement written once and never revisited goes stale as tools and roles change. Build the review cadence into the agreement itself: monthly the team reviews one real AI-assisted decision for quality and fairness, quarterly it looks for error and adverse-impact patterns, and annually it revisits the whole agreement against current law.

A Short Glossary

A team agreement is a shared commitment to how a team will work together responsibly, written by the people who have to live with it. Peer review is the process in which team members review each other's work before it becomes final. Accountability is a clear, stated understanding of who is responsible for what, so that no decision ends up owned by nobody. And continuous improvement is the regular practice of reviewing how you work and adjusting it, which is what keeps the other three from calcifying.

Practice: Draft Your Team's Agreement

Begin by defining your values. Write three to five sentences describing what matters to your team about using AI responsibly, in language your team would actually use out loud. If a sentence sounds like it came from a compliance poster, rewrite it until it sounds like your team.

Then translate each value into practice. For every value, write what it looks like in action, concretely enough that an observer could tell whether it happened. "We prioritize fairness" becomes a specific behavior at a specific point in your process, or it stays a slogan.

Next, build your decision framework. Go through the tasks your team really does and sort them: what can AI do autonomously, what can it support while a human decides, and what may it never touch? Be specific to your roles, because a clinical hiring team and a sales hiring team will not draw the lines in the same place.

With those three pieces, draft the agreement using the sample above as a scaffold. Then share it and refine it. Ask your team the three feedback questions, incorporate what comes back, and iterate until the document is something the team owns rather than something you circulated. Only then finalise it.

Reflection Questions

What would your team say is most important about responsible AI use? Guess first, then ask them, and pay attention to the gap between the two answers.

What norms around AI do you already have, and are they explicit or implicit? Every team has norms; the question is only whether they are written down where a new joiner could find them.

Where do you think your team is most likely to struggle with responsible AI use, and how would you address that specific weakness in the agreement rather than in general terms?

And if you had a strong team agreement on AI in place tomorrow, how would your hiring actually change? If the honest answer is "not much," either your practice is already strong or the draft is not specific enough to bite.

Putting It Into Practice

A strong team agreement is what turns responsible AI use from an individual practice into a team culture, and culture is how quality gets sustained after the initial enthusiasm fades. When responsible use is embedded in how the team works, it stops being the careful exception and becomes the ordinary default.

This week, draft your agreement. Share it with your team, iterate on it until everyone owns it, and then do the part that matters most: use it. An agreement referenced in a real decision within its first fortnight tends to survive; one that is filed after the meeting rarely does.

Related Lessons

Designing Guardrails: What AI Can Do, What Requires Human Approval is the direct companion to the tiered acceptable-use policy here. It goes deeper on how to draw the boundaries between the lanes before you write them into an agreement.

Human Touchpoints: Strategic Moments for Human Review helps you decide where the human-in-the-loop rule should bite. Diane put it on screening rejections; that lesson helps you find the equivalent moments in your own process.

Transparency and Disclosure: Telling Candidates About AI Use expands the disclosure norms section into the practical question of what to tell candidates, when, and in what words.

Legal and Compliance Partnerships: Ensuring AI Use Is Defensible is where the legal floor becomes a working relationship. An agreement your compliance colleagues have read is far more durable than one they discover after an incident.

Feedback Loops: How to Report AI Errors and Improve System Performance supports the practice of reporting errors and bias rather than working around them, and it turns the quarterly error analysis into something with real inputs.

Auditing AI-Assisted Decisions: Sampling Methodology and Fairness Metrics gives the monthly and quarterly review cadence a method, including how to sample decisions and how to test for the adverse-impact patterns the four-fifths rule is meant to surface.

Closing

Reaching a written, owned team agreement is the end of the guardrails work and the beginning of the practice. You have the knowledge, the frameworks, and the tools to use AI responsibly and effectively in recruiting, but knowledge only becomes skill through use. In the months ahead, run these frameworks on real requisitions, refine them where they chafe, and build them into habits until nobody has to consult the document to know what the team does.

The goal was never to be better at operating AI tools. It is to build a recruiting practice that is fairer, more efficient, and more human than the one you had before.

Key Takeaways

• Shared norms beat individual judgment at team scale. Eight competent recruiters acting independently produce eight different standards. The variance, not any one person, is the liability. A written team agreement converts private good judgment into a consistent, enforceable practice.
• Build on the legal floor, not instead of it. NYC Local Law 144 mandates bias audits, disclosure, and human oversight for automated employment decision tools. The EEOC holds employers liable for AI-driven discrimination, with the four-fifths rule as a screening yardstick. GDPR bars solely automated rejection decisions for EU candidates. Treat the strictest applicable rule as the team standard.
• The decision framework is the core. A team agreement is only as strong as its answer to one question: what may AI never decide? Compensation, final hire or reject calls, and unreviewed auto-rejections belong in the "never" tier with a named human owner on every protected decision.
• A tiered acceptable-use policy removes daily ambiguity. Sort every task into AI-may-do, AI-assists-human-decides, and AI-may-never. The structure lets people move fast on low-stakes work while guaranteeing a human signature on every irreversible decision.
• Make the human-in-the-loop rule operational. "A human reviews it" is aspirational until you specify who, when, and within what window. No screening rejection without a second recruiter's review inside one hour is a rule a team can actually run.
• Disclosure is the cheapest trust you can buy. Tell candidates AI assists and humans decide. Tell hiring managers what was AI-summarized. Require that new use cases be discussed before going live, which is also what catches the unaudited tool before it becomes a compliance problem.
• Write it together and review it on a cadence. The agreement people argue over is the one they own. Bake in monthly, quarterly, and annual reviews so the agreement keeps pace with changing tools, roles, and law instead of becoming wallpaper.