HR AI implementation playbook: from approved use case to a live agent

The use case is approved and the budget is signed. Most of what decides whether the agent works happens in the next few weeks, and most of those decisions sit with the people team rather than the builder.

This HR AI implementation playbook follows one agent from the day its use case is approved to the day it runs for a small group of real users. It’s written for a people team at a company of 500 to 5,000 employees. A team that size often has one HRIS owner and no AI team, and its security team may never have reviewed an AI system.

The playbook is six gates. Each one has an owner on the people team, a written output and one question that decides go or no-go. After the gates come a definition of “live”, the first 30 days and a list of what you should hold on go-live day.

One agent runs through every gate as the example: the HR Policy Question Answering Agent from our HR Agent Map. It answers questions like “how much notice do I give” from your own handbook, and it hands anything personal to a person. It’s an illustration, not a client result.

Culturro does this work for people teams, so our Proof and Build steps are the worked example. The gates hold whoever runs them, including your own IT team. If the use case isn’t approved yet, start with how to prioritize HR AI use cases and come back.

What you need before the first gate

You don’t need an AI team. You do need these people named before anything starts:

• An accountable owner who signs each gate. This is usually the CHRO or VP People.
• The process owner, who runs the process today and will own the number after go-live.
• The HRIS owner. They know which fields exist, which ones are reliable and who has access.
• A security contact who sees the architecture in week one.
• Three to five future users who test it before anyone else does.

You also need three inputs. The first is a step-by-step written description of the process as it runs today. The second is an honest read on the data, and our HR data readiness checks cover it. The third is a budget line for the build and for the running costs after it. Budgeting for AI in HR lists the cost lines.

In our strategy step you name an owner on your side for every number before anything is built. Do the same here. A sponsor without a named process owner gets a pilot that nobody runs.

The gates from approved use case to live

Gate	Owner on the people team	Written output	The question that decides go or no-go
1. Scope what the agent may touch	CHRO with the process owner	Scope note: systems, fields, actions, and the points where a person decides	Can we name every action the agent takes and the person who approves it?
2. Test on non-production data with real users	Process owner with the future users	Evaluation set and a go or no-go note	Did the users reach for it, and can we live with the failure cases we saw?
3. Security review	Security contact, supported by the HRIS owner	Architecture document, data-flow diagram, signed review	Will security approve it running inside our environment, on our terms?
4. Run with no action taken	Process owner	Log of agent outputs set against what people actually did	Would the agent’s outputs have been right, for the right people?
5. Go live for a small group	Process owner, who owns the number	Metric with a baseline, pause rule, runbook	Is the number moving with no incident we can’t explain?
6. Widen	CHRO	Usage and outcome report, plan for the next group	Does the report justify more users, and can the owner still keep up?

Gate 1: Write down what the agent may touch and what stays with a person

The CHRO and the process owner write a scope note. It lists the systems the agent reads and the fields it may see. It lists the actions it may take and every point where a person makes the call. It also says what the agent must never see. The policy agent needs an employee’s location, entity and worker type to pick the right policy version, but it has no reason to see case notes or health data.

If you did the strategy step with Culturro, part of this note already exists. Anything a works council or regulator would treat as high-risk gets the human-decides design in writing in that step. Our rule is plain: we don’t build systems that make autonomous decisions about hiring, pay or promotion, and the same goes for discipline and termination. A human decides, always.

A “no” here usually means the use case is wider than anyone said out loud. Cut it to the part you can name and send the rest back to the backlog.

Gate 2: Test on non-production data with the people who will use it

In Culturro’s Proof step the agent runs in one of two places. One is our sandbox on synthetic data. The other is a sandbox in your own cloud tenant, on synthetic or de-identified data your security team approves. Each Proof takes a few weeks at a fixed fee per use case. It ends with a written go or no-go against what your team saw in use: accuracy, the failure cases, cost per query and whether people reached for it. At the end the sandbox is torn down or handed to your team.

The last item matters most. If the future users tried it twice and went back to emailing the HR inbox, accuracy doesn’t matter yet.

Keep the test questions and the answers you expected. That’s your evaluation set, and you’ll run it again at every later gate. For the policy agent it’s a list of questions shaped like last quarter’s inbox, each with the answer an HR advisor would give.

A “no” here is cheap, because you’ve paid for a few weeks of learning instead of a few months of build.

Gate 3: Security review with the architecture and data-flow documents

Security review is where many HR pilots stall, as we describe in why HR AI pilots stall. Bring the documents to the first meeting, not the last.

Culturro’s Build step takes the architecture document and data-flow diagram to your first security review meeting. We build the agent inside your environment under your security team’s controls. The model runs where you choose, on a provider your security team approves. The provider’s no-training and retention terms are in writing, and access runs through your identity and your devices.

The security contact signs the review, often with conditions. Write the conditions down, because they become part of the runbook.

When security says no, ask what would turn it into a yes. A different model provider might do it, or a narrower data set, or a different place to run it. If nothing would, stop and write down why.

Gate 4: Run with no action taken

The agent now runs on live inputs but takes no action, so a person still does the work as usual. Its outputs are logged, and the process owner sets each one against what the person actually did. An advisor still answers every policy question by hand. The agent’s draft answer sits in the log beside each one.

This step is often called shadow mode. It needs three things: a log, a reviewer and time on the reviewer’s calendar.

Check two things. Were the outputs right? Did the agent fire for the people it should, and only for them?

A “no” sends you back to gate 2 with the new failure cases added to the evaluation set.

Gate 5: Go live for a small group

Pick one team, one site or one country. Keep it small enough that the process owner can read a real share of the outputs.

Three things are agreed in writing first: how people will use it, which number says it works and the pause rule. The number needs a baseline from today’s process. At Culturro we agree the first two before anything ships, and our Build includes onboarding for the people who will use it.

The people team picks the number, and the vendor and the partner don’t. The policy agent’s page on the HR Agent Map suggests ticket deflection and first-contact resolution. Hours spent per case or the share of cases reopened suit other agents.

Gate 6: Widen

Widen when the usage and outcome report says so and the process owner can still keep up with the review. Add one group at a time and rerun the evaluation set before each step. A policy agent that works for one country still needs the next country’s handbook tested before that country gets it.

A “no” here is normal. Some agents should stay with one team, and that still counts as finished.

Where the human decides

A rule that lives only in a policy document gets worn down by a busy quarter, so the human-decides rule has to be built in. It shows up at four gates.

• Gate 1 names each point where a person makes the call and who that person is.
• Gate 2 tests the hand-off itself. Can the user see why the agent suggested what it did, and can they overrule it in one step?
• Gate 4 confirms the agent takes no action while a person still does the work.
• Gate 5 puts the hand-off points in the runbook, so a new HR advisor learns them in their first week.

Our guide to AI agents in HR sets out four ways a person can sit in an agent’s flow. A person approves each output before it goes out, or reviews the agent’s work, or sets the rules the agent runs inside. The fourth is none, for a job that isn’t a decision about anyone. Gate 1 is where you pick one for each step and write it down.

The policy agent is a “sets rules” agent. HR owns the source documents and the hand-off rules, and a person takes every hand-off. On hiring, pay and promotion the agent prepares and a person decides. Discipline and termination work the same way, and we design the system so it can’t do otherwise.

What “live” means and the first 30 days

An agent is live when three things are true:

1. Real users use it for real work, on production data, inside your environment.
2. A named owner watches a number against a baseline.
3. Someone can pause it today without calling the vendor.

If any one is missing, you’re still in a pilot.

Here’s the rhythm we’d suggest for the first 30 days. The process owner reviews a sample of outputs every week. The HRIS owner watches access and data flows. The security contact gets a short weekly note, and the CHRO sees the number every fortnight.

Write down what triggers a pause before go-live. These are the triggers we’d start from:

• The agent gives an answer that breaks a policy or a law.
• It acts for someone outside the agreed group, or shows data about them.
• Users stop opening it.
• The number moves the wrong way for two reviews in a row.
• Cost per query climbs well past what the Proof showed.

Watch how often the agent fires, too. A sharp rise or fall usually means something upstream has moved, such as a new field in the HRIS or a renamed team.

The policy agent adds one more thing to watch. Its page on the HR Agent Map names a pitfall in plain words: “Nobody reads the unanswered-question log.” That log tells you which policies are unclear, so give it an owner before day one.

At day 30 the process owner reports the number against the baseline, and the CHRO decides whether to hold, fix or widen.

We don’t give a target number. We haven’t found an independent benchmark for HR agents, and the thresholds vendors publish are their own. Your own baseline is the fair comparison.

What you hold on go-live day, and what a partner should leave behind

Check this list before you sign off a build, whoever built it. Ask why if an implementation partner can’t hand over an item.

• The code in your repository under your accounts.
• Documentation of how the agent works and what it connects to.
• Runbooks for pause and restart, prompt changes, key rotation and incidents.
• Prompts and instructions your team can read and edit.
• The evaluation set from gate 2 with every failure case added since.
• A usage and outcome report your team can pull without the partner.
• The architecture document and data-flow diagram plus the signed security review.
• A clear access position: which accounts the partner still holds and when they end.
• Hosting and model contracts in your name and billed by your own providers.

At Culturro the code, documentation and runbooks transfer to you at handover. The prompts and evaluation sets are yours too, and nothing depends on a Culturro platform. Your team can pull a usage and outcome report without us, and we hold no production credentials after handover. A support period is included. Hosting and model usage are billed by your own providers, so you see the running cost from day one.

IT leaders judge partners on the same thing. CIO.com’s piece on choosing an AI implementation partner (read 20 Oct 2026) lists “willingness to share skills for lasting impact” as one of its criteria. The handover list is the written half of that.

Mistakes we expect at each gate

Several of these match the four stalls in why HR AI pilots stall. The gate is where each one gets caught.

• Gate 1: the scope note lists what the agent does but leaves out what it must never see. Security finds the gap when they ask which fields it reads.
• Gate 2: the test runs on a spreadsheet export, which behaves differently from the live system. Test on the real source or a faithful synthetic copy.
• Gate 2: the build team tests it without the users and learns later that the users never wanted it in that form.
• Gate 3: security sees it last, so weeks of build wait on a review that could have started in week one.
• Gate 4: the team skips shadow mode to hit a date. The first wrong answer reaches a real employee.
• Gate 5: nobody agrees the number, so nobody can defend next year’s budget.
• Gate 6: the team widens because the launch week went well. A good first week says little about month three.

Build it with your own IT team or with a partner

Your own IT team can run these gates if someone there has built and run a system that calls a language model. Security also needs a view on AI providers already, and the team needs the weeks to spare.

An outside partner fits when those don’t hold, or when the people team wants the first agent live before IT can make room. Put the handover list in the contract.

Many companies do both. A partner builds the first agent, and the internal team runs it and builds the second. Culturro’s roadmap and Proof are written so your own team can do the build if you choose.

The partner route has costs. You pay a fee your IT team wouldn’t charge, and some knowledge stays with the partner unless the handover is real. Before you sign, name the person on your side who runs the agent the day after handover.

Where Culturro fits

Culturro works with people teams at companies of 500 to 5,000 employees. We run the Proof and Build steps described above.

• Proof takes a few weeks per use case at a fixed fee. It ends with a written go or no-go.
• Build typically takes a few weeks to a couple of months per use case and is scoped after the Proof. It runs inside your environment under your security team’s controls, and the code, documentation and runbooks are yours at handover.

We hold no certifications and don’t claim any. An optional month-to-month retainer is offered after the strategy step. It isn’t needed to keep a delivered system running. The full steps are on our services page.

If you don’t have a use case yet, the HR Agent Map shows which HR agents are worth building first. If you’re holding an approved one now, book a 30-minute call. We’ll walk through the six gates for it with you and whether to build it on your own systems.

Frequently asked questions