AI in HR: What Works Today, What to Wait On, and Who Decides

More than half of organizations don’t involve HR in their AI strategy. That’s 52%, from SHRM’s research report The State of AI in HR 2026. The same research finds just 39% of organizations have AI live in HR at all, and 56% of HR functions don’t formally measure whether their AI investments work.

If your CEO has asked you for an AI plan for HR, that’s the gap you’re being asked to close.

The demos will keep coming, and most of them show something that works. A demo can’t tell you which idea to start with, though, or who in your company signs off when the system gets something wrong.

Sort AI in HR by the trust it needs

Most guides to AI in HR management sort it by department: recruiting, onboarding, performance, learning. Some sort it by technology: agents, language models, predictive analytics. Both tell you what exists. Neither gives you an order to do it in.

SHRM has done real sorting work. Its researchers sorted 138 AI use cases by risk and cost and validated them with 1,722 HR professionals. It’s serious research, and it sits behind membership, so a CHRO who isn’t a member can’t see the order.

We sequence by trust. Most AI frameworks put the biggest return first. In HR, a wrong answer lands on a person, so we start with what a wrong answer costs.

The 10X HR-AI Framework scores one idea on three lenses. Impact asks which number it moves, how much of the work it takes on, and who would notice. Trust to earn asks what a wrong answer costs, who decides, and how sensitive the data is. Readiness asks where the data lives and whose name is on the number. All three answer to employee lifetime value, or eLTV: how well people perform and how long they stay. Impact raises it, trust protects it, readiness keeps it.

Every idea lands in one of four tiers. Build now needs a real number to move, little or some trust to earn, and data and an owner that are ready. Proof first needs a useful number, the same low trust, and data at least half ready, so it’s safe to try on test data. Build with guardrails needs a real number and data at least half ready, and it has real trust or a lot of trust to earn. The guardrails, the notice and who decides get designed before anything is connected. Not yet means one thing has to change first. We’ve scored three real use cases end to end if you want to see the scoring in full.

Where HR AI actually stands right now

Start with the people doing the work. Pew Research Center found that 21% of US workers say at least some of their work is done with AI, up from 16% roughly a year earlier. Most, 65%, still don’t use AI much or at all in their job.

Inside HR, SHRM’s State of AI in HR 2026 puts the adoption of AI in HR at 39% of organizations. Another 7% plan to start this year, and 31% have no plans. Across the whole business, 62% of organizations use AI somewhere. HR is behind the company it serves.

Where HR teams are using AI, it’s clustered in three places:

• recruiting, at 27%
• HR technology, at 21%
• learning and development, at 17%

The people using it say it helps. In the same research, 87% of HR professionals report improved efficiency, and 75% report better work quality.

What’s proven today, area by area

Our map covers all 122 agents, across 12 HR areas. Every one carries a verdict and a line saying who decides. The count, from our own data: 48 proven, 55 strong case, 7 emerging, and 12 human decides. Proven, strong case and emerging say how established that kind of agent is across HR in general. Human decides means the job itself is a decision about a person. Your own version gets its tier when you score it on the 10X HR-AI Framework, because your data, your owners and your rules change the answer.

The advantages of AI in HR show up first where a wrong answer is cheap to catch. Here’s one proven agent from eight of the 12 areas, with who signs off and the number it’s meant to move.

HR area	What’s proven	Who signs off	What it moves
Workforce Planning & Org Design	Headcount Forecasting Agent	Finance and HR leadership set and change the plan; the agent never adds or removes a position on its own	Headcount and cost variance to plan
Talent Acquisition	Candidate Sourcing Agent	The recruiter decides who’s contacted and sends the outreach	Sourced-to-interview rate
Onboarding	Day-One Provisioning Coordination Agent	IT, facilities and security owners fulfill; the hiring manager sets what access the role needs	Day-one readiness rate
HR Operations & Shared Services	HR Policy Question Answering Agent	HR owns and approves the source documents and hand-off rules; the agent escalates when it’s unsure	Ticket deflection, first-contact resolution
Compensation, Benefits & Payroll	Benefits Q&A Agent	The benefits team and the carrier or provider decide eligibility exceptions, claims and appeals	Ticket deflection, HR hours reclaimed
Performance & Talent Management	Goal Setting Support Agent	The employee and their manager agree the goals; the agent drafts and flags	Goal completion and update rates
Learning & Development	Learning Path Recommendation Agent	The learner decides what to take on; the agent never enrolls anyone	Learning completion and engagement
Manager Enablement & Coaching	Manager Coaching Agent (the one we run, as Merlin)	The manager decides what to say and do; HR never sees what was discussed	Manager capability and confidence

Read the third column first. Each line names a person or a team. The agents find, draft, answer, forecast and chase. A person owns every call that lands on somebody.

The one we run ourselves

The last row is ours. Risely is our own leadership development AI product, live at risely.me. Its coach is Merlin, and Merlin has reached 5,000+ people from 40+ companies, in 40 languages, across 83 workplace skills. Of the 122 agents on the map, the Manager Coaching Agent is the one we verified by running it ourselves.

Here’s how it works. A manager brings a live situation into Slack or Teams: a report who’s gone quiet, a peer conflict, a promotion they can’t grant. Merlin asks the questions a good coach would ask, offers options, and helps them settle on an approach and the words to use.

Whether managers use it comes down to what HR can see. In Risely, a manager’s own coaching conversations are private. When a company assigns a coaching plan, HR receives a session summary and engagement level, never the details of what was discussed. Customer data isn’t used to train the underlying models.

The agent page sets out the rules we’d design into any coaching agent:

• It isn’t kept as an HR record and is never used in a process about anyone.
• It reads no HRIS, performance or pay data. It knows what the manager types and nothing else.
• When a situation needs a formal process, it points the manager to HR, legal or an ethics line.

The way this goes wrong is in the launch. Announce it as a tool HR rolled out, and managers assume HR reads what they type. Then they stay away. The privacy rules have to be built into the system, and managers have to hear them before anyone asks them to use it.

Same department, different verdict

Now take two agents from Talent Acquisition.

The Candidate Sourcing Agent searches your ATS, your CRM and permitted outside profile sources against a role brief. It drafts personal outreach. The recruiter decides who gets contacted and sends the message. If it suggests a weak match, the recruiter skips that one. It’s proven.

The Application Screening Support Agent reads each application against the published must-have criteria, summarizes the evidence for and against each one, and orders the pool for review, with reasons a recruiter can check and overturn. The recruiter or hiring manager decides who advances and who’s declined. They see every application, including the ones at the bottom of the queue. The agent never auto-rejects anyone. Its verdict on the map is human decides.

Same function. Same team. A wrong suggestion from the first costs a recruiter one skipped name. A wrong ranking from the second can shape who moves toward an offer.

Group AI by function and these two arrive together, on one slide, in one rollout. Sequence by trust and they come apart. Sourcing is useful fast once your candidate database is clean and you’ve confirmed your licenses let it search outside profiles. Screening needs candidate notice, and in New York City a bias audit, in place before day one. When we scored a screening idea end to end, it came back Build with guardrails. You can see how the other nine sit in Talent Acquisition’s 11 agents.

What to prove before you scale it

Seven agents on the map carry an emerging verdict, which means agents like these are still new across HR. They’re worth building, and each one needs a contained pilot before it goes any wider. All three below touch sensitive data, so the guardrails get designed before the pilot starts.

The Skills Gap Analysis Agent compares the skills a role needs with what people have declared, shown in projects or completed in learning, and shows where gaps cluster. An inferred skills profile can quietly turn into a performance record, so profiles are opt-in, visible to the employee and open to dispute. The proof passes when most people accept or lightly edit their own profile, and your L&D lead can name three gap clusters they didn’t already know. The employee confirms or disputes their profile. L&D and business leaders decide which gaps get the investment.

The Team Health Signal Agent watches aggregate signals a team has agreed to share, like pulse trends, whether 1:1s are happening, and after-hours load. It tells the manager when something is drifting. It never shows one person, and there’s no leaderboard. Several of those signals come from work systems that count as monitoring, which is why it comes after agents that stay private to the person using them, like the coaching one, have earned some trust. The manager decides whether and how to act. HR decides which signals are in scope and the minimum group size.

The Pay Transparency Compliance Agent checks that job postings carry the pay range each jurisdiction requires and drafts answers to employees’ requests for pay information under the EU Pay Transparency Directive and similar rules. The rules are still settling. A draft that goes out unreviewed is a legal disclosure. Total rewards and legal decide what gets disclosed, and they approve every response before it goes out, while the agent checks, drafts and keeps the log.

What stays a person’s call

The sharpest disadvantages of AI in HR sit in one kind of job: the ones where the work is a decision about a person. Twelve agents on our map carry the human-decides verdict. They’re built as decision support on purpose, and that’s permanent.

The Application Screening Support Agent is one, covered above. It’s also an automated employment decision tool under NYC Local Law 144, and high-risk under the EU AI Act.

The PIP Documentation Support Agent helps a manager and HRBP structure an improvement plan once they’ve decided one is warranted. It drafts the document and checks it’s complete. It never suggests starting a PIP, never assesses whether the plan has been met, and never drafts an outcome. The manager and HRBP decide all three.

The Disciplinary Process Support Agent checks that each step of the procedure was followed before the next one starts. It drafts the letter for the decision the hearing manager has already made. The hearing manager decides whether there’s a case to answer and what sanction, if any, applies, with HR advising. The agent never recommends a sanction.

The Restructure Impact Analysis Agent works out the roles in scope, costs, notice periods and consultation timelines, and flags adverse impact in a proposed pool. Leadership, HR and legal decide the pool, the criteria and every individual outcome. It never selects or ranks anyone for redundancy.

Your own version of one still gets scored. Ideas like these shape a decision about a named person, so they have at least a lot of trust to earn. One that stays at a lot, with a real number and data at least half ready, can score Build with guardrails. One that needs investigation or protected-characteristic data, like most disciplinary and restructure work, carries the most and comes back Not yet.

An AI that auto-rejects a candidate, sets one person’s pay with no human proposing it, or picks who’s in a layoff is a different case. It isn’t on our map at all. We don’t build systems that make autonomous decisions about people. Score an idea like that and it comes back Not yet, whatever else it scores.

That’s one of four ways the 10X scorer returns Not yet:

• The idea, as imagined, would make a decision about a person on its own.
• It carries the most trust the framework can score: a wrong answer would bring in a regulator, a works council or a lawyer, or it needs health, protected-characteristic, grievance or investigation data.
• The number it moves is too small, or too small for the trust it would have to earn.
• The data lives in spreadsheets or nobody’s sure where, and nobody has put their name on the number yet.

Each Not yet comes with the one thing to fix first. Fix it and score the idea again.

The gap between live and working

SHRM reports its 39%, 56% and 52% side by side and doesn’t claim one causes another. Our reading: when HR isn’t in the room where the AI strategy gets set, the tools that arrive in HR come with nobody’s name on the number they’re supposed to move. Live and working are different states, and the second one needs an owner.

Every agent above carries a line that names who decides. Write that line before you build, and put a name on the number it’s meant to move beside it.

None of this puts HR back in the AI strategy conversation at a company that’s left it out. That gap belongs to the executive team, and we don’t expect it to close quickly. What you control is your own list: which idea goes first, which waits for proof, and whose name sits on each one. If the board is asking, here’s what to tell the board about where you stand. If something’s already running and has gone quiet, start with why pilots stall.

---

The business case for AI in HR starts with knowing what to build first. A product demo can’t give you that.

In a strategy month, we take your shortlist of ideas, score each one against the 10X HR-AI Framework, and sequence the list. You leave with a written plan and a named decision-owner for every use case. The month is a fixed fee. Each use case we build after the month is a fixed fee for an agreed scope, quoted once the proof is done. Hosting and model usage are billed by your own providers.

If you’d like to see where your list lands, book a 30-minute call.

Frequently asked questions