AI in the HR Stack: What Works, What Doesn’t, What to Pilot
An honest map of where AI is delivering value in HR today, where vendors are overselling, the risk and regulation landscape, and a 30/60/90 pilot framework.
Every HR vendor now has an ‘AI’ tab. Most of what works is unglamorous: better summarization, smarter search, faster drafting. The risk is treating AI as a decision-maker for hiring, performance, or termination — which is increasingly illegal and almost always a mistake.
Where AI helps today
| Workflow | What AI does well | Time saved |
|---|---|---|
| Sourcing | Boolean-free search, candidate matching, outreach drafting | 30–60% of recruiter sourcing time |
| Resume / scorecard summaries | Synthesize across candidate materials | 5–10 min per candidate |
| Job description drafting | First draft from a scorecard + JD library | 60–80% draft time |
| Interview note synthesis | Turn recorded interview into structured scorecard prompts | 10–15 min per interview |
| Policy + handbook Q&A | Conversational HRIS / policy search for employees | Deflects 30–50% of HR tickets |
| Manager enablement | Drafting feedback, prepping difficult conversations, summarizing 1:1 history | Highest leverage; most underused |
| L&D content | Skill-tagged learning paths, summarization of long content | Faster curation, not creation |
| Analytics narratives | Natural-language summaries of dashboards | Hours per leadership update |
Where AI is oversold
- ‘AI-driven hiring decisions’ — predictive scoring without explainability is legally risky and rarely outperforms structured-hiring + work samples
- Video-interview personality inference — repeatedly shown to be biased and unreliable; effectively banned in NYC for hiring use without bias audit
- Engagement ‘sentiment AI’ that reads private chat — culture-destroying and often illegal
- ‘Autonomous performance reviews’ — reviews are accountability artifacts, not generated text
- ‘AI HR copilot that replaces your HRBP’ — copilots augment, they don’t replace judgment
Use AI to accelerate human decisions in hiring, performance, and discipline. Do not use AI to make those decisions. The legal, reputational, and ethical exposure is not worth the marginal speed.
Risks and regulation
| Jurisdiction | Rule | What it requires |
|---|---|---|
| EU | EU AI Act (in force 2024–2026) | HR uses classified as ‘high risk’: documentation, bias monitoring, human oversight, conformity assessment |
| EU | GDPR Article 22 | Right not to be subject to solely automated decisions with legal/significant effects |
| New York City | Local Law 144 (AEDT) | Annual bias audit + candidate notice for automated employment decision tools |
| Illinois (US) | Artificial Intelligence Video Interview Act | Notice + consent + reporting for AI video interview analysis |
| Colorado | SB24-205 (AI Act, effective 2026) | Developer + deployer duties for ‘high-risk’ AI including hiring |
| EEOC (US) | Technical assistance on AI + Title VII | Disparate-impact liability for AI hiring tools |
| UK | ICO + DSIT guidance | Lawful basis + DPIA + human oversight for AI in HR |
Treat any AI in HR as ‘high-risk by default’: document it, monitor it for adverse impact, give humans the final call, and let candidates and employees know it’s being used.
A 30/60/90 pilot framework
- 1Day 0–30 — Define and de-riskPick one workflow, one team, one metric. Write the data flow. Run a DPIA / bias-risk assessment. Decide what ‘human in the loop’ means concretely.
- 2Day 30–60 — Pilot in production with guardrailsRoll out to one team. Compare AI-assisted vs control. Track the metric weekly. Capture qualitative friction.
- 3Day 60–90 — DecideKeep / kill / expand. If keeping, write the runbook, train users, set the governance owner, schedule the next audit.
Resist the urge to pilot ten use cases at once. The work is not the model — it’s the governance, change management, and integration around it.
Governance you actually need
- An inventory of every AI feature in your HR stack (including ones turned on by vendors by default)
- A named owner for AI risk in HR (often the Head of People, with Legal/Privacy support)
- A standard intake — vendor questionnaire on training data, bias testing, explainability, data residency, opt-out
- DPIA template for any new AI use touching personal data
- Annual bias audit for any AI used in hiring or promotion decisions
- Candidate and employee notice language in standard policies
- An ‘off switch’ — vendor must let you disable the AI feature without losing the rest of the product
What to watch in 2026
- Agentic AI in HRIS — multi-step actions across systems (provisioning, onboarding, comp-cycle support)
- Skills inference at scale — turning roles + projects into a live skills graph
- AI coaching for managers, embedded in 1:1 + feedback flows
- Continued regulatory expansion — assume more states/countries will follow NYC + EU
- Open-source HR copilots — privacy-friendly alternatives to vendor lock-in
Read next
All playbooksEvery category of HR software, what it actually does, where the boundaries blur, and how to think about a stack that grows with you.
The minimum operating system for fair, fast, and predictive hiring at any company size.
A vendor-evaluation framework you can run in two weeks — covering job-to-be-done, scoring, references, security review, contracts, and the exit path.