Twelve months after AI agents went from keynote demo to production workflow, HR teams have a clear-eyed view of what works, what costs more than advertised, and where a human still has to sit.
Recruiting is where agents actually landed. Scheduling, candidate screening questions, sourcing outreach, and status-chaser communications are now autonomously handled at a growing share of large employers. The wins are real: recruiter time freed from logistics, faster candidate response, fewer falls-through-cracks moments. The candidates mostly don't object — they object to silence, and agents are very good at not being silent.
The integration bill surprised everyone. Vendors sold agents as plug-in features; enterprises discovered that an agent is only as useful as its access to the ATS, the calendar system, the HRIS, and the career site — and wiring that together cleanly ate months of IT capacity. Teams that budgeted for licenses but not integration work have a graveyard of half-connected pilots to show for it.
Data quality turned out to be the tax on everything. An agent acting on stale requisition data, wrong hiring-manager assignments, or inconsistent job families doesn't fail loudly — it fails convincingly. The organizations that got value fastest were the ones that spent the unglamorous first quarter cleaning the data the agents would read.
On bias and compliance, the picture matured. The bias-audit and transparency regimes that started in New York and the EU pushed vendors to ship logging and explanation features, and buyers learned to ask for them in every contract. Nobody thinks the compliance question is settled; everybody treats audit trails as a prerequisite now.
Human-in-the-loop settled into a consistent pattern. Agents propose; humans decide. Screening shortlists, interview summaries, and offer-letter drafts run through machine hands first, but selections, rejections, and anything touching adverse action stay with a person — partly for legal cover, partly because trust in the process demands it.
The capability gap inside HR teams is the quiet headline. Prompting, supervising agents, auditing outputs, and redesigning workflows around them are new skills, and most functions are training reactively rather than building programs.
The lesson heading into 2026: the technology was never the hard part. Data, integration, and operating model were — and the teams that treated AI as an operations project rather than a software purchase are the ones now scaling with confidence.