HR Advisory · AI & Technology

AI Agents in HR: What Actually Works in 2026 (and What's Still Hype)

Over 90% of Indian firms have piloted generative AI in HR. Only 38% find it highly relevant. The gap between those two numbers is where the real lessons live.

There is a number worth sitting with. Over 90% of Indian firms have piloted generative AI in HR functions, but only 38% report high relevance for their organisation today. That is not a story of failure — it is a story of a field still separating what works from what merely demos well. If you are deciding where to spend your team's limited attention, the gap between 90% and 38% is the most useful thing to study.

The shift from chatbots to agents

The first wave of HR AI was chatbots that answered policy FAQs. Useful, narrow, and largely solved. The 2026 shift is toward AI as a coordinator that handles the grunt work between steps — the document chasing, the scheduling loops, the first-pass screening — rather than just answering questions. Think of it less as a smarter search box and more as an invisible junior coordinator who never forgets a follow-up.

What actually works today

Across deployments we have seen and the broader market evidence, four use cases consistently repay the effort:

• Resume screening and matching. Parsing and shortlisting against a clear rubric is genuinely faster with AI, provided a human owns the rubric and the final call.
• Interview scheduling. Multi-stakeholder coordination — the back-and-forth that eats a recruiter's afternoon — automates cleanly.
• Exit-interview analysis. Thematic analysis across hundreds of exit conversations surfaces patterns a human reading them one at a time would miss.
• Performance-review summarisation. Turning scattered manager notes into structured summaries saves real time, with the manager editing rather than drafting from scratch.

What is still hype

Equally important is naming what to be skeptical of. "AI that predicts who will quit" sounds compelling and usually rests on data too thin and too biased to trust for individual decisions. "Fully autonomous hiring" ignores that the legal and fairness risk of a bad automated rejection sits with you, not the vendor. And any tool promising to "remove bias" by adding more AI deserves hard questions about what it was trained on.

The part everyone skips: governance and DPDP

Here is the failure mode we see most: a team deploys a screening agent that quietly sends candidate data to a third-party model, with no record of consent, no retention policy, and no audit trail. Under the DPDP Act, that is a problem before it is a productivity gain. Employee and candidate data is personal data; feeding it to an AI system is processing it; processing requires a lawful basis and a governance framework.

The governance work is not optional and it is not glamorous: know what data each agent touches, where it goes, how long it is kept, and who can see the output. Build that before the tool, not after the incident. We go deeper in our DPDP and employee data guide.

Where to start if you are starting now

Pick one use case with a clear human owner and a bounded data footprint — interview scheduling is the safest first win. Run it on real data, measure the time saved, and write the governance note as part of the pilot, not as a follow-up. The teams that move from "piloted" to "high relevance" are not the ones with the most tools; they are the ones who picked fewer use cases and deployed them properly.

This is exactly the territory our AI for HR cohort is built around — participants build five working agents on their own data, with a governance module rather than a vendor pitch.