An AI interviewer works in four stages: it builds an interview plan from a role description, conducts a live voice or video conversation with the candidate, evaluates the responses against that plan in real time, and produces a report with evidence for every score. The technology underneath has changed a lot in the last two years, and that shift is the reason this generation of tools is worth taking seriously in a way earlier chatbot-style screeners weren’t.

Before any candidate joins, the system needs to know what “good” looks like for the role. The weakest tools require a recruiter to fill out a form: pick competencies from a dropdown, set weightings, submit. That works for simple, well-defined roles, and breaks down for nuanced ones where the right evaluation criteria depend on team context, tech stack, and seniority.

Better tools use a conversational setup: paste in a job description, describe the context, and the system proposes a structured plan you can revise before the first interview runs. This matters more than it sounds. It’s the difference between filling out a form and actually briefing an interviewer on what you need.

This is where the underlying AI capability shows up most clearly. Early AI interviewers were essentially scripted chatbots: a fixed list of questions, asked in order, regardless of what the candidate said. That’s not evaluation, it’s a survey with a voice interface.

What’s changed is how the best tools use large language models during the live conversation. Instead of scoring keywords against a script, a well-built AI interviewer engages with the substance of an answer: it notices when a response is vague and asks a follow-up, tracks what’s already been covered so it doesn’t repeat itself, and adjusts the difficulty or direction of the next question based on how the candidate is performing.

The test for whether an AI interviewer is doing real evaluation isn’t whether it can ask a good question. It’s whether it can recognize a weak answer to a good question and probe further. That adaptiveness is the single biggest differentiator between tools in this category.

As the interview happens, the system scores what’s said against the plan from Stage 1. This is typically continuous, not a single evaluation at the end. The model is tracking coverage of each competency area throughout the conversation, which is also how it decides what to ask next.

The output is a structured report: an overall recommendation, scores by competency area, and ideally a full transcript with citations linking each score to the specific exchange that produced it. This last part, evidence, is what separates a report a hiring manager can defend from one they have to take on faith. If a score can’t be traced back to something the candidate actually said, there’s no way to audit it, and no way to catch it when it’s wrong.

This is the right question to ask. Keyword-matching against a rubric isn’t evaluation, and if a candidate figures out the rubric, that kind of system is easy to game.

The distinction that matters is whether the system is reasoning about the content of an answer, adaptively, in context, or just checking for the presence of expected phrases. A system that presses a vague answer for specifics, tracks conversational context across 20 minutes, and adjusts its line of questioning based on candidate performance is doing something closer to real evaluation than pattern matching, even though both are technically “AI.”

Chakra, HackerRank’s AI interviewer, is designed around that adaptiveness principle: it builds a structured interview plan from a conversational setup, adjusts its questions in real time based on what the candidate says, and produces reports where every rating cites the specific transcript moment that supports it. If a candidate gives a high-level answer to a technical question, Chakra probes further before moving on. That’s the behavior that separates evaluation from a scripted survey with better production value.

→ See the setup and reporting flow at chakra.sh