Think about the best interview you've ever been part of — either side of the table. What made it good? Almost certainly not the questions themselves. It was what happened around them: the interviewer who heard something interesting in your answer and pulled on it. The one who gently pushed back on a claim you made, not to trip you up, but because they wanted to understand if you'd thought it through. The one who noticed when your energy changed as you described a project you were genuinely proud of.
That's a real interview. It's a conversation with stakes, with momentum, with a human on the other side reading signals you didn't know you were sending.
Now think about most AI interview tools. A screen. A timer. A question you've seen variants of on dozens of prep sites. A blinking cursor waiting for you to fill a text box, or a camera recording your answer with no one watching, no one reacting, no one there.
The gap between those two experiences isn't a technical detail. It's the entire thing.
What a Real Interviewer Actually Does
When we talk about a great interviewer, we're describing a cluster of behaviours that are deceptively hard to replicate:
They probe. Not with a pre-planned follow-up, but in direct response to something specific you said. "You mentioned the migration was complex — what made it complex? Was it the data model, the sequencing, the team coordination?" The question changes because the answer changed it.
They counter. A skilled interviewer will occasionally push back on something you've said — not aggressively, but deliberately. "That's an interesting approach. What would you say to someone who argued that caching at that layer would create consistency problems?" This isn't a gotcha. It's a test of whether you hold your view for good reasons, or just haven't been asked to defend it yet. A candidate who hasn't thought it through will wobble. A candidate who has will push back with clarity.
They read the room. Confidence isn't just what you say — it's the speed at which you say it, the absence of hedging, the willingness to commit. A good interviewer notices when a candidate lights up talking about something they built versus going flat describing something they observed from a distance. They notice hesitation that isn't uncertainty — it's someone choosing words carefully because the detail matters to them. They notice the candidate who's performing versus the one who's just talking.
They follow energy, not the script. The interview that was supposed to cover five topics ends up spending twenty minutes on one because something interesting surfaced. That's not a failure of process — it's the process working. The interesting thing is almost always more valuable than the fifth checkbox on a rubric.
The signal in a great interview isn't in the answers. It's in how a candidate responds to being pushed, surprised, and genuinely engaged.
Why Almost All AI Interview Tools Miss This Completely
The dominant model in AI interviewing today — asynchronous video screening — is almost the exact opposite of a real interview. The candidate answers into a camera. No one responds. There is no room to read. There is nothing to follow. The "conversation" is a monologue with a structured prompt.
As we've covered in our guide to AI screening, these tools were designed to solve a volume problem, not a quality problem. They move candidates from 500 to 30 faster than a recruiter could. That's valuable. But it's not an interview — it's a triage mechanism that happens to look like one.
Even the more sophisticated platforms — coding assessments like Codility or HackerEarth — are measuring task completion, not thinking. A candidate who gets the right answer through memorisation and one who got it through genuine understanding of the underlying data structures look identical in the output. There is no counter, no probe, no way for the system to find out which is which.
The result is a generation of AI hiring tools that have automated the easy part of screening while leaving the hard part — the part that actually predicts who will do well in the role — entirely to a human interview that often happens too late, too inconsistently, and with too little structure to be reliable.
The Three Things an AI Interviewer Needs to Actually Work
Building an AI that interviews like a real interviewer requires solving three distinct problems that the current generation of tools hasn't seriously attempted.
It needs to listen, not just hear. Most AI tools treat candidate responses as inputs to a classifier. Did they mention the right keywords? Did they stay within the expected range? A real interviewer extracts meaning from the specific way something was said — the detail chosen, the framing, what was left out. An AI that interviews well needs to construct an understanding of what the candidate actually communicated, not just what they technically said.
It needs to respond, not just advance. The next question in a real interview is a function of the previous answer. In most AI tools today, the next question is predetermined. Even tools with branching logic are following a decision tree, not generating a genuine response. A meaningful AI interview requires the system to formulate follow-up questions that are directly reactive to the specific content of what was just said — not from a bank of pre-written alternatives, but composed in the moment.
It needs to hold a position. This is the hardest part. A counter in an interview isn't just a question — it's a position. "I'd push back on that" means the interviewer has formed a view and is testing whether the candidate can engage with it. For an AI to do this well, it needs to reason about the answer it just received, identify something worth challenging, and frame the challenge constructively. That's a significantly higher bar than question generation.
Why This Is Now Technically Possible
For most of the history of AI hiring tools, these three capabilities were out of reach. Building them required language models sophisticated enough to reason over a free-form conversation in real time, generate genuinely novel follow-up questions, and hold context across an extended back-and-forth. That wasn't available at the quality level required for professional hiring.
It is now. The generation of language models that emerged over the last two years is capable of all three — not perfectly, but well enough to be useful. The question isn't whether the underlying technology can do it. It's whether anyone is building a product designed around doing it well, specifically for the hiring context, rather than repurposing a general-purpose chatbot with a set of interview prompts bolted on.
What Diyam AI Is Building With Ray
This is the bet that Diyam AI is making with Ray — that the right design goal for an AI interview tool is not "screen candidates efficiently" but "replicate the things that make a great human interviewer great."
Ray conducts live conversational interviews that adapt based on what the candidate says. When an answer is thin, it probes. When a claim seems underexplored, it counter-questions. When a candidate describes something they clearly understand deeply, it goes further into that territory rather than advancing to the next topic on a checklist. The conversation has momentum — it goes where the signal is.
For technical roles specifically, this matters more than it does for general professional hiring. A candidate describing a distributed system they built is either someone who understands the trade-offs they made, or someone who was in the room while someone else made them. A set of fixed questions can't reliably tell the difference. A conversation that probes what happened when the cache was cold, or what they'd change if they did it again, or what the team argued about before making the final call — that can.
Ray is also designed to pick up on confidence signals that async video tools completely miss. Not facial expressions or tone of voice as proxies for personality — the kind of pseudo-scientific analysis that drew legitimate criticism and has been largely abandoned — but the structural confidence in an answer: the willingness to commit to a position, the speed and specificity of technical claims, the difference between someone who says "we used Kafka for the message queue" and someone who says "we considered Kafka and RabbitMQ, and chose Kafka because of the retention guarantees we needed for the audit trail."
The second candidate didn't just use Kafka. They chose it. That distinction is everything in a senior hire — and it's exactly what a tool designed to automate volume screening will never surface.
The Hard Problem That Remains
Building toward a real-interview experience in AI is the right goal. But it comes with honest challenges that teams evaluating these tools should understand.
Reading energy and confidence isn't just about what candidates say — it's about pace, silence, the physical signals of certainty or hesitation. A voice-based AI interview captures some of this; a text-based one captures almost none. The closer AI gets to replicating a real interview, the more format matters, and video or voice is much closer to the signal-rich environment of a human conversation than a text interface.
There's also the candidate-experience dimension. A real interview is a two-way process: candidates are also evaluating the company, and the interview is part of that impression. An AI that probes and counter-questions well can feel engaging and intellectually serious. One that does it clumsily can feel adversarial. Getting that calibration right — rigorous without being hostile, persistent without being repetitive — is a design and product problem, not just a model capability problem.
And transparency remains non-negotiable. As we've documented, candidates who don't know they're being interviewed by AI report significantly worse experiences than those who are told upfront. The closer an AI gets to feeling like a real interviewer, the more important it is that candidates know it isn't one.
The Direction of Travel
The tools that will define AI hiring over the next few years won't be the ones that screen most efficiently. Efficiency at the volume stage is a largely solved problem. The tools that will matter are the ones that can do what currently only a skilled human interviewer can do: find out what a candidate is actually like when they're pushed.
That's a harder problem. It requires better models, better product design, and a clearer theory of what a great interview is actually measuring. Diyam AI's Ray is one of the most deliberate attempts we've seen to build toward it — designed explicitly around the idea that the goal is a real conversation, not a more efficient form submission.
We're still early. But the direction is right. And in a market full of tools that automate the easy parts of hiring while leaving the hard parts unchanged, a tool built around getting the hard parts right is worth paying attention to.
Learn more about Ray at diyamai.com. For how Ray compares to the current field of AI interview tools, see our full comparison piece.