What Happens to Candidate Questions During an AI Interview

Candidate questions during an AI interview arrive somewhere in the middle of a screening call, when a candidate stops answering and asks something. What is the salary for this role. Is it work from home. Who will I speak to after this. How many rounds are there.

The short answer: the agent answers company questions from documents your organisation has uploaded to its Knowledge Base, quotes only what they say, and points the candidate to the hiring manager for the rest. The line between what it can and cannot answer is drawn by what you have given it.

This post is about that line, and about where the answers the agent cannot give have to live instead. The wider design of the screen, described in the AI interview without a human interviewer, depends on getting this part right rather than leaving it to chance.

What Questions Do Candidates Actually Hold?

Candidate questions fall into four groups, and it helps to sort them before you decide who should answer each one.

Money comes first, usually framed as a check rather than a demand: is this in the range I saw. Then logistics: shift, location, whether hybrid means two days or four. Then process: how many rounds, how long, when will I hear. And now and then a question about the work itself: what will I actually be doing in the first three months.

Notice that most of these are not assessment questions at all. They are decision questions. The candidate is weighing up whether to keep spending time on you, and they are making that call in the same twenty minutes you are using to evaluate them.

How Does the Agent Answer a Company Question?

The agent has a tool built for exactly this moment. When a candidate asks about the company, it searches the documents your organisation has uploaded to its Knowledge Base and answers from what it finds.

Three rules govern that answer, and all three are worth knowing before you upload anything.

First, it is told to use only what the documents say. It is not meant to fill gaps from general knowledge, and when a document gives a specific number, name or date, it is told to quote it rather than blur it into something vague.

Second, when the documents do not contain the detail asked, it is told to say so and point the candidate to the hiring manager. It is told not to guess. A candidate who asks about a benefit your documents never mention should hear that the agent does not have that detail, not an invented answer.

Third, there is a limit. One call takes only a small number of company questions, and once a candidate goes past it the agent closes the call instead of letting a screening interview turn into an open question session. That is deliberate. The call exists to assess, and the limit keeps it that way. Answers are also short, a sentence or two, in the tone of someone answering in passing.

Why Is the Knowledge Base a Publishing Decision?

Whatever you upload can be spoken to a candidate, in a recorded call, in your company’s name, and a stale figure is quoted as confidently as a current one. Treat the upload as a publishing decision, not a filing exercise.

That sounds obvious until you look at what teams tend to have lying around. An old policy document with last year’s leave numbers. A pay grid that was meant for internal use. A benefits summary from before the last restructure. Because the agent is told to quote specific numbers, the candidate hears any of them as your offer. Load what you would be comfortable seeing on a printed sheet handed to every applicant, the way you would for a careers page:

If a fact should be said only by a person who can qualify it, keep it out and let the agent hand that question to the hiring manager, which is exactly what it is told to do.

There is a liability point underneath this. A recorded conversation at scale is a paper trail. Every statement the agent makes about pay, benefits or policy is a statement attributed to your company, repeated for every candidate who asks. I would rather hear it say the hiring manager can fill them in a hundred times than hear it quote a stale number once.

Why Put the Answers in the Invitation?

The invitation is the last thing a candidate reads before deciding whether to show up, and most teams waste it. It says congratulations, here is your link, it is fifteen to twenty minutes.

Everything that determines whether a candidate should attend belongs there instead: the shift pattern, the location and whether it is negotiable, the range if you publish one, the number of rounds after this, and the honest fact that this first conversation is with an AI screening call rather than a person. We have written separately about what belongs in the interview invitation email, and the candidate-facing version of the same ground in talking to someone anxious about an AI call.

A booking confirmation is the second slot. It carries a calendar attachment with a reminder, which a candidate will reopen on the morning of the call, so check that what it says matches what your invite promised.

For roles where email is not where your candidates live, WhatsApp carries the invite, the reminder and the booking confirmation on a per-role basis. There is a real constraint worth understanding before you plan around it. Business-initiated messages have to use templates approved in advance under Meta’s WhatsApp Business rules, and only four slots change: who the candidate is, which role, which company, and the link or time.

You cannot improvise a paragraph about the shift allowance into a template. That fixed structure is exactly why the job description and the invite email have to carry the detail. The delivery mechanics are covered in how WhatsApp invites and reminders work.

How to Write a Job Description for the Second Read

Most job descriptions are written to attract applications. Almost none are written for the moment that decides things here, which is a candidate re-opening the post an hour before their screening call to remind themselves what this job is.

That second read is where your unanswered questions get settled, one way or another. If the post says competitive compensation, the candidate fills the gap with the last number they heard from a similar company, and they will arrive at the call anchored to a figure you never offered. If it says flexible working, half your applicants have decided that means fully remote.

I would rather a job description lose me some applications in week one than produce a screening queue full of people whose expectations were set by their own optimism. Name the shift. Name the city and how many days on site. Name the number of rounds and roughly how long the process takes. If you publish a range, publish the real one.

There is a selfish argument for this beyond fairness. Every fact you place in the post is a fact the agent never has to be asked about, and a constraint the candidate has already accepted before they booked. The brief that produces the screening questions is the same document, which is why a vague brief produces a vague screen in both directions at once.

Which Questions Do Phase-0 Knockouts Answer First?

Some candidate questions get resolved without ever being asked, because the disqualifiers at the top of the call surface the constraint first.

A candidate who is asked in the opening minute whether they can work a rotational night shift now knows the role has rotational night shifts. Someone asked to confirm they can be on site in Pune five days a week has just learned it is not hybrid. The question they were holding until minute four has been answered in minute one, from the other direction.

This is a real argument for writing disqualifiers as plain, specific questions rather than coded ones, which is the design pass in knockout questions. Can you start within thirty days tells a candidate something true about your timeline. A vaguely phrased availability question tells them nothing and leaves the assumption intact.

What to Do With Questions That Arrive Anyway

Some will land regardless, and the cheapest mistake in this entire flow is sending invitations from an address nobody reads.

Find out where replies to your invitations land, and make sure a person reads that inbox. A candidate who replies with a question before the call is showing you more intent than the ones who booked silently, and in high-volume hiring that signal is nearly free to collect. For WhatsApp, each attempt is stamped on the candidate’s row, so did this person actually get pinged is answerable per candidate rather than per campaign.

Then make the human round genuinely answer things. A candidate who clears the screen with a strong verdict should reach a person who is briefed and willing to talk about money, not a second structured assessment. The handoff after that point is its own discipline, which we covered in what to do after a Strong Go.

What Do Candidates Ask After the Call?

Once the conversation ends, candidates want three things: confirmation it was received, a sense of when they will hear, and, for some of them, feedback. The first is easy and mostly automated. The second is a promise you have to actually keep. The third matters because a candidate who asks how they did is still interested, and a non-reply turns that interest into a story they tell their friends about your company.

The operational trap is the On Hold verdict. On Hold is where candidates go to be forgotten. Someone sitting in that band for eleven days has effectively been rejected without being told, and by the time you come back to them they have accepted something else. Give every held candidate a date by which they get a real answer, which is the discipline we set out in every On Hold candidate needs a decision date.

What silence costs you

The failure mode is not complaints. It is a candidate who quietly decides you are not serious. Someone who asked about the shift and got nothing back does not send a follow-up. They stop replying, accept the offer from the company that answered, and never appear in your funnel data as a loss. They appear as a no-show, which is easy to file under candidate behaviour rather than under something you caused. We have written about the larger version of this in the applicants you never reply to.

SHRM’s 2026 analysis of AI in HR is framed around a move from hype to measured, human-centered impact, which is a fair description of what candidates expect from an automated first call too. Candidates notice when a process has removed the people. They just rarely tell you.

Two adjacent cases carry the same duty. An agency desk owes candidates clarity on where a scorecard goes, as the submission document explains, and a founder with no recruiter has nobody but themselves to catch an unanswered email, as running a first round alone tends to prove.

How to Audit Your Own Invitation in Five Questions

Pull up the last invitation your team sent and read it as a candidate who has three other processes running. Then ask:

1. Can you tell what the salary range is, or at least that there is one and when it will be discussed?
2. Can you tell where the job is and how many days a week you would be there?
3. Do you know whether the first conversation is with a person or a system, and how long it lasts?
4. Do you know what happens after it, and roughly when?
5. Is there a reply address that reaches a human, and does anyone read it?

If your invitation answers only one or two of them, fixing that takes an afternoon and costs nothing, and it removes most of what candidates would otherwise try to ask a voice agent that may not have the answer.

How to Test What the Agent Tells Your Candidates

Before you rely on any of this, put it to the test. Run the demo the way a candidate would, and ask the agent the three questions your own candidates ask most often. Listen to whether it answers, what it quotes, and what it hands back to a person. What it can quote for your candidates depends on what your organisation has uploaded.

The live demo interview runs about five minutes and needs no signup, and it is built on the same agent a real candidate meets.

Keep the questions where you can control them

An unattended screen is a measuring instrument that can also answer a few questions, and the second job is only as good as what you have loaded into it. It will quote what your documents say, hand back what they do not, and stop after a handful of questions.

So do the unglamorous work around it. Decide deliberately what goes into the Knowledge Base, put the decision-relevant facts in the job description, repeat them in the invite, encode the hard constraints as questions the candidate hears in minute one, and make sure a person reads the reply address. Where a candidate’s questions need a person from the start, see roles that need a human first round.

The screen buys you the right to have fewer of those conversations. It does not buy you the right to have none.