The Better Job Interview Is Structured, Not Scripted

The candidate gives a polished answer about resolving a difficult launch. It has all the right shapes: a tense deadline, a cross-functional team, a decisive intervention, a successful result. The interviewer nods, writes “strong communicator,” and moves on.

What is still missing? Almost everything needed to judge the work. Which conflict did the candidate actually resolve? What did they say? What changed because of it? What would their colleagues describe differently?

Interviewers often treat structure and curiosity as opposites. One camp wants every candidate to receive exactly the same questions. The other wants a natural conversation that follows interesting threads. The better design combines both: standardize the evidence you need, then adapt the route you take to find it.

That combination matters as AI enters recruiting. A July 2026 working paper, Voice AI in Firms, reports a natural field experiment in which 70,000 applicants were randomly assigned to interviews led by human recruiters or live AI voice agents. Human recruiters still reviewed the interviews and made the hiring decisions. Applicants interviewed by the AI agents were 12 percent more likely to receive offers, with gains carrying into job starts and retention and no measured decline in the productivity of hires.

The useful finding is not simply that AI “won.” The researchers attribute the result to controlled variance: the agents were more structured and consistent while still responding to individual answers. That is a design lesson human interviewers can use without handing the relationship or the decision to a machine.

Fix the evidence, not every word

A structured interview should begin with a shared evidence plan. Before the first candidate arrives, choose the few capabilities the role genuinely requires and define what convincing evidence would look like.

For a customer success role, one capability might be recovering a strained account. Strong evidence could include diagnosing the source of friction, naming the tradeoff offered to the customer, coordinating internal owners, and measuring whether trust improved. “Great with customers” is not evidence. Neither is an interviewer’s feeling that the conversation flowed well.

Give every candidate the same core opportunity to demonstrate each capability. Use the same criteria and roughly the same amount of time. But do not confuse consistency with reading identical sentences in identical order. Two candidates can answer the same opening question at very different levels of detail. The interviewer’s job is to keep the target stable while choosing the next probe intelligently.

This distinction is especially important when candidates use AI to prepare. LinkedIn’s 2026 global research found that 81 percent of respondents had used or planned to use AI in their job search, and nearly half said it increased their interview confidence. A rehearsed opening answer is therefore a weak signal on its own. The goal is not to catch someone using a tool. Preparation is reasonable. The goal is to move from portable language to evidence rooted in a particular person’s experience.

Use a five-step follow-up ladder

When an answer sounds relevant but remains vague, climb the same ladder every time:

1. Claim: What are you saying you accomplished?
2. Scene: What was happening at the moment your contribution became necessary?
3. Choice: Which options did you consider, and why did you choose that one?
4. Consequence: What changed, and how did you know?
5. Reflection: What would you repeat or revise now?

The ladder creates comparability without turning the interview into an interrogation. You may start on a different rung depending on the answer. If a candidate already gives a specific scene, ask about the choice. If they name an impressive result, ask how it was measured and what else may have caused it.

Consider the polished launch story. A useful follow-up sequence might be: “What did engineering and sales disagree about?” Then, “What did you personally propose?” Then, “What did the team reject?” Finally, “What happened in the two weeks after launch?” Each question asks for observable detail. None assumes the candidate is exaggerating.

Probe the gap, not the person

Good follow-ups should sound curious, not prosecutorial. Interviewers sometimes become suspicious when an answer is fluent, then start testing the candidate’s credibility rather than collecting job-relevant evidence. That shift can introduce inconsistency and reward candidates who are comfortable sparring.

Name the information gap instead. “I understand the team result, but I do not yet understand your part.” Or, “You described the decision. Could you walk me through the evidence available at the time?” This keeps the conversation anchored to the criterion.

Also make room for honest limits. A candidate who says, “I supported the analysis, but my manager made the call,” may be giving more useful evidence than someone who absorbs the whole team’s achievement into a tidy first-person story. Attribution is itself a workplace skill.

For panel interviews, agree in advance on who owns each evidence area. Otherwise one interviewer may probe deeply while another repeats a surface question, and the group later mistakes airtime for coverage. A shared live recap can help the panel see what has been established and which criterion still lacks evidence.

Do not automate away the candidate experience

The case for structure is not a case for removing people from hiring.

A separate 2026 field experiment involving 3,296 real applicants found that one-way asynchronous interviews reduced continuation by about 45 percentage points, or 53 percent, compared with no interview screen. The researchers also found that the format deterred highly qualified applicants and was perceived as less fair and more competitive. The Monash University working paper studied a different setting from the live voice-agent experiment, which is precisely why the contrast matters.

Responsiveness changes the experience. A live interviewer can explain an ambiguous question, notice that the role was described unclearly, make space for the candidate’s questions, and repair a misunderstanding. A one-way prompt collects answers but cannot build much trust.

Teams should therefore decide separately where AI helps with consistency and where a person creates necessary reciprocity. Scheduling, note capture, evidence coverage, and suggested follow-ups are different jobs from evaluating someone’s potential or representing the company they may join.

Score the evidence before discussing the person

Structure can disappear as soon as the candidate leaves. One interviewer remembers an energetic opening. Another fixates on a weak final answer. A third has conducted four interviews that afternoon and can no longer recall which example belonged to whom.

Before the panel discussion, have each interviewer record evidence by criterion and score it independently. Separate observation from interpretation:

Observation: The candidate identified two causes of churn, changed the escalation cadence, and reported renewal improvement across six accounts.
Interpretation: Strong evidence of diagnosis and cross-functional follow-through.
Missing: No example of handling an executive-level conflict.

This format makes disagreement productive. The panel can debate the quality of evidence instead of trading adjectives such as “strategic,” “senior,” or “not quite a fit.” It also creates a clean question for the next round rather than letting uncertainty harden into a negative impression.

Let AI protect attention, not replace judgment

The best human interviewers are doing several things at once: listening closely, managing time, choosing a follow-up, taking accurate notes, checking coverage, and representing the company with care. Something will usually slip.

An AI copilot can reduce that load if its role is clear. It can maintain a live recap, surface a possible follow-up when an answer lacks evidence, help the interviewer check which criteria remain uncovered, and preserve accurate context for the debrief. The interviewer can stay present and decide whether the suggested question fits the moment.

That is a natural use for Caspi, which supports real-time recap, suggested questions, contextual chat, proactive flags from connected tools, post-call action items, and persistent meeting memory. The aim is not a robotic interview. It is a more attentive human conversation with a steadier evidence trail: structured where fairness needs consistency, adaptive where understanding needs curiosity.