Would you walk out of an AI interview? 11% say yes

Two things landed on 12 November: an agentic recruiting product that runs the hiring pipeline for weeks, and a UK survey showing one in nine workers would walk out of an AI-run first interview. Both concern me personally, because I sketched the InterviewBot years ago. Here is what I think my own design brief got wrong, the three conditions I would now insist on, and why the action log, not the model card, is the governance artefact worth auditing.

Two things landed on the same morning, 12 November, and they argue with each other. One is a product. The other is a verdict on the product's entire category.

The product is Eightfold AI's "AI Recruiting Teammate". As DHRMap reported, the agent "understands an organization's specific hiring process and autonomously drives work toward the hiring goal itself", from intake through sourcing, screening, scheduling and follow-up, working "across hours, days, and weeks". It keeps a reviewable action log, and its matching model was independently audited for bias by BABL AI Inc. under New York City's Local Law 144.

The verdict is a Careerminds UK survey of 1,000 full-time UK workers, released the same day and carried by The Global Recruiter:

• 69.1% are uneasy about a first interview run by AI rather than a person
• 33.4% would go ahead but feel uncomfortable
• 24.5% say it would depend on the role or the salary
• 11.2% would withdraw their application — 5.5% definitely, 5.7% probably

On a role with 500 applications, that is roughly 56 people gone before anyone has spoken to them.

I must declare an interest. In my National Employment Grid whitepaper (July 2025) I listed, as an attraction, that "InterviewBots conduct first-round interviews using voice and video", with percentile scores returned instantly. My ARDIS–ARGIS sketches were about reading resumes, not interviewing, but the same impulse ran through them: take the manual labour out of screening. I designed for the employer's queue and assumed the candidate would be grateful for the speed.

This survey says otherwise. Three conditions I would write into the spec today:

1. Disclose before the interview starts — in the invitation itself, not in a privacy page. Say that the interviewer is software, and what it will and will not judge.
2. Publish how answers are scored: which dimensions, what weighting, and whether voice or face is analysed at all.
3. Name the step at which a human enters. "A recruiter reviews shortlists" is not enough; say which stage, and keep to it.

Then the harder point. In When AI Becomes Your Boss (9 May 2026) I asked who owns the decision when an algorithm rejects a person, and noted that in many systems candidates "don't know why a score changed or why a person was filtered out". That problem has now been industrialised into an orchestration layer. When one agent sources, screens, schedules and follows up over weeks, no single step is the decision any more — which makes it harder to point at and easier to deny.

So the audit log becomes the real governance artefact, not the model card. A model card states what the vendor intended. A log states what happened to applicant 217 on a Tuesday morning. Only the second survives a tribunal, or a candidate's polite question.

A Local Law 144 bias audit is welcome and not sufficient: it audits the ranking model, not the orchestration around it. If I were buying this, I would ask for the log before the accuracy figure — and show it to the candidate who asks. Careerminds' own expert, Amanda Augustine, puts the remedy plainly: be upfront about when AI is used, how answers are assessed, and when a person will be there. I would only add a fourth item. Keep the receipts.

With regards,
Hemen Parekh
www.Iam-Immortal.ai / www.IndiaAGI.ai / www.YourContentCreator.in
13 November 2026