Employees Aren't Short of Opinions, They're Afraid to Speak: How to Design Anonymous AI Interviews on a Corporate Intranet

What Comes Back from an Anonymous Survey Is Mostly the Safe Answer

Any HR lead who has run an internal survey knows the feeling: the response rate is decent, the scores are acceptable, and reading the results tells you nothing. The complaints cluster around benefits and parking spaces. The real process problems, the friction between departments, the thing everyone knows about but nobody writes down: not a word of it appears on the form.

The reason is not complicated. The form says "anonymous," but employees do the practical arithmetic: the handwriting is gone, the phrasing habits are not, and cross-tabulating two fields, department and years of service, makes it obvious who wrote what inside a small unit. Our own product documentation puts it in one line: pseudonymous is not anonymous. Taking the name off does not mean nobody can trace it back, employees understand that better than anyone, and so what they hand in is the version that will not cause trouble.

A verbal promise will not get you the truth. The protection has to be system behavior: even if a manager wants the transcript, the system has no path that hands it over. Only at that point do employees have a reason to speak.

How an AI Interview Differs from a Survey

A survey is limited by questions fixed in advance. A question the designer never thought of will never get an answer, and what an employee actually wants to raise is usually not among the options.

An AI interview inverts that: one question at a time, following wherever the answer leads. Our interview engine works through four layers. It starts with what the person says on the surface, asks how it felt at the time, then presses for a concrete scene, and only then reaches the root cause. It also avoids broad questions like "what are the pain points in your current process." Abstract questions produce abstract answers, so the engine follows one specific thing: the last form the person handled, the last approval that stalled.

An interview runs 12 to 15 minutes over roughly 14 to 22 turns, one question per turn, never multiple choice. The interviewee decides when it ends. The AI confirms only when the person clearly wants to wrap up, and never signs off on its own. Within 48 hours of the interview, the system produces two layers of output: a structured analysis of each session, with every conclusion marked with the part of the conversation it came from, and a draft action plan consolidated across interviewees that management can take straight into a meeting.

Dimension	Anonymous survey	AI interview
Questions	Fixed in advance	Follow-ups driven by the answers
Depth	Options and rating scales	Four layers down to the root cause
Time	A few minutes to fill in	12–15 minute conversation
Anonymity protection	A verbal promise	Access tiers and stripping built into the system
Output	Scores and pie charts	Sourced analysis plus a draft action plan

"No Penalty for Telling the Truth" Has to Live in the System

Our engine builds interviewee protection into three designs that cannot be worked around. The first is tiered access: a corporate account receives the analysis only, while transcripts and verbatim quotes are stripped out on the back end. It is not an agreement not to look, it is an inability to produce them. The second is a minimum group size: any breakdown with fewer respondents than the threshold (currently set at three) is not displayed, and the threshold can only be moved in the stricter direction, because interviewees spoke on the promise of that number. The third is written into the contract: the company may not compel employees to take part, and may not factor participation into performance reviews or shift scheduling.

Two things we choose to state before the interview begins. First, analysts on the platform side can see the full conversation. The goal of anonymity is specific: what you say will not come back to your manager. The fact that someone on the platform side reads it for analysis is stated plainly. Second, in the cloud edition the AI analysis is processed through an external AI service, that processing may take place outside Taiwan, and the interviewee consent flow says so explicitly.

Want to see what an interview like this surfaces? Book a demo and we will walk through the interview flow and the report format with you.

Book a Demo

Intranet Deployment: Data Never Leaves the Company

Systems like this used to come in one form only, the cloud, which made sending data out of the company a precondition for adoption, and many evaluations stopped at that information security question. On-premises AI models have matured quickly over the past year or two, and running a full interview and its analysis inside a company data center is now a practical option.

The design principle of the on-premises edition we are planning is that transcripts, analysis results and the data accumulated over time all stay inside the company environment, never passing through an external cloud, with access governed by the company information security policy. The plan reuses the same engine and the same protections as the cloud edition: tiered access, the minimum group size and the interviewee right to end the session do not change with the deployment model. For industries sensitive to data leaving the building, such as finance and manufacturing, this turns the information security review from a wall into a box that can be ticked.

Accumulating Over Time: From an Annual Survey to Talent Signals

A one-off survey answers the question of where it hurts right now. Accumulated interviews answer a different class of question: which themes come back every quarter, which units show shifting signals, and who has put forward a concrete, workable idea in an interview.

That last item is the direction we are developing. Once enough interview data has built up, it can support matching people into innovation task forces, along the lines of a management trainee program, with the difference that the candidates are not circled by their managers. They surface from the people who genuinely raised ideas, interview after interview. This is a planned capability, not a feature we deliver today. But it explains why interviews are worth running as a standing practice rather than a once-a-year event.

What to Settle Before You Roll This Out

This is not a monitoring tool. There is no live per-person dashboard, and it will not tell you who said what. If that is what you want, the access design of this system cannot give it to you.
The invite list has to cover everyone who touches the process, at every level. Interviewing only the representatives a manager picks puts you straight back into the old survey problem.
There should be an internal letter signed by senior leadership stating that participation is voluntary and that the results do not feed performance reviews. The interview is exactly as credible as that letter.
Leave the link open for about a week so people can pick a time that suits them. An interview should not feel like being called into a meeting room.

Frequently Asked Questions

Can my manager really not see what I said?

The analysis a corporate account receives contains no transcript, and verbatim quotes are stripped on the back end; any breakdown below the minimum group size is not shown at all. Analysts on the platform side can see the full conversation, and that is stated before the interview starts. If a company cannot accept data passing through an external platform at all, on-premises deployment is the form we are developing, and we can discuss feasibility during the evaluation.

Will employees just go through the motions with an AI?

Yes, and the first few turns are often safe answers. That is why the engine avoids broad questions and follows specific documents and incidents; once it reaches the level of a concrete scene, going through the motions is hard to sustain. Anyone who really does not want to talk can end the session at any time, and that decision belongs to the interviewee. What comes from the people who volunteer twenty turns is far more useful than a survey everyone completed by marking 4 on every item.