Why we let AI run the interviews (and why we never let it pretend to be human)

Ask any consultant how long it takes to interview twelve executives at a mid-sized business, and the honest answer is never “a day.” It’s three weeks of calendar tennis, a handful of reschedules, one person who “will send notes instead,” and a final report built on whoever actually showed up. The interviews themselves might take six hours combined. Getting them booked, run and written up eats a month.

We’ve been testing a chunk of that discovery work through AI-conducted interviews instead of a round of consultant meetings, on our own operation first. The results have been good enough, often enough, that the format is now part of our discovery offer, and it’s shaping up as one of two related offerings, at opposite ends of an engagement, built on the same underlying discipline.

The bottleneck was never the questions

Nobody struggles to write a good interview guide. The bottleneck is logistics: fitting a structured conversation into the diary of someone who is already the busiest person in the building, twelve times over, without the findings going stale between the first interview and the last.

AI-conducted interviews remove the logistics problem almost entirely. Twelve conversations that would take three weeks to schedule around executive calendars can run in days, because the interviewer doesn’t need a diary slot; it needs whenever the participant has twenty minutes. That alone changes what’s possible: a business can get a genuinely large-scale discovery exercise done inside a sprint, not a quarter.

Speed isn’t the only thing that changes. Every participant gets the same structure, the same core questions, asked with the same framing. A dozen human interviewers, however skilled, drift: different follow-ups, different emphasis, different read of when to push and when to let something go. An AI running the same guide across every conversation produces something a spreadsheet of human interview notes rarely does: findings that are actually comparable to each other, not just twelve separate impressions loosely stitched into a report.

Ego-less data extraction

There’s a third effect, and it’s the one that surprised us most. We call it ego-less data extraction, and it’s not a euphemism.

Ask a staff member what’s actually broken in their workflow, in front of a consultant they’ve just met, and you get a filtered answer: filtered by what makes them look competent, what won’t land a colleague in trouble, what won’t contradict what their manager said in the room next door twenty minutes earlier. None of that is dishonesty. It’s ordinary workplace self-protection, and it’s completely normal. It’s also exactly the material a discovery process most needs, and least often gets.

An AI asking the same question isn’t reading the room, doesn’t remember who’s popular, and isn’t going to mention the answer at the next BNI meeting. That absence of social stake seems to lower the guard in a way a human interviewer, purely by being a person with a reputation and relationships of their own, structurally cannot. The awkward questions get asked, and answered, without the politics, the defensiveness, or the calendar drain of a round of consultant meetings.

To be clear about what we’re claiming and what we’re not: this is what we’ve observed testing the format, not a track record of client engagements and not a settled finding from the literature. The research on AI-conducted interviews and disclosure effects is still thin and still catching up. We’d rather say that plainly than borrow authority from a body of evidence that doesn’t fully exist yet.

The technology caught up

A year or two ago, “let the AI run the interview” was a harder pitch, because the AI would occasionally invent an answer, misfire on a follow-up, or simply stall. That’s no longer the normal experience. Delays and hallucinations in a well-scoped, well-grounded interview flow are now rare rather than routine, rare enough that we’re comfortable putting it in front of a client’s staff without a human minding every exchange.

Rare isn’t zero. We still design the flow with guardrails: a fixed question set, defined boundaries on what the interviewer will and won’t speculate about, and a human review pass on the compiled output before it becomes anyone’s deliverable. The reliability improvement is real, but it’s a reason to trust the tool more, not a reason to stop checking its work.

We never pretend it’s human

This is a line we don’t cross, and we don’t want to be vague about it: every AI-conducted interview tells the participant, plainly, at the start, that they’re talking to an AI. No persona, no name that implies a person, no pretending. It’s a transactional exchange, and we say so. It’s the same rule we apply to every assistant we run, and what happens when a business breaks it is well documented.

Some people assume that honesty costs you depth, that participants will hold back once they know there’s no human on the other end to build rapport with. Our testing runs the other way. The transactional framing doesn’t compromise depth; if anything, it’s part of what removes the ego complications in the first place. A bot with no reputation to protect and no relationship to manage can go straight at the area that needs challenging, without the softening a skilled human interviewer might apply out of instinct or courtesy. That precision, asking the pointed question cleanly without the social hedging humans apply almost unconsciously, is shaping up as an advantage, not a compromise.

Two products, one discipline

The same instrument shows up twice in how we structure engagements, at opposite ends.

At the front, it’s Discovery: structured interviews, AI-conducted and in person, alongside focus groups and surveys, used at the Assess stage to build the picture a strategy gets built on. This is the large-scale, fast-turnaround version: the one that used to take weeks of executive calendars and now takes days.

At the back, it’s a Retain-stage instrument: the same interview and survey capability run again after a build has shipped, measured against the original baseline. Did the thing we built actually change what people reported the first time round? A post-build survey and a follow-up focus group answer that with the same rigour as the original discovery, not a satisfaction-score afterthought.

Same underlying capability, two very different jobs: one opens an engagement, the other proves what the engagement delivered.

Where we’re being careful

A few honest caveats, because we’d rather state them than have a client discover them.

This is new for us, and we say so. We don’t claim a client track record on the AI-conducted interview format. It’s a live instrument we’ve proven on our own operation, not a decade-proven one, and our public materials say that plainly rather than imply otherwise.

The disclosure standard has to keep pace with the rules, not just our comfort level. As regulation around automated decision-making and AI disclosure continues to develop, we’re treating that as ongoing work, not a box ticked once. What “plainly told they’re talking to an AI” needs to look like in practice is something we expect to keep tightening, not something we consider settled.

Pricing is scoped after a conversation, not fixed on a rate card. Some of the surrounding instruments (in-person interviews, focus groups) are delivered through third-party facilitators, and survey costs depend on respondent count and delivery method. We’d rather scope it properly than publish a number that doesn’t hold up across a genuinely varied set of engagements.

Our confidence is observation, not literature. Everything above about depth, honesty and reduced defensiveness is what we’ve seen testing this on our own work, not a citation. We think it’s real. We’re also watching for research that might complicate it, and we’ll say so if it does.

None of that is a reason to hold off using it. It’s the reason we can recommend it without overselling it.

Perth AI Consulting runs AI-conducted interviews as part of Discovery, and again at Retain to measure what a build actually changed. Start with a conversation.