---
title: "What AI-run interviews do to an agency's margins"
description: "Interview hours cap what a research agency can bill. See how AI-run, disciplined customer interviews change discovery pricing, scoping, and margin."
url: "https://luc.so/articles/ai-customer-research-agency-economics/"
category: "Agency economics"
date: "2026-06-29T19:05:52.900387"
source: luc.so
---

# What AI-run interviews do to an agency's margins

**TL;DR.** For a research, UX, or brand agency, interview hours are the cost of goods, and they cap how much discovery you can sell. Traditional one-on-one interviews run roughly $800 to $1,500 per respondent, so a 20-interview project lands near $15,000 to $30,000, most of it moderation and recruiting labor. When an AI interviewer runs the volume under enforced craft, the bottleneck moves off your researchers' calendars. That changes four things: what you can scope, what you can price, how fast you deliver, and how much of the work is judgment instead of repetition. The catch is the same one that governs all of this: volume only pays off if the discipline holds.

## Why interview hours cap agency margin

Run the math on a single discovery project and the constraint shows up fast. A senior researcher can moderate maybe three or four good interviews in a day before fatigue starts leading the witness. Recruiting, scheduling, no-shows, and write-up eat the rest. So the interview count on any project is set by how many human hours you can bill into it, not by how much evidence the question actually needs.

That ceiling forces three bad trades, and every agency principal recognizes at least one of them. You under-scope: you sell eight interviews when the decision wanted twenty, then hedge the readout. You hand the load to whoever is free, so a junior moderator runs the conversations that a senior should have. Or you quietly swap a survey in for interviews and call it research, which trades the one thing qualitative work is for. None of these are laziness. They are what a labor-bound cost base does to a project plan.

The deeper problem sits underneath the count. Discipline degrades at volume. A tired interviewer starts paraphrasing answers back, planting the answer inside the question, and accepting a compliment as a finding. The eleventh interview of the week is rarely as clean as the first. So scaling the human way costs you twice: more hours, and softer data per hour.

## What changes when the interview runs itself

When an AI interviewer carries the moderation load, the count stops being a function of your team's calendar. You can run the volume a project needs in parallel, overnight, across a wider sample, without asking a researcher to sit through the fortieth session of the week.

The part that matters for an agency is not raw speed. It is that the craft stays constant. A disciplined AI interviewer holds the same rules on every conversation: it asks what someone did, not what they would do; it follows the last real project instead of a hypothetical; it never feeds the answer inside the question; it treats enthusiasm as noise until a behavior backs it; and it follows the hesitation instead of the script. That consistency is exactly what a tired human loses at volume. For an agency, [discipline is the product](/manifesto), and an instrument that enforces it on conversation fifty as firmly as conversation one is what makes scaling safe rather than reckless.

Your researchers do not disappear from the work. Their hours move to where judgment actually pays: writing the discussion guide, reading the transcripts for the jobs and forces underneath them, deciding when the data has saturated, and presenting the verbatims that move a stakeholder. The repetitive moderation comes off their plate. The thinking stays on it.

One honest line here, because the category is noisy about it. AI-run interviews mean a real person answering a machine-moderated conversation about their own behavior. That is not the same as a model imagining what a customer might say. [A simulated answer is a hypothetical with a UI](/articles/synthetic-users-vs-real-interviews/), the precise bad data the Mom Test throws out. The margin story in this piece depends on real respondents. Synthetic ones do not count.

## Repricing discovery: cost versus what you bill

Here is the move most agencies miss when a labor constraint lifts. Your discovery price was never really the sum of your interview hours. It was the value of the decision the research de-risks: the positioning bet, the feature cut, the campaign spend the client is about to commit. Hours were just how you backed into a number you could defend.

So when moderation stops being the cost driver, you have a choice, and price-cutting is the worst of the available options. You can hold the value-based price while your cost base drops, which widens margin on every qualitative project you already win. You can widen the sample at the same price, so the evidence is deeper and harder for a stakeholder to dismiss, which makes the deliverable more defensible and the next sale easier. Or you can take on the larger discovery phases and enterprise samples you used to decline because the interview load would have burned out the team.

Most agencies will blend these. The point is that the conversation with the client stays anchored on the outcome, not the unit. If you reprice down to your new cost, you teach the buyer that research is a commodity priced by the interview. Keep selling the decision.

## Scoping a disciplined discovery sprint

With the moderation ceiling gone, you can scope a sprint around the question instead of around your team's availability. A workable shape looks like this.

Start with the decision. Write down what the client will do differently depending on what you find. That single sentence sets the sample and the guide.

Set the sample by saturation, not by budget math. The research is fairly settled here: meaning saturation tends to land around 16 to 24 interviews for a varied sample, and closer to 12 for a homogeneous one. Use that to size the run, and read [how many interviews a project actually needs](/articles/how-many-customer-interviews-are-enough/) before you commit a number, because the count is a trap when the craft underneath it is loose.

Build one rigorous discussion guide. This is where good research is won or lost, and it stays human work. The guide encodes the jobs you are hunting and the switch moments you want respondents to walk you through.

Run the volume in parallel under enforced craft, then put your senior people on synthesis: coding the jobs, the forces of progress, the pains, and pulling the verbatims that will land in the readout. The sprint compresses because the slow, serial part got parallel, not because anyone cut a corner.

## Selling AI interviews to skeptical clients

Some clients will flinch at the words "AI interview." Two worries sit under that flinch, and both have straight answers.

The first is depth: will a machine-moderated conversation go off-script when something interesting surfaces? The honest answer is that an undisciplined AI interviewer will not, and a disciplined one will, because following the hesitation is the whole design. Show them a transcript. Let the client read a real probe chasing a contradiction the respondent did not expect to reveal. A transcript settles this faster than any deck.

The second worry is rigor: is this softer than a human moderator? Reframe it. A human moderator's discipline varies with fatigue, mood, and seniority. An enforced-craft interviewer applies the same non-leading standard to every conversation, which makes the method more consistent, not less. You are selling repeatable rigor at a sample size a human team could not have reached. Name what the tool refuses to do, in plain terms: it will not accept a compliment as evidence, it will not plant the answer in the question, and it will not let a vague answer slide.

Sell the deliverable, not the machinery. The client is buying jobs, forces, pains, and verbatims they can act on. How the volume got moderated is your production detail, the same way they never asked which transcription tool you used.

## Delivering evidence stakeholders act on

The output that earns the next project is not a stack of transcripts. It is structured evidence a stakeholder can carry into a room and defend. That means the jobs the customer is hiring for, the forces pushing them toward and away from a switch, the pains in their own words, and the verbatims that make a finding impossible to wave off.

This is the part agencies often eat as unbilled hours: pulling the quotes that prove the value, formatting the readout, building the throughline from raw conversation to recommendation. When the moderation load comes off your team, those hours are available for the synthesis that actually persuades. The deliverable gets sharper at the exact moment it gets cheaper to produce. That is the margin story stated plainly. Lower cost of goods, higher-value output, and a sample size that makes the conclusion hard to argue with.

## Frequently asked questions

### How can a research agency run customer interviews at volume without losing quality?

Move the moderation load off tired humans and onto an interviewer that enforces the same craft on every conversation: ask about past behavior, never plant the answer in the question, treat compliments as noise, and probe the hesitation. Quality drops at volume when discipline depends on who is in the room and how many sessions they have already run that day. An AI interviewer holds the rules on conversation one and conversation fifty alike, and your researchers spend their hours on synthesis and the few conversations that decide something.

### How do you price a customer-research project for an agency client?

Price the decision the research informs, not the hours it takes. Traditional one-on-one interviews run roughly $800 to $1,500 per respondent, so a 20-interview project lands near $15,000 to $30,000, most of it moderation and recruiting labor. When moderation stops being the constraint, you can hold the value-based price while the cost base falls, or widen the sample so the evidence is harder for a stakeholder to wave away. Either way, anchor the conversation on the deliverable and the risk it removes.

### How do agencies scale qualitative research?

Most agencies scale qualitative work by adding contractors, narrowing scope, or substituting a survey for interviews, and all three cost either margin or rigor. The durable path is to scale the moderation, not the headcount: run the volume a project needs in parallel under enforced interview craft, then keep human judgment on synthesis and saturation. Volume only helps if the discipline holds. Faster bad interviews are still bad interviews.

## Where this is headed

Lùc is in closed beta, built for exactly this buyer: the agency principal or research lead who can feel interview hours capping the work they could otherwise win. The vision we are building toward is simple. Run the volume a project needs, hold every conversation to interview craft, and give your researchers their hours back for the thinking that clients pay for.

If interview load is the thing forcing you to under-scope or turn work down, that is the constraint we want to remove. Join the [private beta waitlist](/) and we will get you in as access opens.