The customer lens is what makes a CDD different from a well-researched equity report, and it's the one most often done badly under time pressure. Voice-of-Customer (VoC) means structured interviews and, where volume allows, surveys with the target's actual customers, former customers, and sometimes non-customers who evaluated and rejected the target. The seller's data room can't manufacture this kind of primary evidence, because the seller doesn't control who you call or what they say. Expert calls corroborate and contextualize what customers say. You might talk to former employees, competitors' staff, industry specialists, regulators, or channel partners. Expert input is triangulation, never a substitute for the customer evidence. Give a single expert's opinion the weight you'd give a single data point. It's useful, but it's never sufficient.
How do you design a voice of customer sample?
Sampling design comes first, and it's a hypothesis-driven choice, not a headcount. Before you dial a single number, decide which segments could tell different stories. Think of customers acquired recently versus long ago, customers in the segment the thesis leans on hardest versus the rest, satisfied customers versus those who churned, and customers reached through the seller's introduction versus those you reached yourself. The seller will always offer to "help set up some calls". Accept a handful for calibration, then insist on building an independent list from the CRM export or public records. A sample built entirely from management's referrals is a sample built to make management look good.
A workable rule of thumb is enough interviews per segment to see a pattern repeat. Roughly 8 to 15 per segment gets you qualitative saturation, and you'll want more if a survey is layered on top for a statistical read. Spread them across every segment a hypothesis depends on, and never bunch them in the easiest-to-reach group.
Advertisement
How do you write the discussion guide?
The discussion guide is where diligence teams do the most damage to their own evidence, entirely by accident. Take a leading question such as "You've been happy with the service since the ownership change, right?" It invites agreement and manufactures the answer the interviewer expected to hear. Social-desirability bias is a well-known human tendency. Faced with a stranger asking a leading question, most interviewees will simply agree to avoid creating friction. The fix is to anchor questions to recalled behavior instead of soliciting an opinion. Ask "Walk me through your last three purchases from this category, which supplier, and why that one, each time?" and you get a revealed-preference account. It's far harder to bias, and far more useful for a market model than a satisfaction score. Build the guide behavior first, opinion second, and always leave room for the interviewee to volunteer something you didn't think to ask. In a real diligence, the most valuable finding is almost always something nobody on the team had written into the guide.
Which biases affect voice of customer research?
Defend against four biases by name, because each one has a specific, repeatable fix:
- Selection bias. The sample doesn't represent the population you're trying to characterize. You interview only large accounts, say, when churn is concentrated among small ones. Fix it by stratifying the sample deliberately across the segments that matter, including the ones that are hard to reach.
- Sponsor or referral bias. The seller, or the target's own management, hand-picks who you talk to. Fix it by building at least half the sample independently, from a CRM export, a public customer list, or cold outreach, and then comparing the two halves' answers explicitly. A large gap between the referred sample and the independent sample is itself a finding.
- Small-n overconfidence. You draw a firm conclusion from four or five interviews that happened to agree. Fix it by treating any finding from under roughly eight to ten interviews as a hypothesis to corroborate further before you write it down as a conclusion.
- Social-desirability bias. Interviewees soften criticism because they're talking to a stranger, or because they suspect (correctly) that the call is deal-related and don't want to jeopardize a relationship. Behavior-anchored questions fix part of it. You fix the rest by never revealing which side of a deal is paying for the call, unless professional norms require it.
What are the ethics of expert calls?
Expert-network ethics is not a footnote. An expert on a paid call is bound by whatever confidentiality obligations they carry to their current employer. Your job is to extract judgment and market context, and never to solicit a breach of that confidentiality. Asking a competitor's current employee to disclose their employer's unpublished roadmap, pricing strategy, or specific client list is unethical. In most jurisdictions, and under most reputable expert-network platforms' compliance terms, it's also against the rules of engagement. A well-run call stays in the zone of general market knowledge, publicly observable behavior, and the expert's own past experience. A former employee discussing their former employer, within the bounds of any exit agreement, is on different ground from a current employee discussing a current employer's confidential plans. The moment a question drifts across that line, redirect it, right there on the call.
{
"skill": "Synthesizing raw VoC/expert-call transcripts and notes into coded themes without losing the evidence trail",
"use": "Feed anonymized interview notes or transcripts to an AI assistant to get a first-pass thematic coding — recurring topics, sentiment, contradictions — across many interviews faster than a first human pass, never as the final synthesis.",
"tools": ["Chat assistants — Claude, ChatGPT", "Document/notebook AI — a notebook tool grounded on your own transcript files"],
"prompt": "Here are N anonymized interview notes from a commercial due diligence on <TARGET>. Identify the 4-6 recurring themes, flag any theme mentioned by only one interviewee (weak evidence), quote the exact supporting lines for each theme, and flag any two interviews that directly contradict each other.",
"verify": "Re-read every quoted line against the original transcript yourself before it enters the report — a summarizer can merge two different interviewees' points into one false consensus, or produce a supporting quote that sounds plausible but was never actually said. Trace every theme back to a named interviewee and transcript line (Guardrail #1, Trace) before it appears in a deliverable; never let AI-coded themes substitute for your own read of the raw notes.",
"diy": "You must be able to code interview themes by hand first — reading transcripts, tagging repeated ideas, spotting contradictions is the actual analytical skill this section teaches. AI accelerates the first pass across volume; it never replaces your judgment on what counts as a theme.",
"market": "US"
}