Recorded-Customer Digital Twin
Turn interview transcripts into a queryable voice-of-customer panel
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 6
- Confidence
- 96%
The Recorded-Customer Digital Twin converts first-party interview transcripts into a reusable voice-of-customer system. The team records and transcribes customer conversations, uploads them to a grounded AI workspace such as NotebookLM, and asks the model to role-play each customer using that person's documented statements as context. Teams can then test proposed courses, features, messages, or personalized outreach against those virtual customers. The transcript grounding makes the responses more specific than generic personas, especially for niche businesses with limited public data. Aggregated answers reveal common demand, while customer-level answers expose differences useful for personalization. The system reduces repeated interruptions and helps teams avoid obviously weak questions, but its conclusions remain predictions that should be confirmed with real people before consequential action.
Origin
On Marketing Against the Grain, Mike Taylor showed customer-interview transcripts in NotebookLM and used them to simulate how individual subscribers might react to proposed courses.
Core principles
- 01Capture customer conversations as durable first-party data
- 02Ground simulated responses in each customer's actual transcript
- 03Use digital twins to sense-check ideas before contacting customers again
- 04Combine proprietary context with the model's external knowledge
- 05Treat simulated answers as hypotheses rather than customer consent
How to run it
- 1
Capture customer conversations
Record interviews, sales calls, support conversations, or research sessions with appropriate permission. Preserve enough context to identify the customer's goals and constraints.
Pro tip Use a consistent recording and transcription workflow so the archive grows automatically.
Watch out Do not record or reuse conversations without appropriate consent and privacy controls.
- 2
Build the transcript library
Transcribe each conversation and add it to a grounded AI notebook or retrieval system. Keep sources separate so answers can be traced to individual customers.
Pro tip Label sources with customer type and interview date rather than relying on file order.
Watch out Poor transcripts can cause the model to attribute views incorrectly.
- 3
Define the decision
Ask a narrow question about a proposed product, course, feature, message, or offer. Specify whether the model should answer for each customer or for the group.
Pro tip Compare two concrete options when possible.
Watch out Do not ask the system to infer sensitive traits that customers did not disclose.
- 4
Role-play each customer
Instruct the model to answer as each customer using the corresponding transcript as evidence. Require reasoning connected to the customer's recorded needs.
Pro tip Ask the system to distinguish direct evidence from inference.
Watch out A simulated customer can still hallucinate preferences absent from the transcript.
- 5
Aggregate and segment insights
Identify recurring preferences, objections, vocabulary, and differences across customer types. Use these patterns to shape the next experiment or conversation.
Pro tip Preserve outliers when they represent strategically important customer groups.
Watch out Aggregation can hide conflicting needs behind a false consensus.
- 6
Sense-check and validate
Use the virtual panel to reject weak ideas and refine promising ones before returning to real customers. Confirm important predictions through genuine interviews or observed behavior.
Pro tip Bring customers a stronger, more specific proposal after the simulated pass.
Watch out Never represent a digital twin's answer as a statement the real customer actually made.
In the wild
Taylor uploaded transcripts from customers of his marketing education business and asked NotebookLM to role-play them. He proposed possible courses such as marketing mix modeling and customer retention, then used the transcript-grounded reactions to estimate which subject would resonate before bothering customers with another interview.
→ The existing interview archive became a reusable source of directional product feedback.
After identifying interest in a marketing mix modeling course, Taylor asked the system to write an email to a specific customer using that customer's transcript. He treated the output primarily as insight into the appropriate angle rather than blindly sending the generated copy.
→ The customer context supplied a personalized positioning angle for human-written outreach.
Common mistakes
Failing to record conversations
Uncaptured calls cannot contribute to future analysis, content, or product research. Establish recording and transcription as part of the customer workflow.
Sending generated outreach unchanged
Transcript grounding can reveal a useful angle, but generated copy may be too long or misstate the customer. Review and write the final message deliberately.
Confusing prediction with permission
A simulated answer is not consent or a binding statement from the customer. Validate consequential decisions with the actual person.
Is it for you?
Best for
It is best for teams with recorded interviews, calls, or support conversations that need repeated voice-of-customer input.
Not ideal for
It is not ideal when recordings lack consent, transcripts are unreliable, or the decision requires a customer's current explicit approval.
From the transcript
“the deeper uh the Persona information you give it uh then the better the more accurate the response is”
“role play as each of my customers uh and tell me would they be happy to see a new course on um say like marketing…”
“I can also just ask those people questions without having to bother them again”
From the episode
This AI Prompt Gets You Customer Insights in 5 Minutes (Free Tool)