Strategy & Research

From30recordedcallstoastructuredproductbrief.Withoutanyonepressingplay.

The product team runs 30 customer calls. Today, those recordings sit in a folder until someone finds time to listen. This agent processes every call, extracts quotes by theme, maps what customers are asking for against the current roadmap — and calls out the gaps no one planned to solve. The output lands directly in the product management tool as tagged, prioritized notes. PMs open their existing workflow and the research is already there, structured and actionable.

Key Takeaways

Themes extracted, not just keywords

The automation identifies recurring patterns across calls — not just repeated words. A pain point mentioned seven different ways in seven calls becomes one theme.

Linked to source clips

Every insight in the brief is traceable to the specific call and timestamp where it was said. Product managers can verify and share the evidence.

Days, not weeks

Thirty calls that would take a researcher two weeks to code and synthesize are processed overnight. The brief is ready for the sprint planning meeting.

Feature signal quantified

Each feature request or pain point is tagged with frequency — how many calls mentioned it, in what context, and with what emotional weight.

Segment breakdowns included

Insights are broken down by customer segment, company size, or persona — whatever dimensions the research team defines — so prioritization decisions are grounded.

Listening to thirty calls is not research. Synthesizing them is.

Product teams run discovery calls to understand what customers actually struggle with. But the insight isn't in any single call — it's in the pattern that emerges across thirty of them. Extracting that pattern manually means listening to recordings, taking notes, coding themes, and then synthesizing across a spreadsheet that no one agrees on. That process takes weeks and still depends heavily on who did the coding.

The automation processes every recording simultaneously. It transcribes, extracts pain points and feature requests with their surrounding context, groups them into themes using semantic similarity rather than keyword matching, and quantifies frequency across the call set. Every theme in the brief is linked back to the specific timestamps where it appeared — so the evidence is verifiable, shareable, and usable in a roadmap discussion.

Product managers stop spending time on synthesis. They spend it on the brief itself — deciding what to build, not counting how many people asked for it.

One call for the full set

Pass the recordings folder and a research frame defining the themes and segments of interest. The automation transcribes every call, codes insights, clusters themes across the full set, and produces a product brief with frequency data and source-linked evidence.

python

import dspy
from predict_rlm import File, PredictRLM


class SynthesizeDiscoveryCalls(dspy.Signature):
    """Transcribe all call recordings, extract pain points and feature requests
    with surrounding context, identify themes across the full call set using
    semantic clustering, and produce a product brief with frequency data
    and timestamp-linked evidence."""

    recordings: list[File]   = dspy.InputField(desc="Discovery call recordings (audio or video)")
    research_frame: File     = dspy.InputField(desc="Research questions, segment definitions, and theme taxonomy")
    product_brief: File      = dspy.OutputField(desc="Structured product brief with themes, frequency counts, and source clips — Word")
    evidence_table: File     = dspy.OutputField(desc="All coded insights with call ID, timestamp, and theme tag — Excel")


agent = PredictRLM(SynthesizeDiscoveryCalls, lm="openai/gpt-5.4", sub_lm="openai/gpt-5.1", max_iterations=25)

result = agent(
    recordings=[File(path="recordings/")],
    research_frame=File(path="research/q2-discovery-frame.pdf"),
)
# result.product_brief → shared with product team
# result.evidence_table → linked from brief sections

Built on predict-rlm — open source. github.com/Trampoline-AI/predict-rlm

What the product team receives

A structured brief organized by theme, not by call. Each theme section includes a description, the frequency count across the call set, representative quotes with speaker context, and a breakdown by customer segment. Every quote links to the call timestamp where it was said. A second file — an Excel evidence table — contains every coded insight from every call, tagged by theme, call ID, and segment, for teams that need the raw data.