The Brief Is the Product. The Pipeline Is How We Build It.
The previous capability-led partner post in this series walked through what's inside a single ICP Intelligence Brief. This one walks through how it gets built — the five-stage pipeline that turns a long-form executive interview into archetype-level structured data we can deliver by API.
The reason the methodology matters is that data licensing is a trust transaction. A buyer evaluating whether to integrate our briefs into their ICP scoring system, message generator, or AI sales training tool is implicitly trusting that the data behind those briefs was produced rigorously. The structured payload is what they consume. The pipeline behind the payload is what they're underwriting.
This post is the pipeline.
Go deeper: The GTM Enrichment partner page covers deployment patterns, integration options, and how the briefs plug into AI products.
Stage 1: Ingestion
The pipeline begins with audio. We ingest long-form executive interviews — podcasts, video conversations, webinar recordings, recorded panels — where senior leaders speak at length about their work. Interviews under twenty minutes are excluded. The structural argument from our methodology post earlier this week applies here: short-form material doesn't produce the specificity that makes briefs useful.
For each ingestion source, we capture the audio file, the source URL, the publication date, the speaker identities where available, the role and industry signal where the source provides it, and the basic metadata that lets us later associate the interview with a specific archetype.
Sources are selected for both volume and quality. We don't ingest every podcast that interviews CFOs. We ingest podcasts that interview CFOs about CFO work — distinct from podcasts that interview CFOs about general leadership or personal-development topics. The first kind produces working-vocabulary material. The second kind produces aspirational-vocabulary material. Briefs need the first kind.
The ingestion stage is also where we filter out non-interview content — vendor-sponsored episodes that are functionally advertisements, panels where the dialog is too crowded to extract individual voice patterns, and content that fails initial quality filters. The output of stage one is a corpus of long-form interview audio matched to source-level metadata.
Stage 2: Transcription and Speaker Segmentation
The audio gets transcribed using current speech-to-text systems calibrated for long-form business dialogue. The output is a time-aligned transcript that preserves speaker turns and natural disfluencies.
The fact that the transcript preserves disfluencies matters. When a CFO interrupts themselves to revise a thought ("we were planning to expand into — well, we were considering expanding into Europe before we changed direction"), the revision is signal. The interview is showing how the leader actually thinks, and the model behind the brief shouldn't smooth that signal out.
Speaker segmentation runs alongside transcription. The output is a transcript where each leader's turns are labeled and where multiple speakers in a single interview are distinguished. For multi-speaker conversations, we then identify which speaker matches the target archetype and run subsequent stages only on that speaker's content.
The output of stage two is a clean, speaker-segmented transcript with timestamps and metadata.
Stage 3: Role and Industry Classification
Every interview gets classified into a role-and-industry pair. Role classification draws on the speaker's self-description, the interviewer's framing, the speaker's described responsibilities, the company they're affiliated with, and the language patterns in their answers. Industry classification draws on the company's primary industry, the topics the conversation covers, and the working vocabulary the speaker uses.
Both classifications run through a structured taxonomy. Roles include CEO & Founder, CFO, CIO, CRO, CMO, CISO, COO, President, Board Member, Chief People Officer, Advisor & Consultant, VP-level functions, and adjacent categories. Industries include Tech/SaaS, AI/SaaS, Cybersecurity, Health Tech, Healthcare Services, Health Systems & Providers, Manufacturing, Logistics, Financial Services, FinTech, Consulting, Professional Services, Food & Hospitality, Retail & Consumer, Media & Entertainment, Venture Capital & PE, and the rest of the working taxonomy.
Classification confidence is logged alongside the classification itself. Interviews where the role or industry classification is below confidence threshold get flagged for review. They're not silently included in archetype briefs; they're either excluded or re-classified before being used.
The output of stage three is a transcript labeled with role, industry, company-stage indicators, and classification confidence.
Stage 4: Phase 1 Extraction — The Structured Layer
This is the analytical core of the pipeline. Each classified transcript runs through what we call Phase 1 extraction — a structured analysis that produces four parallel data outputs:
Factor scoring. Every interview is scored on the seven behavioral dimensions on a 1–5 scale: Narrative orientation, Operational philosophy, Data philosophy, Technology orientation, Risk calibration, Growth orientation, Stakeholder orientation. The scoring uses defined rubrics for what each dimension is measuring, anchored on representative examples from prior interviews. The output is seven scores per interview with confidence indicators.
Thematic extraction. The transcript is parsed for the speaker's expressed priorities, pain points, success metrics, decision frameworks, and red flags. These aren't extracted as keywords — they're extracted as bounded statements with associated context. "Reduce unnecessary variation in patient care" is a priority statement, not a keyword match on the word "variation." The output is a structured set of statements per category per interview.
Linguistic extraction. The transcript is parsed for the speaker's power words (vocabulary they reach for as emphasis), negative words (vocabulary they use to describe failure or rejection), jargon (technical or industry-specific terms they use without explanation), and stories or analogies (extended metaphors and reference frames they deploy). Each item is captured with the surrounding context so later analysis can verify how the speaker actually used the term.
Buyer journey extraction. The transcript is parsed for buying signals, evaluation criteria for tools and partners, and leadership style indicators. This is the section that's most variable in density — some interviews surface dense buyer-journey content, others barely touch it. The extraction captures what's there and marks confidence accordingly.
All four extraction outputs are stored as structured JSON associated with the transcript, the speaker, the role, and the industry classification.
Stage 5: Archetype Assembly
The final stage rolls individual-interview extractions up to archetype-level briefs.
An archetype is the intersection of role × industry × company stage. CRO at Tech/SaaS at growth stage is one archetype. CFO at Health Systems at enterprise scale is another. For each archetype, the pipeline aggregates the Phase 1 extractions from every interview matching that archetype.
Aggregation does several specific things:
Factor scores are averaged with confidence intervals. The CRO at Tech/SaaS archetype has 134 interviews in our corpus contributing to its factor scores. The brief shows the mean and the distribution. Where individual factor scores vary widely within the archetype, the brief notes the spread.
Vocabulary items are counted, weighted, and ranked. Power words appear in the brief with prevalence counts — not as a binary "this archetype uses this word" but as "this archetype reaches for this word in X% of interviews." The ranking lets downstream consumers weight the vocabulary by importance.
Thematic statements are clustered and deduplicated. Where multiple interviews surface the same priority in different phrasing, the cluster is unified. Where priorities recur in identical phrasing across multiple interviews, the brief notes the recurrence count.
Buyer-journey content is surfaced where dense and flagged where sparse. Some archetypes have rich buyer-journey content; others have thinner. The brief surfaces what's available and is transparent about density.
Time-bounded refreshes are tracked. Each brief is regenerated on monthly cadence as new interviews enter the corpus. The brief history is preserved so consumers can see how the archetype's vocabulary, factor scores, and priorities have drifted over time.
The output of stage five is the structured brief we deliver by API.
What Quality Means in This Pipeline
Several specific quality controls run alongside the five stages.
Source diversity. No single podcast or interview series dominates the corpus for any archetype. If one source contributes more than a threshold percentage of interviews for an archetype, the brief weights the contribution down to prevent any single editorial voice from dominating the structured data.
Time recency. Each brief preserves the rolling window of interviews it draws from. Older interviews are included but weighted lower. The brief is responsive to recent vocabulary shifts without being whiplashed by one quarter of unusual interviews.
Confidence transparency. Every brief surfaces confidence indicators at the factor level, the thematic level, the vocabulary level, and the buyer-journey level. Consumers know what they're getting and where the brief is confident versus where it's flagging lower-density material.
Classification audits. Random samples of classified transcripts are re-classified by independent runs to verify classification stability. Where stability is below threshold, the underlying classification logic is reviewed.
Brief-level review. New archetypes don't enter the API until their underlying briefs have been reviewed for face validity — meaning a human can read the brief and confirm it describes a coherent buyer the team can recognize.
What the Methodology Doesn't Do
Several common assumptions about "AI-powered buyer data" don't apply to how we work, and it's useful to be explicit about them.
We don't generate synthetic interviews. Every interview in the corpus is a real long-form executive conversation that happened. We license the analysis of the conversations; we don't license the conversations themselves. We never synthesize buyer dialogue from prompts.
We don't use LLMs to produce the briefs from scratch. Large language models are used in extraction stages — for structured parsing of transcripts — but the briefs themselves are structured aggregations of extracted content, not generated text. The brief's claims are traceable back to specific interview passages.
We don't claim representation we can't verify. Where the corpus is dense for a particular archetype, the brief reflects that density. Where the corpus is sparse, the brief is transparent about it. We don't extrapolate a sparse-corpus archetype into a full-density brief.
What the Pipeline Enables
The methodology produces structured data that AI products can deploy. Specifically:
- AI sales training tools consume briefs to simulate real buyer language in roleplay scenarios.
- ICP scoring systems consume briefs to evaluate prospects against behavioral profiles rather than demographic tags alone.
- Message generators consume briefs to produce outreach that matches the buyer's actual vocabulary, priorities, and tone.
- Persona models consume briefs to ground their personas in real interview data rather than aggregate persona templates.
In every case, the structured nature of the brief is what makes it usable as an API input. The methodology is what makes the brief credible as a source of truth about how real buyers actually talk.
If your team is building AI features that depend on knowing how senior buyers actually talk this quarter, the GTM Enrichment partner page is the place to start. The methodology is the answer to the question you'll have before you sign anything — where does this data come from, and how do we know it's any good? This pipeline is the answer.