Call-Transcript Signal Extraction

The workflow

Play 20

Call-Transcript Signal Extraction

1Collect the transcripts

Recorded calls

DiscoveryDemoClient calls

Every call names people nobody ever follows up with. Any recorded call works.

From

Fathom

2Extract the people named

Third-party extraction

PersonCompanyTitleVerbatim quote

Finds everyone mentioned who is not on the call. Built from what was said, never from speaker labels, and tolerant of names the transcript mangled.

Read

Claude

3Deduplicate and merge

Mention foundWhat happens

Same name, same company→Exact match, merged

Misspelled name, same company→Fuzzy match, merged

Three mentions, three quotes→One record, every quote kept

Context quotes are merged, never dropped.

4Apply the resolvability gate

Gate1

Full name

First and last name both heard. A first name alone is held, not guessed.

× Held: first name only

Gate2

Named company

A real company is attached to the person.

× Held: no real company

Gate3

Inferable title

The role can be worked out from the call.

× Held: no title

5Resolve, enrich, check it is current

Match person to company

Only rows that passed the gate get here, so only they cost money.

Match

ClayApollo

Verified email, current role

People are often mentioned by an old title. Confirm the role before anyone writes.

Find

Findymail

Role

LinkedIn

Verify

BounceBan

6Push to Clay and tell the rep

Webhook into Clay

One call per resolvable lead, with the quote attached. Clay owns the sequence from here.

Push

Clay

Tell the rep

The rep who ran the call sees who came out of it.

Notify

Slack

Named leads, with the quote attached

The stackFathomClaudeClayApolloFindymailLinkedInBounceBanSlack

Oriel · GTM plays library

Open full size

The diagram as text

1. 1

   Collect the transcripts

   Recorded calls: Discovery, Demo and Client calls. Every call names people nobody ever follows up with. Any recorded call works. From: Fathom. Tool: Fathom.

2. 2

   Extract the people named

   Third-party extraction: Person, Company, Title and Verbatim quote. Finds everyone mentioned who is not on the call. Built from what was said, never from speaker labels, and tolerant of names the transcript mangled. Read: Claude. Tool: Claude.

3. 3

   Deduplicate and merge

   Mention found → What happens

   • Same name, same company → Exact match, merged
   • Misspelled name, same company → Fuzzy match, merged
   • Three mentions, three quotes → One record, every quote kept

   Context quotes are merged, never dropped.

4. 4

   Apply the resolvability gate

   Gate 1: Full name. First and last name both heard. A first name alone is held, not guessed. Rejected: Held: first name only.

   Gate 2: Named company. A real company is attached to the person. Rejected: Held: no real company.

   Gate 3: Inferable title. The role can be worked out from the call. Rejected: Held: no title.

5. 5

   Resolve, enrich, check it is current

   Match person to company: Only rows that passed the gate get here, so only they cost money. Match: Clay and Apollo.

   Verified email, current role: People are often mentioned by an old title. Confirm the role before anyone writes. Find: Findymail. Role: LinkedIn. Verify: BounceBan.

6. 6

   Push to Clay and tell the rep

   Webhook into Clay: One call per resolvable lead, with the quote attached. Clay owns the sequence from here. Push: Clay. Tool: Clay.

   Tell the rep: The rep who ran the call sees who came out of it. Notify: Slack. Tool: Slack.

   Outcome: Named leads, with the quote attached. Tools: Clay.

The stack: Fathom, Claude, Clay, Apollo, Findymail, LinkedIn, BounceBan and Slack.

Open the diagram full size (1180 px wide)

The play in six steps

1. 01

   Step 1: Collect the transcripts

   Every call already contains people the buyer named, and almost none of them are ever followed up.

   • Pull discovery, demo and client-call transcripts from Fathom.
   • It works on any recorded call, not only sales calls. Onboarding and QBR calls name people too.
   • Run it daily on new transcripts so a name is worked while the context is fresh.
2. 02

   Step 2: Extract the people named

   The value is in the third parties, the people floated on the call who were not on it.

   • One Claude pass per transcript returns person, company, title and the verbatim quote they came from.
   • Third parties only: work from what was said, never from the speaker labels.
   • Expect transcription to mangle names. Ask the model to keep its best reading and the raw text side by side.
3. 03

   Step 3: Deduplicate and merge

   The same person can be mentioned three times in one call with three spellings.

   • Exact match first, then a fuzzy match on name plus company.
   • Merge the context quotes instead of keeping only one.
   • Merge across calls too: the same name from two customers is a stronger lead.
4. 04

   Step 4: Apply the resolvability gate

   The gate is what makes the output usable instead of noisy, and it keeps enrichment spend on rows that can resolve.

   • Require a full name, a named company and a title that can be worked out from the call.
   • Hold first-name-only mentions. Never guess a surname.
   • Hold obvious junk company names, such as a team or a city heard as a company.
5. 05

   Step 5: Resolve, enrich and check it is current

   Only the resolvable rows cost money here, and a mention of 'my old manager at X' points at someone who may have left.

   • Match person to company through Clay and Apollo.
   • Find a verified email with Findymail and re-verify it with BounceBan on send day.
   • Confirm the current role on LinkedIn before the lead moves on.
   • Attach the original quote as context for the angle.
6. 06

   Step 6: Push to Clay and notify the rep

   Clay owns the sequence from here, and the rep should know who came out of their own call.

   • Send one webhook call per resolvable lead into Clay.
   • Post a Slack message to the rep who ran the call, listing who was pulled from it.
   • When the mention was a recommendation, ask the customer for the intro first. Never quote a private call to a stranger.

Tools used

The eight tools this play names. Each links to the other plays that use it.

FathomClaudeClayApolloFindymailLinkedInBounceBanSlack

Related plays

Play 21

Pain-Qualified Segments from Calls

1Gather calls and customers

Sales callsBuyer side

DiscoveryDemoClosing

What was true for the buyer when they chose to talk.

From

Fathom

Closed won and lostBacktest set

Close dateReasonAccount

The customers every segment is checked against.

From

HubSpot

2Mine the buying situations

One schema per call

SituationWhy nowAlready triedVerbatim painOutside-in fingerprint

Every situation is counted across calls. The buyers' exact phrases go into a language bank.

Read

Claude

3Situation plus tension

Segment anatomyTwo facts pulling apart

Doing

The thing they are doing

Hurts

The thing that makes it hurt

Now

A deadline, target or season

Number

The figure that costs the quarter

4Membership test, field by field

Condition isUse it as

Public: the field states it→Filter and copy

Proxy: the field suggests it→Score and ask, never assert

Not public→A question for the call

5Measure before building

Gate1

Count it

Free counts, with the query and the date recorded.

× Too small or too broad

Gate2

Eyeball five rows

Open the evidence. Two of five wrong means the test is wrong.

× Fix the test

Gate3

Backtest for lift

Customers in the 90 days before they bought, against a TAM sample.

× Low lift

6Build the list and the opener

Refreshable list

Discover on one filter, score on the rest. Unknown never passes. Emails re-verified on every refresh.

Find

Explorium

Evidence

Exa

Build

Clay

Verify

BounceBan

Opener in their words

Names the situation, says where you saw it, asks if the read is right.

Email

Smartlead

LinkedIn

HeyReach

Segments with measured lift, in the buyers' own words

The stackFathomHubSpotClaudeExploriumExaClayBounceBanSmartleadHeyReach

Oriel · GTM plays library

Play 21Outbound

Pain-Qualified Segments from Calls

6 steps · Email, LinkedIn

+3 more

Play 19

Meeting Prep and Follow-Up

1Pull tomorrow's meetings

Booked meetings with an external attendeeDaily run

2Research the company, the people, the history

Company

NewsHiringFunding

Searched inside an absolute date window.

CRM history

Deal stageOwnerLast emails

Falls back to attendee email when no deal is linked.

Past calls

NotesOpen items

What they already said, so nobody asks twice.

3Around the call

The briefDay before

Who is coming, what changed, what is still open, a suggested agenda.

Write

Claude

Post

Slack

Save

HubSpotlink on the deal

Reminder to the attendeeDay of

A short note with the time and the link, sent to the number they gave when booking.

Text

WhatsAppTwiliobooked meetings only

The transcriptAfter

A webhook fires when processing ends. Poll until it has sentences.

Fetch

Fathom

The call

4One meeting, one record

Gate1

Same meeting id

A repeat webhook for one transcript is dropped.

× Repeat webhook

Gate2

Same client, title and date

Several recorders in one call make several transcripts with different ids.

× Duplicate recording

5Recap, route, follow up

Recap

Decisions and open questions, posted to the account channel.

Post

Slack

Action items

Each one goes to the person who committed to it, else the owner.

Extract

Claude

CRM and follow-up

Next step on the deal. The email is drafted, a human sends it.

Log

HubSpot

Draft

Gmail

Text

WhatsAppexisting chats only

Every call prepped the day before, followed up the same day

The stackGoogle CalendarExaHubSpotFathomClaudeSlackWhatsAppTwilioGmail

Oriel · GTM plays library

Play 19Inbound · Outbound

Meeting Prep and Follow-Up

6 steps · Email, CRM, WhatsApp, SMS

+3 more

Play 17

Automated Reply Enrichment

1Fire the moment someone replies

Prospect replies to a campaign

Reply event to a Clay webhookInstant

No polling and no hourly batch. Speed to lead is the whole point.

2Normalise every reply

One consistent shape

EmailLinkedIn URLTitleCompanyReply text

Deduped against contacts already in the table before anything reads it.

3Classify, then enrich only Interested

Classify the reply

InterestedNeutralNot interestedOut of office

Everything the rep needsInterested only

Direct phoneFundingHeadcountTech stack

The rep opens the record and it is already complete. The phone is checked live before it is shown.

Phone

BetterContactFullEnrich

Company

Clay

Suppress or re-queueEverything else

Not interested is suppressed. Out of office is re-queued for the return date. Neutral gets a human read.

Queue

Smartlead

4Sync to HubSpot without duplicates

Match foundHubSpot action

Email matches a contact→Update that contact

No email match, domain matches→Create contact on that company

No match at all→Create contact and company

Reply text and classification stamped on the record

5Alert the rep

Slack alertUnder a minute

The full enriched profile in the alert, not just a name, with links to act straight away.

Links

HubSpotLinkedIn

Call

Nooks

Interested replies answered in minutes

The stackSmartleadClayClaudeBetterContactFullEnrichSlackHubSpotLinkedInNooks

Oriel · GTM plays library

Play 17Outbound

Automated Reply Enrichment

6 steps · Email, CRM, Call

+3 more

All GTM plays