Internal meeting transcripts with decisions, 27 firms in 10 industries

Summary

Transcripts of internal meetings from 27 US firms of 15 to 150 staff in 10 industries: weekly operations reviews, production huddles, estimating meetings and project kickoffs. Each transcript keeps speaker turns and timestamps and is joined to its agenda or notes doc, so decisions and action items can be checked against what the team wrote down. Good for training models to follow multi-party discussion and draft follow-ups that match what happened next. Audio and video are never collected, one-on-ones and HR meetings are excluded, and names, customers, emails, phone numbers and addresses are replaced with placeholders.

Collected to order. We collect it from US companies for your order. Scope, volume and timing are confirmed in writing before you sign.

Pricing on request

Request scope and pricing

At a glance

Planned scope. Volume, sources and timing are confirmed in writing before you sign.

Records: 24Kmeeting transcripts
Size: 1 GB
Tokens, estimated: 200M
Years: 2020 to 2026
Firms: 27

Owner profile: 27 US firms, 15 to 150 staff each, in 10 industries including commercial printing, civil engineering, freight brokerage and specialty food manufacturing
Industry: Multiple industries
Data type: Company operations archive
Headcount: 15 to 150 staff per firm
Region: United States
Languages: English
Formats: JSONL
Updated: 2026-09-26

Sources

System	Kind	Records
Zoom transcripts	Call transcripts	23.8K transcripts
Shared drives	Shared files	17K agendas and notes docs

Good for

extracting decisions and action items, checked against the team's own notes
multi-party dialogue understanding with speaker turns
drafting meeting follow-ups that match what happened next
evaluating summarization of operational meetings

Fields

Field	Type	Description
transcript_id	string	Stable transcript id
firm_code	string	Code of the contributing firm
industry	enum	One of 10 industry codes
meeting_type	enum	weekly_ops, weekly_production, estimating, kickoff or project_review
started_at	timestamp	Start time in UTC
duration_minutes	int	Meeting length
speakers	int	Number of distinct speakers
turns	array<string>	Speaker turns with offsets, excerpted in this sample
decisions	array<string>	Decisions recorded in the notes doc, where one exists
action_items	array<string>	Action items recorded in the notes doc

Example record

{
  "transcript_id": "[MEET_4410]",
  "firm_code": "F-14",
  "industry": "commercial_printing",
  "meeting_type": "weekly_production",
  "started_at": "2025-06-02T13:00:00Z",
  "duration_minutes": 38,
  "speakers": 6,
  "turns": [
    "[PERSON_3] 00:04:12: Press 2 is still waiting on plates for the [COMPANY_8] job, prepress says noon.",
    "[PERSON_9] 00:04:30: If plates are noon we lose the second shift run. Can we swap in the folder job and push theirs to Wednesday?",
    "[PERSON_3] 00:04:51: Only if sales tells them today. I'll call [PERSON_17]."
  ],
  "decisions": [
    "Run the folder job on press 2 second shift",
    "Move [COMPANY_8] ship date to Wednesday if the client agrees"
  ],
  "action_items": [
    "[PERSON_3] to call [PERSON_17] by 11",
    "Prepress to confirm plate time by 10"
  ]
}

One record, de-identified. Placeholders such as [PERSON_12] and [COMPANY_4] stay the same across the whole dataset, so threads still read correctly.

Licenses available

Exclusive
One buyer. Nobody else can license it during the exclusive term, usually 2 years.
Non-exclusive
Licensed to more than one buyer.
First look
One buyer gets 2 to 4 weeks to make an offer before anyone else sees the dataset.

What ships with it

Datasheet: what it is, what it is good for, verified size, sources, years, format and limits.
Provenance file: the owner shown by code, staff notice and consent status, exclusions, mapped to the EU AI Act Article 53 training-data summary and California AB 2013.
Manifest with a hash for every file.
QC report and de-identification report.
Your own copy, with unique canary records.

How it is de-identified

Default exclusions removed, files flattened to text, identifiers found in three layers and replaced with consistent placeholders, and the result measured against a 99% bar for each identifier type in each source.

The full method