Skip to content
DataDouble Get a valuation
Working with frontier AI labs

Your company’s data is worth more than you think.

DataDouble works with frontier AI labs to buy and anonymize data from companies like yours. We remove every name and identifier, license what’s left to the labs, and pay you for it. Your customers stay private.

  • Customer names and details removed
  • Raw customer records never reach buyers
  • Buyers barred from re-identifying anyone
support_thread_48213.txt
8 identifiers found
From: maria.okafor@willowbrookdental.com

Hi team, this is Maria Okafor from Willowbrook Dental. Invoice INV-20931 was charged twice to the card ending 4417 on March 3. Could someone call me at (503) 555-0148? I’m at the Alder Street office until five.

Detect Replace Generalize Attack Approve
Raw record. Never leaves your environment.
Removed before release Names Emails Phone numbers Addresses Account numbers Card numbers Government IDs IP addresses Device IDs Faces Voices Signatures
Why now

The public web is running dry. The next models need what’s inside companies.

Frontier models learned to write and reason from the open internet. To learn how real work gets done, they need records that were never published: support resolutions, sales calls, code reviews, contracts. Labs now buy that data directly from the companies that hold it.

Six figures
Reported size of some deals for a single defunct startup’s Slack, email and Jira archives. Forbes, April 2026
$203M
Aggregate data-licensing contract value Reddit disclosed in its IPO filing. Reddit IPO filing, via LLM Pulse
2026–2032
The window in which Epoch AI projects AI developers will exhaust the stock of public human-written text. Epoch AI, via AP
What labs are buying

The everyday records your teams create are training data.

If it shows how skilled people solve real problems, a lab probably wants it. Most companies have years of it sitting in tools they already pay for.

Customer conversations

Support tickets, live chats and email threads, with the resolution attached.

Zendesk · Intercom · Front

Sales calls

Call recordings and transcripts, plus the CRM notes and outcomes that followed.

Gong · Salesforce · HubSpot

Code and engineering

Repositories, pull requests, code reviews, tickets and incident write-ups.

GitHub · GitLab · Jira · Linear

Team communication

Channels, threads and email archives that show how decisions get made.

Slack · Teams · Gmail · Outlook

Internal knowledge

Wikis, runbooks, playbooks and the decision docs behind them.

Notion · Confluence · Google Drive

Documents and contracts

Contracts, redlines, filings, reports and completed forms.

SharePoint · Box · Drive

Operational records

Orders, transactions, schedules, claims and logistics events.

Snowflake · BigQuery · Postgres

Recordings of expert work

Screen sessions, inspection photos and field video of skilled people at work.

Session logs · Media libraries
Calculator

How much could your data earn?

Set the sliders to roughly match what your company holds. The estimate updates as you go.

250K
Support tickets, chats and email threads
5K hours
Sales and support calls, measured in hours
50K
Contracts, reports, runbooks and internal docs
25K
Issues, code reviews and incident write-ups
6 years
How far back your records go
Estimated licensing value
Your data could be worth $508,800 to frontier AI labs. Likely range $356K to $712K
Customer conversations$69K
Recorded calls$275K
Documents and wiki pages$55K
Tickets and pull requests$110K
Get the full breakdown
Estimate only, not an offer. Actual value depends on data quality, uniqueness and current lab demand.

How we estimate

Rates are inferred from publicly reported AI data deals.
Conversations $0.25 each

Text sells for about $0.001 per word through data marketplaces, and a support thread runs a few hundred words.

Reuters via The Decoder, 2024
Recorded calls $50 per hour

Call-center recordings are listed at about $15 per call on data marketplaces, and video footage sells for $1 to $4 a minute.

Datarade listingBloomberg via The Decoder, 2025
Documents $1 each

At about $0.001 per word, a typical 1,000-word internal document is worth around $1. Books have licensed for $5,000 per title.

The Decoder, 2025
Tickets and PRs $4 each

SimpleClosure says it is on track to return more than $10M to 350+ startups whose archives hold about 200,000 pull requests. We price against that conservatively.

Business Wire, September 2026

Each year of history adds 5%, from 0.85× at one year up to 1.5×. A specialized field multiplies the total by 1.4× and exclusivity by 1.75×. Those two premiums are our judgment, since no public benchmark exists for them.

How it works

Four steps from first call to first payment.

You stay in control at every step, and your team’s part is small.

  1. 01

    Valuation

    Describe your data at the schema level: systems, volume and history. No raw records needed. We come back with a value range and the labs most likely to buy.

    Your part: one call
  2. 02

    Rights review

    Our counsel reviews your privacy policy, customer terms and contracts to confirm what you can license. If something can’t be sold, you hear it before you sign.

    Your part: share your policies
  3. 03

    Anonymization

    We de-identify every record in your cloud or an isolated environment, attack the output for re-identification, and send you a sampled report.

    Your part: approve the report
  4. 04

    Payment

    Choose an outright sale, a revenue share or an advance plus royalties. Every license and payment is reported to you.

    Your part: choose how you’re paid
Privacy by construction

Five layers between your records and any lab. You sign off on the last one.

Raw data never leaves your control, and no dataset ships until it survives a re-identification attack and your review.

  1. 01 Detect Pattern rules, entity models and LLM reviewers flag names, contact details, account numbers and other identifiers in text, transcripts, documents and metadata.
  2. 02 Replace Each identifier becomes a consistent surrogate. [PERSON_1] stays [PERSON_1] across a whole thread, so the data still reads naturally.
  3. 03 Generalize Exact dates, rare job titles, small towns and deal sizes are coarsened until no combination points to one person or company.
  4. 04 Attack A separate red team runs linkage and inference attacks against the output. Anything that can be re-identified goes back through the pipeline.
  5. 05 Approve You review a sampled report and approve the release. Nothing ships without your sign-off.
DD-RPT-0142 Sample
De-identification report
Customer support threads · Prepared for seller review
Records processed
1,248,310
Direct identifiers removed
3,904,117
Quasi-identifiers generalized
611,842
Residual identifiers in human review
0 of 2,000
Attacks run
Linkage · Inference · Membership
Re-identification result
Passed
Awaiting your approval
Sample report. Figures are illustrative.
Your data, your rules

You decide who buys it, how it’s used and for how long.

Approve every buyer

See which lab wants your data before any deal closes. Decline anyone.

Exclude competitors

Block named companies or whole industries from ever licensing your data.

Limit how it’s used

Restrict use to training, evaluation or research. Resale is never allowed.

Ban re-identification

Every license prohibits attempts to identify any person or company in the data.

Stop at any time

End future licensing whenever you choose. Licensed copies are deleted when each license ends.

See everything

Every dataset, buyer, license and payment is reported to you in plain terms.

For research labs

Real-world data, with the paperwork your counsel needs.

We source proprietary data from operating companies and deliver it de-identified, rights-cleared and documented, so your legal and safety teams can sign off.

  • Chain of title and seller attestation for every dataset
  • A de-identification report with red-team results
  • Exclusive and non-exclusive licenses
  • Custom sourcing to your specification
Explore data for labs
Example dataset card DD-DS-0142
B2B support conversations
Modality
Text and metadata
Records
1.2M threads
Time span
2017 to 2026
Languages
English, Spanish
Rights
Seller-attested, counsel-reviewed
De-identification
Five layers, red-team passed
License
Non-exclusive, training and eval
Formats
JSONL, Parquet
{
  "thread_id": "t_8f21c0",
  "industry": "dental_practice",
  "turns": 6,
  "resolution": "refund_issued",
  "messages": [
    { "role": "customer",
      "text": "Hi team, this is [PERSON_1] from [ORG_1]…" },
    { "role": "agent",
      "text": "Thanks, [PERSON_1]. I’ve refunded the duplicate charge…" }
  ]
}
FAQ

Questions sellers ask first

Something else on your mind? Write to hello@datadouble.us

Is it legal to sell our company’s data?
Often, yes. It depends on your privacy policy, your customer contracts and the laws that apply to your data. Every engagement starts with a rights review by our counsel, so if something can’t be licensed, you know before you sign.

Will our customers or employees be identifiable?
No. Every dataset passes through our five-layer pipeline and a re-identification attack before release, and every buyer is contractually barred from trying to identify anyone.

Do we have to hand over raw data?
No. We can run the entire pipeline inside your own cloud account, so raw records never leave your infrastructure. Only the approved, de-identified output is exported.

Who buys the data?
Frontier AI research labs and model developers. You see every prospective buyer and can decline any of them.

How much is our data worth?
It depends on volume, history, domain, uniqueness and exclusivity. Deep, specialized data that was never published commands the highest prices. The calculator above gives a rough figure, and a free valuation gives you a real one.

How do we get paid?
Choose an outright sale, a revenue share as your data is licensed, or an advance plus royalties. Your valuation includes the options that fit your data.

Could a competitor end up with our data?
Only buyers you approve can license it. You can exclude any company or industry, and every license prohibits resale.

What does a valuation cost?
Nothing. Valuations are free and confidential, and you’re under no obligation to sell.

Find out what your data is worth.

A valuation is free and confidential, and it needs no raw data. If the numbers don’t work for you, you walk away.

Get a free valuation or email hello@datadouble.us