Second Exit.
We buy your company’s working record to train AI.
Your threads, tickets, code, documents, photos and videos show how real work gets done. AI labs license that record to train agents. We anonymize one workflow as a sample, grade it and quote a firm price per gigabyte. The sale covers an anonymized copy of the record. Your shares, customers and original files stay with you.
- company
- B2B SaaS · 8 people · 3 years
- workflow
- customer bug → spec → fix → review → release → reply
- teams
- support, product, engineering
- artifacts
- 214 messages · 9 tickets · 31 commits · 6 screen recordings
- removed
- 1,208 identifiers · 14 secrets · 23 faces blurred
- residual
- 0 identifiers found in human review
- grade
- A · continuous, cross team, with media
▸ anchor estimate ≈ $500K · firm quote issued per GB
Why AI labs buy company records
Language models were trained on what the public internet contains: articles, forums, open source code, books. The next generation is trained to act as agents that take a task and finish it inside a company. For that, a model needs examples of tasks being finished: the request, the discussion that clarified it, the ticket, the code change, the review, the release and the reply to the customer, in order, across the people who handled each step.
That sequence sits inside company tools that search engines never index. AI labs and agent builders price it by scarcity. An acquirer prices a company on revenue and growth, which puts a small startup with modest revenue at a modest valuation. A lab prices the same company’s record on how rare its workflows are, and for a company with years of linked history that figure exceeds the valuation.
After cielo24 shut down, its former chief executive sold 13 years of Slack, email and Jira history for hundreds of thousands of dollars (Forbes, April 2026).
Source
- 2028
Median year in which the stock of public human text (about 300 trillion tokens) is fully used for training, per Epoch AI.
Source - $200 to $2,000
Price of a single reinforcement learning task sold to a lab, up to $20,000 for the hardest, per Epoch AI (January 2026).
Source - $1B+
What Anthropic leaders discussed spending on training environments for agents within a year, as reported by The Information (September 2025).
Source
Who pays whom, and where our margin comes from
Three parties take part in each sale: you, La Boétie, and the model builders who license the corpus.
- 01
You
- Gives
- A licence to an anonymized copy of your working record.
- Receives
- A fixed price, paid on delivery, set per gigabyte from the reviewed sample.
- 02
La Boétie
- Gives
- The extraction, the anonymization, the packaging into training sets and task environments, and the risk of reselling.
- Receives
- Licence fees from model builders. Our margin is that income minus the price paid to you and the cost of processing.
- 03
Model builders
- Gives
- A licence fee per dataset. Their contract forbids any attempt to reidentify a person or a company.
- Receives
- Anonymized, continuous workflows they can train and evaluate agents on.
We license the same anonymized corpus to several buyers. The income from several licences funds a price above what a single buyer would pay. We carry the resale risk: your price is fixed at signature and paid on delivery, whatever number of licences we sell afterwards.
The sample: one workflow across teams
To price your record we ask for one transversal workflow: a single task your company carried from A to Z through the teams that touched it. The sample shows us the continuity, depth and cleanliness of the rest of the record.
- 01 →
Request
The customer email, the sales call note or the internal ask that started it.
- 02 →
Discussion
The Slack or Teams threads where people argued the scope and made decisions.
- 03 →
Specification
The ticket, the spec document, the design file, with their revisions.
- 04 →
Execution
The commits, the pull request and its review, or the field report and photos.
- 05 ■
Delivery
The release, the handover, the invoice, the reply that closed the loop.
How photo and video are priced
Multimodal and robotics models train on footage of real tasks: a technician’s phone video of a repair, the photos of a site inspection, a screen recording of an analyst in a spreadsheet. Licensed video already trades by the minute, at about $6 per HD minute (Variety, 2024). In your record, the gigabytes that hold photo and video get a higher rate than text.
SourceThe six steps of the anonymization pipeline
The same pipeline processes the sample and, later, the full record. You receive the anonymized output of your sample before our quote and check it yourself.
- 01
Read only export
We take exports from the tools you already use (Slack, Teams, Google Workspace, Microsoft 365, Jira, Linear, GitHub, GitLab, Notion, Drive, mailboxes, photo and video libraries). We read from your systems without writing to them or installing software.
- 02
Detection
Every document goes through entity recognition and pattern rules for names, emails, phone numbers, postal addresses, bank and card numbers, national identifiers and customer references. A separate scanner finds credentials: API keys, tokens, passwords, private keys.
- 03
Consistent pseudonyms
Each person and each customer becomes one stable placeholder across every tool, so the thread in Slack, the ticket in Jira and the commit in Git still point to the same actor.
- 04
Media treatment
Faces and licence plates are blurred, metadata and GPS coordinates are stripped, customer data visible on screen recordings is masked, and speech is transcribed then scrubbed like any other text.
- 05
Human review
A reviewer reads a random share of every batch looking for anything the automated passes missed. A batch with a residual identifier goes back through the pipeline before it moves on.
- 06
Key destruction
The table linking placeholders to real identities is used during processing only and deleted once the batch is accepted. After deletion, no stored table links a placeholder to a person.
The perimeter of the sale
In the perimeter
- Internal conversations: Slack, Teams, email threads between colleagues.
- Work tracking: tickets, boards, specifications, meeting notes.
- Code and its history, including reviews and incidents, if you choose to include it.
- Documents, spreadsheets and design files with their revisions.
- Photos, screen recordings and videos produced during the work.
Outside the perimeter
- Your shares: no investor enters your capital.
- Your customers and contracts: none are transferred.
- Personal data of your customers and staff: it is removed before delivery.
- Secrets and credentials: they are detected and deleted.
- Anything you mark out of scope, such as a core algorithm or a client under strict NDA.
You keep your data and every right to use it. The transaction covers a licence to an anonymized copy. Your company runs the day after signature as it did the day before.
The five steps of a sale
The estimate, the sample and the quote are free and leave you able to stop at any point. You decide whether to sell once you hold a firm price.
- 01
Estimate
Answer the five questions of the estimator below. You get a range, with the multiplier each answer applied.
- 02
Sample
Under a confidentiality agreement (NDA), you pick one workflow and export it. We anonymize it and send the anonymized version back to you.
- 03
Grading and quote
We grade the sample on continuity, depth, media share and cleanliness, and we issue a firm price per gigabyte of anonymized output.
- 04
Extraction
We process the full record in batches. Each batch is measured in gigabytes after anonymization, and you see the running total.
- 05
Payment
You are paid the quoted rate times the delivered volume. Processing, storage and review are at our cost.
The estimator and its five factors
The reference company is a software startup of under ten people, three to five years old, whose record is text and code and can be followed across teams. Its record prices around $500K. Each of the five answers below multiplies that anchor.
Buyers pay more for domains the public web documents poorly, such as industrial engineering and manual trades.
More people means more handoffs between roles, and handoffs are the part of a workflow a model cannot learn anywhere else.
A longer history holds more volume and more revisions: decisions revisited, systems migrated, errors found and fixed.
Multimodal and robotics models need footage of real work, and each gigabyte of it gets a higher rate than text.
A buyer trains on the whole path of a task. At equal volume, a record with broken links between tools is worth less.
$350K to $700K
- Anchor
- $500K
- Domain
- ×1
- People
- ×1
- Age
- ×1
- Media
- ×1
- Continuity
- ×1
- Central figure
- $500K
The range spans 0.7 to 1.4 times the central figure. The sample review replaces it with a firm price per gigabyte of anonymized output.
| Factor | Options and multipliers |
|---|---|
| Domain | Software and SaaS ×1 · Professional services ×1.3 · Health and life sciences ×1.6 · Engineering and industrial R&D ×2 · Manual trades, field work, manufacturing ×2.5 |
| People | Under 10 ×1 · 10 to 49 ×2 · 50 to 249 ×4 · 250 and more ×8 |
| Age | Under 3 years ×0.6 · 3 to 5 years ×1 · 6 to 10 years ×1.5 · 11 to 20 years ×2 · Over 20 years ×2.6 |
| Media | Text and code only ×1 · Some photos or screen recordings ×1.25 · Photo and video are central to the work ×1.6 |
| Continuity | Siloed tools ×0.7 · Inside one team ×0.85 · Across teams, end to end ×1 |
Questions from sellers
How can the record be worth more than my company?
Two different buyers price them. An acquirer prices your company on revenue, growth and risk. A model builder prices your record on scarcity, because continuous examples of real tasks finished across teams sit inside companies. Epoch AI puts the price of a single training task sold to a lab at $200 to $2,000 (January 2026). A record holds thousands of such tasks, which is how its price passes the valuation of a young company with modest revenue.
What does a typical deal look like?
Our reference point is a software startup of under ten people, three to five years old, with a record that can be followed across teams: around $500K. Technical and manual domains, more people, more years and more photo and video raise that figure, and the estimator shows by how much. The firm price comes per gigabyte after the sample review.
Why price per gigabyte?
Because it is measurable by both sides. The sample review fixes a rate that reflects quality (continuity, domain, media share). The volume is measured after anonymization, batch by batch, so you are paid for what is actually delivered and can check every figure.
Who buys the data?
AI labs and companies building agents buy the corpus to train and evaluate models on realistic work. Anthropic leaders discussed spending over $1 billion on such training environments within a year (The Information, September 2025). Every licence forbids attempts to reidentify people and any use outside training and evaluation. Buyer names stay confidential under their contracts.
Is it legal under the GDPR?
The raw export contains personal data, so we process it as your processor under a data processing agreement, as Article 28 of the GDPR (General Data Protection Regulation) requires. Pseudonymized data still counts as personal data. Recital 26 places anonymous data outside the regulation, which is why we remove identifiers and delete the mapping table before delivery. The contract lists the excluded categories before the first export.
Can my competitors read my internal discussions?
Buyers receive the corpus under a licence limited to training and evaluating models. Before delivery, our pipeline replaces your company name, product names, customers and people with placeholders. You choose the perimeter and can exclude any project, client or piece of code from it.
My company shut down. Can I still sell?
Yes, as long as the exports still exist and the seller holds the rights to them. After a closure, the liquidator or the former directors hold those rights.
Do I lose the use of my own data?
No. The contract grants a licence to an anonymized copy of the record. Your originals stay on your systems, and you keep every right to use, sell or delete them.
Put a price on your record.
Run the estimator, then send us the request. We reply with the NDA and the export instructions for your sample.
Estimate your record