What data buyers want

Records of how the work actually got done

AI models are good at answering questions and still weak at doing a day's work inside a real company. Records that show real requests, decisions and outcomes help developers build practice tasks and tests for AI agents.

Who buys

AI developers ("labs")

Companies that build AI models and license data to train and evaluate them. They often prefer to work through a small number of vetted suppliers.

Practice-environment builders

Firms that turn real records into simulated workplaces, with tools, tasks and pass/fail checks, where AI agents practice.

Data companies & marketplaces

Intermediaries that evaluate, prepare and license company data to the groups above. Some list datasets for buyers to browse.

These describe types of buyers in the market. Legacy Signal doesn't claim partnerships with any of them, and terms differ widely from buyer to buyer.

Records that tend to fit

Chat & email

Threads where problems were diagnosed and fixed.

Tickets

Support and IT tickets, especially with resolutions.

CRM & sales

Deal history, activities, notes, won and lost.

Procedures

SOPs, playbooks and escalation guides, more useful when linked to real cases.

Projects

Tasks, handoffs and timelines.

Knowledge base & files

Internal wikis, templates, training material.

Finance & operations

Invoices, work orders, dispatch, quotes.

Code

Repositories with issues and reviews.

Specialist systems such as CAD/BIM (e.g. Revit), ERP and supply-chain records interest some buyers, but it varies by buyer and isn't assured. We list them separately and check fit case by case.

What makes records more useful

  • Several tools that refer to the same customers, jobs or tickets
  • Years of continuous history, not a few months
  • Outcomes on record: resolved, paid, won, lost
  • Real messiness: exceptions, overrides, workarounds
  • Desk-based work by many people
  • Clear rights and no earlier exclusive deal

What's usually left out

  • Client-owned or client-confidential files
  • Privileged legal communications
  • HR and payroll records
  • Health records, payment cards, ID numbers
  • Anything about children
  • Passwords, keys and credentials
  • Public material buyers already have

What a de-identified record looks like

Fictional example. Buyers usually replace names and details with consistent placeholders, so the work still makes sense without exposing people.

Original · stays with you

From: Jenna Ruiz ([email protected]) to Carl Banks, Banks Dental, Elm St.
"Tech Marco (555-0142) replaced your rooftop unit compressor Tue; warranty #W-88213."

De-identified copy

From: [Staff_04] ([EMAIL]) to [Customer_217], [Business_type: clinic], [Street].
"Tech [Staff_11] ([PHONE]) replaced the rooftop unit compressor [Day]; warranty [WARRANTY_ID]."

No method removes every risk of re-identification. You review the method and an approved sample before anything is released.