Training and evaluation data by use case

Each guide maps a model or agent use case to the business datasets that train and evaluate it best, explains why that data rarely exists publicly, and lists what good data for the use case looks like.

• Coding agents

  Coding agent training data and private eval tasks: linked issues, pull requests, review threads, CI runs and tests, licensed from software companies.

• Customer support agents

  Customer support AI training data from real operations: resolved tickets, call transcripts, policies, QA scores and the back-office actions behind each case.

• Enterprise and computer-use agents

  Enterprise AI agent training data from real business processes: system-level task trajectories, SOPs, files, approved email and review outcomes.

• Document AI, enterprise search and RAG

  RAG evaluation datasets and document AI training data from real company archives: files, spreadsheets, decks, SOPs and contracts with versions and permissions.

• Finance and accounting agents

  Finance and accounting AI training data from accounting firms and finance teams: GL coding, reconciliations, close workpapers and reviewer sign-offs.

• Legal AI

  Legal AI training data from real contract work: draft-to-signature redline histories, playbooks, matter files and lawyer review, with rights reviewed.

• Voice agents and speech models

  Voice agent training data from real business calls: two-party audio, transcripts, turn timing, dispositions and QA scores, licensed with spoken PII masked.

• Sales agents

  AI sales agent training data from real B2B pipelines: CRM stage histories, call transcripts, email threads and proposals, linked to won and lost outcomes.

• Healthcare administration AI

  Healthcare administration AI training data: de-identified prior authorizations, claims and remittances, denials and appeals, payer calls and billing playbooks.

• Robotics and embodied AI

  Robotics training data for embodied AI from real work: new first-person video of skilled trades, plus field service, quality, construction and CAD records.

• Private evaluation sets

  Build private, held-out AI evaluation datasets from real business work: support cases, code changes, workflows and redlines, with outcomes and expert grades.

Updated 3 October 2026.