Tool-use data: how AI models learn to operate business software

What is tool-use training data?

Tool-use training data is a record of an actor choosing a software action, supplying the right inputs and getting a result back. It teaches AI models and agents when to call a function, which arguments to pass and what to do with the response. Companies already generate this material every day as audit logs, workflow histories and ticket actions, which is why it matters to a referral partner.

Public function-calling datasets tend to describe individual calls. What is harder to find is a real business sequence: who did what, in which system, in what order, and whether it worked. That gap is where company records become interesting to AI labs and data buyers.

What do the records look like inside a company?

Each business system keeps a trail of actions. The value is in the trail, not in any single entry.

Admins sometimes call this an audit log. AI developers call the same thing a trajectory: a sequence of actions with state before and after.

What makes it useful for training and evaluation?

Four properties separate a useful trail from a noisy one:

1. Order and timestamps. Preserved sequence turns scattered rows into a workflow. The note on why metadata raises the value of business data explains how.
2. Named actors and roles. Whether a person, a bot or an integration acted affects what a model can learn.
3. Outcomes. A call that succeeded, failed or was reversed gives a label that training and checking can use. See verifiable rewards and checkable business outcomes.
4. Context links. The action matters more when it connects to the ticket, customer thread or invoice that prompted it.

Not every log qualifies. A raw server access log with no business context is thin. A support platform's action history linked to the conversation is rich.

Where do these records live?

Most live in the products a company already pays for, and most strong companies run 10-15+ systems. Look for:

• Admin audit exports from CRM, helpdesk, HR and finance platforms
• Workflow automation and integration run histories
• Change and approval logs in project and engineering tools
• Shared-drive activity histories where retention is on
• Archived systems from earlier migrations, which extend history

Retention differs by product, plan and configuration, and nobody outside the company should assume what a given vendor keeps. The company's own administrator can say what can actually be exported.

What are the rights and privacy considerations?

Action trails often include employee names, customer names and message snippets. Agreeing de-identification and redaction rules with the company is a step that happens before any work starts, and data is delivered only after an executed agreement and the company's authorization. Partners never see or handle the records. Also watch for logs generated by a vendor under terms that limit reuse, and for client data held on behalf of customers, which needs consent. The explainer on why AI developers want data from US companies covers the clean-rights angle.

How can a partner recognize a company with deep action trails?

Ask these questions in a first call, without asking to see anything:

• Does the company use at least five systems of record, such as CRM, helpdesk, ERP, project tools and chat?
• Has it kept the same main systems, or archived old ones, for several years?
• Is there a named administrator who can pull audit or activity exports?
• Do workflows run through the systems, rather than through email and memory?
• Is it a US company with 50+ full-time employees at peak (contractors excluded)?
• Is an owner, CEO, CFO or authorized representative willing to consider a license?

As a rule of thumb, not a SourceX score, four or more yeses is a reasonable prompt to run the company fit checker together; it is a preliminary, non-binding screen.

What to say

When this is not a fit

• The company runs mostly on spreadsheets and email with no system logs.
• Audit history was disabled or purged and no export exists.
• The logs belong to a customer or outsourcer who has not agreed.
• The company has already licensed the same records for AI training.

How rewards work

Partners earn 25% of the eligible platform fees SourceX actually collects from the referred company's licensing deals, capped at $100,000 per referred company. The reward is payable only after the buyer pays and SourceX receives its fee, and no reward is guaranteed. It is a share of SourceX's fee and is never deducted from what the company receives.

Next step

Pick one client or portfolio company with a clear system footprint and run the questions above. Then register as a partner and introduce them, or read how it works first.