# Browser and computer-use agent training data

Canonical: https://refer.sourcex.si/data-types/browser-agent-training-data
Last updated: 2026-10-20

Computer-use and browser agent training data records a goal, the starting state, the actions taken in software and the verified outcome. Managed service providers hold the nearest real equivalents in helpdesk tickets, change logs and runbooks, but client-owned content needs consent, and partners never handle the records.

## Which records count as computer-use and browser agent training data?

Computer-use and browser agent data records what a person or system did inside software: the goal, the screens or interface states seen, the actions taken (clicks, entries, selections, commands) and the result. For a managed service provider (MSP), the closest real-world equivalents are helpdesk tickets with remediation steps, admin console change histories and onboarding runbooks, because they document real software workflows with outcomes.

Browser and computer-use agents operate software the way a person does. To learn, they need examples of tasks completed in actual SaaS tools, not only descriptions of those tools. That is a gap licensed operational records can fill.

## What is the structure of a useful example?

Every useful example answers four questions: what was the goal, what was the starting state, what steps were taken and what was the final state. The table maps the elements to what an MSP typically holds.

| Element | What it means | MSP record that carries it |
|---|---|---|
| Goal | The task to accomplish | Ticket title and description, request type |
| State | What the system looked like before | Configuration snapshots, tenant settings, device or user attributes |
| Actions | The steps performed | Technician notes, runbook steps, change logs, scripts run |
| Outcome | Whether it worked and how it was verified | Ticket closure code, validation notes, reopen status |
| Context | Constraints and policy | Approval records, SLA terms, customer-specific procedures |

Records with all five are rare in public text. Screen-level captures are even rarer, so most licensable material is text-based history of actions rather than video of screens, and what exists depends on what the company actually kept.

## Which MSP and SaaS-admin histories document real software workflows?

- **Onboarding and offboarding tickets:** account creation, license assignment, group membership and device setup, repeated across many clients.
- **Password reset and access requests:** short, high-volume workflows with clear completion codes.
- **Change management records:** requested change, risk notes, approval, implementation steps and rollback.
- **Migration projects:** mailbox, file share and application moves with sequenced steps and post-checks.
- **Security alert triage:** alert, investigation steps, containment actions and closure rationale, where records do not expose client secrets.
- **Runbooks and SOPs:** the written procedure that technicians follow, paired with tickets showing how it was applied.

The [managed service provider partner page](/for/managed-service-providers) covers the role from the partner side, and our guide on [how AI coding agents are trained](/questions/how-are-ai-coding-agents-trained) shows a parallel case in software engineering.

## Who owns these records? Vendor-owned versus customer-owned

This is the central rights question for an MSP. Three categories exist, and only one is plainly the company's.

| Category | Examples | Whose is it |
|---|---|---|
| The MSP's own operational records | Internal runbooks, its own ticket narratives, its staff's notes, its billing and project records | Generally the MSP, subject to its contracts |
| Client-owned content | Client mailboxes, files, user data, configuration details, credentials | The client; the MSP cannot license it without consent |
| Vendor-owned material | Software documentation, vendor knowledge base text, proprietary tooling output | The vendor, per its terms |

Many MSP contracts treat client information as confidential, so even ticket text that mentions a client can be restricted. The company, with counsel, decides what is clearly its own, what needs client consent and what to exclude. See [who owns company data stored in SaaS tools](/questions/who-owns-data-in-saas) for the general principles. If the MSP's value comes mainly from client data, it may not qualify. This is general information, not legal, tax or financial advice.

## How do you recognize an MSP with deep records of this type?

- [ ] The firm has operated for several years and has 50+ full-time employees at peak (contractors excluded).
- [ ] Tickets sit in a professional services automation or helpdesk tool with fields for category, resolution and time.
- [ ] Runbooks exist and are linked to tickets or projects.
- [ ] Old tools were archived rather than deleted.
- [ ] The firm has an administrator who can run a full export.
- [ ] Contracts allow the MSP to use its operational records, or a clear path exists to obtain consent.
- [ ] The owner or an executive sponsor will consider an exclusive AI-training license.

If the first three are true but the rights question is open, suggest the firm review contracts first rather than dropping the idea.

## What does a partner never do?

Partners make the introduction and give basic fit information only. A partner never exports, uploads, reads or describes confidential records, never touches client tenants and never asks technicians for ticket contents. The company completes a data inventory with SourceX, and de-identification and redaction rules are agreed before any work begins. Data is delivered only after an executed agreement and the company's authorization.

## How does synthetic data compare?

Generated browser tasks can scale, but they reflect their authors' assumptions about how software behaves. Real tickets include the odd configurations, partial failures and client-specific rules that scripts miss. The [comparison of synthetic and real business logs](/compare/synthetic-data-vs-real-data-ai-agents) explains where each helps. For the owner worried about what comes next, [will AI trained on our data replace us](/questions/will-ai-trained-on-our-data-replace-us) and [whether licensed data can be removed from a model](/questions/can-data-be-removed-from-ai-model) answer the common objections, and [the case that agents need work data](/proof/ai-agents-need-work-data) gives the broader argument.

## Next step

If you advise or work with an MSP that has years of ticket and runbook history, run it through the [company fit checker](/tools/company-fit-checker), review [how the process works](/how-it-works) and [register as a partner](/register) to introduce the sponsor. A reward becomes payable after SourceX has paid the referred company under its signed agreement and the payout conditions in the program terms are met, including finance approval and your tax and payment details.
