Which jobs are AI agents learning first, and what records sit behind them?

The short answer: tasks first, jobs later

AI agents are learning tasks, not whole jobs. The first wave targets computer-based work with clear inputs, repeatable steps and a result someone can check: triaging support tickets, matching invoices to purchase orders, updating CRM records, scheduling, writing and reviewing code, and assembling standard documents. How much a given job changes depends on how much of it is made of tasks like these.

Each task is learned from examples of the work done well and done badly. Those examples live in company systems, so the question of which jobs agents are learning is also a question about which companies hold the relevant records.

The four-trait task test

A practical way to see where agents are pointed first is to break a job into tasks and score each one on four traits.

A task that scores high on all four is a likely early target. A task that fails one trait, most often verifiability, is harder to hand to an agent and tends to be automated only in part.

Which workflows are in focus, and what sits behind them?

The table maps common back-office and professional workflows to the tasks agents attempt, the records that teach them and the kinds of companies where partners usually meet them.

The engineering row has the most public detail: how AI coding agents are trained explains why repository history is such an effective teacher.

How do public evaluations measure progress?

Three styles of evaluation shape which tasks get attention.

1. Occupational task evaluations ask a model to produce a real deliverable, such as a memo, schedule or spreadsheet, and have practitioners judge it. The page on what GDPval is explains the best-known example.
2. Coding evaluations check whether a proposed change passes a repository's tests.
3. Agent workflow evaluations score multi-step tool use toward a goal, often inside simulated software.

All three favor work that can be checked, which reinforces the four-trait pattern. None measures whether a whole job can be automated. For sourced figures on how far agent deployment has actually gone, use AI agent adoption statistics.

Which tasks are still hard for agents?

The hard cases are the ones that fail the four-trait test, and they explain why records of experienced people matter so much.

Illustrative and fictional: a regional distributor's collections team handles routine reminders that an agent could plausibly take over, but the calls that matter involve a long-standing customer in temporary difficulty. Whether to extend terms depends on years of payment history, past promises kept and the account manager's notes. That decision trail is exactly the kind of record agents lack.

Why is the data behind these tasks scarce?

Finished documents are common online; the work around them is not. To learn invoice exceptions, an agent needs to see the mismatched invoice, the email to the supplier, the approval override and the credit that followed. That trail exists only inside the company that did the work, which makes it first-party data in the strict sense: created by the business in its own operations. What first-party data means in AI licensing covers the term.

Simulated workplaces fill part of the gap, with trade-offs explained in how simulated workplaces compare with real logs. Real records remain the reference for how the work is actually done.

What this means for partners

Use the workflow table to scan the companies you know. A company deserves a closer look when:

• One or more of its core workflows appears in the table and runs through software rather than paper.
• Those workflows have several years of history, including systems that have since been retired.
• Outcomes are recorded: resolution codes, payment status, win or loss, change order approval.
• The records were created by the company, not supplied by its clients.
• It is a US company with 50+ full-time employees at peak (contractors excluded) and an owner or executive who can authorize a license.

If it ticks most boxes, try the company fit checker, then introduce it. How it works shows what follows: qualification by SourceX, the company's own data inventory, agreed price and terms, buyer review, then delivery and payment. You never handle the records.

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 becomes payable only after the buyer pays and SourceX receives its fee, and no reward is guaranteed.

What this does not mean

• It is not a job-loss forecast. Tasks being learned is different from roles disappearing; many jobs mix automatable tasks with judgment, relationships and physical work.
• It does not make every archive valuable. Thin, templated or client-owned records rarely qualify.
• It does not mean the company loses control. Records are licensed, not sold; the company approves scope, redaction rules and price, and nothing is binding until it signs.

Next step

Choose two companies in your network that match a row in the workflow table and check them against the baseline. Then register as a partner and make the introductions, or share your referral link so each owner can apply at sourcex.si/apply.