Back-office AI agents in portfolio companies: the records that train them

What back-office AI agents actually do

A back-office AI agent takes a goal, works through several steps across business systems and hands exceptions to a person. Where a chatbot drafts an answer, an agent might read an unmatched invoice, look up the purchase order in the ERP, email the vendor and route the case to an approver when the amounts still disagree.

Operating partners meet these tools in vendor pitches, portfolio AI roadmaps and budget requests. The table shows where they appear most often in a mid-market back office and what each one depends on.

How agents learn the work

Builders need records of how the work was really done, not descriptions of it. The useful material comes in five layers, and the more layers a company keeps for the same case, the more useful the case becomes.

1. The starting situation. An invoice with no PO, a customer disputing a charge, a ticket about a failed integration.
2. The step trail. Who did what, in which system and in what order, as shown in audit logs, ticket histories and email threads.
3. Decisions and reasons. Approval comments, exception notes and escalation messages that explain why a person chose one path.
4. The outcome. Paid, credited, reversed, resolved, reopened or written off.
5. The rules in force. SOPs, approval matrices and delegation-of-authority documents that set the boundaries.

The same records serve evaluation as well as training: a builder can hold back real cases and check whether an agent reaches the outcome a skilled employee reached. Records like these are thin on the public web. Epoch AI researchers project that, if current trends continue, language models will fully use the stock of public human-written text between 2026 and 2032, a forecast with wide uncertainty. The US Copyright Office's report on generative AI training, released in pre-publication form in May 2025, also notes that model performance depends heavily on data quality.

Why this matters for an operating partner

The pressure to show operational gains is not easing. McKinsey's Global Private Markets Report 2026 finds that multiple expansion and cheap leverage, which accounted for 59 percent of PE returns between 2010 and 2022, have faded, making operational value creation likely the primary source of returns, and that sponsors are applying AI to operating levers.

Most of that work sits on the buy side: choosing which agents to deploy. The AI value creation playbook and the AI use case prioritization framework cover those choices. The less obvious point is that the records an agent rollout depends on can also be an asset in their own right.

Both columns reward the same housekeeping: clean ticket histories, coded exceptions and preserved archives. A portfolio company that has invested in one is usually better placed for the other.

Which portfolio companies sit on the supply side

The fit test is about the depth and ownership of records, not about whether the company has bought an agent yet.

• Scale and age. US companies with 50+ full-time employees at peak (contractors excluded) and several years of documented operations.
• Recorded workflows. Payables, receivables, support and procurement run through systems with histories, not through personal inboxes alone.
• Outcome codes. Tickets closed with resolution codes, disputes closed with a credit or payment, requests approved or rejected with reasons.
• Breadth. Records across many systems, from email and chat to ERP, CRM and service desk tools.
• Rights. The company created the records itself and its contracts allow licensing.
• A sponsor. An owner, CEO, CFO or authorized representative open to an exclusive AI-training license for an agreed term.

Distribution, logistics, IT services, B2B software and professional services back offices often screen well. Platforms that have absorbed several add-ons may also hold M&A integration records, a related set of multi-step project histories.

What it means for a referral partner

Your role ends at the introduction. You share your referral link or submit the company through the referral form; SourceX then confirms fit, the company builds a data inventory, price and terms are agreed, AI labs and data buyers review the opportunity, and the company delivers data and is paid once the agreement is signed. You never touch 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, and rewards become payable only after the buyer pays and SourceX receives its fee. An introduction or meeting alone earns nothing, and rewards are not guaranteed. The reward is a share of SourceX's fee and never reduces what the portfolio company receives.

The network opportunity finder is a quick way to list which portfolio and network companies to raise this with first.

Limits and open questions

• One-time, not recurring. Treat a license as non-recurring cash and do not model repeat deals.
• Rights come first. Support tickets and dispute threads often contain customer details. The guide to data rights documentation for PE-backed companies sets out what to gather.
• Exclusivity terms. Deals are typically exclusive for AI training for an agreed term. Ask early how those terms interact with the company's own internal AI projects; that is settled in the agreement, not assumed.
• Forecasts are forecasts. Projections about public data running short carry wide ranges, and agent adoption inside any one company may stall.
• Not every archive survives. Companies that cancelled old tools without an export may have little history left to offer.

Next step

Pick one portfolio company whose back office runs on well-kept systems and confirm it meets the baseline on who qualifies. Then register as a partner and introduce the CEO or CFO, or share sourcex.si/apply with your referral link so the company can apply itself.