Private equity and portfolios
AI value creation in portfolio companies: start with operating records
By SourceX Editorial · Updated
Short answer
AI value creation in private equity portfolio companies should start with operating records, not tools. Screen each company's support, sales, project, job and quality records for history, linkage and rights first. That one screen shows which internal AI projects are feasible and whether the same records could also be licensed to AI developers.
Key takeaways
- The records a company keeps decide which AI projects are feasible, so screen records before choosing tools.
- One record screen informs two decisions: internal AI projects and external data licensing.
- Internal projects aim at cost, speed and quality; licensing can add a revenue line while the company keeps ownership.
- Records that show a request, a decision and an outcome are the most useful for both.
- Rights and approvals are checked company by company, with the sponsor coordinating rather than signing.
Why should AI value creation start with operating records?#
AI value creation should start with operating records because every AI use case, internal or external, depends on what a portfolio company has retained and is allowed to use. A support assistant needs resolved tickets. A quoting tool needs past quotes linked to won and lost orders. A scheduling model needs job histories with real durations and outcomes.
Portfolio AI efforts that begin with tool pilots, such as chat assistants or copilots, often stall when the underlying records turn out to be thin, scattered across retired systems or restricted by contracts. Starting with records changes the first question from which vendor to buy to which work this company documents well.
The record-first framework in four stages#
The record-first framework runs in four stages: inventory, screen, choose and govern. Each stage can be run across several companies at once with the same questions, which makes the results comparable for an operating partner or value creation lead.
None of the stages needs data to leave the company. The first two work from metadata and conversations with each company's COO, CTO or IT lead.
- Inventory: list each company's systems and record families, such as tickets, CRM activity, jobs, orders, RFIs, NCRs and code reviews.
- Screen: rate each record family on history, linkage, rights, sensitivity and how easily it can be exported.
- Choose: match strong record families to internal AI projects, and flag those that could also be licensed.
- Govern: set an AI use policy, an approval path for each company and a reporting line to the sponsor.
What does the record screen look at?#
The record screen looks at six dimensions that decide whether records can support AI work. Ask the same questions of every company and record unknown answers as unknown rather than guessing.
Linkage carries the most weight. A ticket that links to the engineering issue, the fix and the customer's reply teaches far more than a ticket that stops at closed.
| Dimension | What to ask the company | What a strong answer sounds like |
|---|---|---|
| History | How many years of these records can you still open or export? | Several years, in the current system or a preserved archive |
| Linkage | Can you follow one request through to its outcome? | Ticket to issue to release, or estimate to job to invoice to callback |
| Coverage | Do the records capture the real work or only the paperwork? | Notes, comments and decisions, not just status fields |
| Rights | Which customer contracts, vendor terms or notices might limit use? | Known, listed and mostly company-owned internal records |
| Sensitivity | How much personal, health or financial detail is mixed in? | Limited and removable without destroying meaning |
| Access | Who can export these records, and in what format? | An admin who has run full exports before |
Where the screen usually points, by company type#
The screen usually points to different record families depending on the business model. The same company can hold one family suited to internal tools and another suited to licensing.
| Company type | Records that usually matter | Internal AI use | Licensing angle |
|---|---|---|---|
| B2B or vertical software | Jira or Linear issues, GitHub reviews, Zendesk or Intercom tickets | Ticket triage, release notes, code review help | Linked support-to-fix histories |
| Engineering or architecture firm | RFI logs, submittals, review comments in Procore or Bluebeam | Drafting RFI responses, QA checklists | Internal review reasoning, with client deliverables carved out |
| Home services or trades | ServiceTitan or Jobber calls, estimates, jobs, callbacks | Call booking help, estimate options | Linked job histories from call to warranty |
| Distribution or 3PL | Orders, EDI messages, exception notes in WMS or TMS | Exception routing, order entry | Exception-to-resolution records |
| Manufacturing | NCRs, CAPAs, maintenance work orders, quote revisions | Quality triage, maintenance planning | Quality and maintenance decision records |
How licensing sits beside internal AI projects#
Licensing sits beside internal AI projects as a second use of the same inventory, with a different goal and a stricter rights test. Internal projects aim to cut cost, speed up work or improve quality inside the company. Licensing provides prepared records to AI developers under a contract, can add a revenue line, and leaves ownership with the company.
The rights questions differ. Using records inside the company is mainly governed by its own policies, customer confidentiality terms and privacy notices. Licensing to a third party adds permitted-use, consent and exclusivity questions, which counsel assesses deal by deal. A company can pursue both, either or neither, and a few decision rules keep the choice consistent across the portfolio.
- Strong records, clear rights: run the internal project and request a metadata fit check for licensing in parallel.
- Strong records, restricted rights: keep the records for internal use only and log the restriction in the AI register.
- Thin or unlinked records: fix how the work is captured first, such as requiring resolution notes before a ticket or NCR can close.
- Records at risk in a system retirement: export full history before deciding anything, because neither use is possible once it is gone.
- Records that differentiate the business, such as pricing logic: decide whether to exclude them from any license before scoping starts.
Illustrative: a lower-middle-market fund applies the record screen#
Illustrative: a fictional fund owns a property management software company, a commercial roofing contractor and a specialty industrial distributor. Its first AI plan was a copilot rollout across all three. The value creation lead pauses it and runs the record screen instead.
The software company holds several years of Zendesk tickets linked to Jira issues and GitHub pull requests. It picks ticket triage as an internal project and requests a metadata fit check for licensing the linked support-to-fix history. The roofing contractor's inspection photos and estimates suit an internal estimating aid, but counsel finds that its subcontracts give general contractors rights over many project files, so licensing is parked. The distributor's order exceptions are rich, but older history sits in an ERP retired before the acquisition; the team recovers an archive before deciding anything.
The fund ends with two internal projects tied to records that exist, one licensing fit check, and one archive saved.
Governance the sponsor should set early#
Sponsor governance should set who approves AI use at each company and what gets reported upward, without moving decisions away from the people who hold the records. Each portfolio company's CEO and authorized signer decide what is licensed; the sponsor coordinates and may need to consent.
Write a short AI use policy that covers which tools may touch customer records, how outputs are reviewed, and when legal review is needed. Review loan agreements and fund documents for restrictions on IP licensing before any licensing discussion, and keep a register of AI projects and licenses for the board and for exit diligence.
How SourceX fits a record-first AI program#
SourceX handles the licensing side only. It does not build a company's internal AI tools, and its own rights in a deidentified dataset are set out in the signed supplier agreement. For companies the screen flags, SourceX runs a fit check on metadata, then the SourceX five-step transaction: Supply, Rights, Preparation, Approval and Delivery, and the company signs off at every stage.
The SourceX Enterprise Data Value Framework gives operating teams a shared way to describe record depth and linkage, so the licensing view and the internal AI view use the same language.
Frequently asked questions
Do portfolio companies need a data warehouse before starting?
No. The record screen works from descriptions: which systems hold records, how far back they go, what kinds of records exist and any restrictions leaders already know about. Many useful records sit in the systems of record themselves, such as the help desk, FSM or ERP. A warehouse can help later projects, but it is not a prerequisite for deciding where AI work should start.
Will licensing records weaken a company's competitive position?
It can if scope is careless, so scope is decided deliberately. Companies can exclude pricing, customer names and proprietary methods, choose non-exclusive terms, and limit permitted use. The question is a judgment for each company's leadership, weighing how much the records differentiate the business against the value of licensing them.
Who should own AI value creation at the sponsor?
Usually the operating partner or value creation lead owns the program, with each company's CTO or COO owning execution. The group CFO joins for deal structure and accounting, and counsel joins for rights review. Keeping one owner at the sponsor prevents each company from running unconnected pilots.
Can internal AI tools and licensing use the same exports?
Partly. Both start from the same systems, but a licensed package goes through preparation that removes personal and confidential details and documents provenance and permitted use. Internal tools follow the company's own access controls. Keep the two flows separate so internal convenience never becomes an unreviewed external release.
What should portfolio CEOs hear first about this?
That the first step asks for descriptions, not files, and that their company approves anything licensed. CEOs worry about effort, brand risk and control. A short metadata questionnaire, a clear statement that ownership stays with the company, and one pilot company make the program concrete before asking for broader time.
Related resources
See if your company qualifies
A short company assessment. No data uploads are needed.