Private equity and portfolios
Build an AI product on portfolio data or license the records out?
By SourceX Editorial · Updated
Short answer
License the records out first when a portfolio company holds deep, linked history but has no funded AI product team; build an AI product first when it already owns a workflow customers pay for. Building needs engineers, live data and paying users; licensing needs clear rights and documented history. A non-exclusive license can keep the build option open.
Key takeaways
- Building an AI product spends engineering and infrastructure money before any revenue; licensing spends rights review and preparation effort instead.
- Licensing usually comes first for companies with deep linked history, no AI product team, and records more useful to model developers than to their own customers.
- Building usually comes first for software companies whose records feed a workflow they already own and sell.
- A non-exclusive, field-limited license with a reserved rights clause can keep the company free to build later.
- Customer contracts may allow internal use to improve a service but not disclosure to a third party, so each path needs its own rights check.
What is the real choice between building and licensing?#
The real choice is who turns a portfolio company's records into something useful. Building means the company trains or tunes models on its own records and sells the result as a feature or product. Licensing means an AI developer receives a prepared, time-limited license to a defined set of records and builds with them, while the company keeps ownership.
The two are often framed as rivals, but they rarely need the same records in the same way. A product team needs live, current data flowing through its own software. A model developer licensing records usually wants deep, linked history: support tickets with resolutions, jobs with callbacks, orders with the exceptions that followed.
For an operating partner the useful question is sequencing across the portfolio. Which companies have a credible build path, which have a licensing path, and which have neither yet?
Build or license: the decision table#
The build-or-license decision table compares the two paths on the factors an investment committee will raise. Neither column wins in general; the answer depends on each company's team, records and rights.
Rights are the row most often misjudged. A customer contract that lets the company use data to improve its service may support an internal AI feature yet forbid disclosure to a third party, and the reverse also happens when a contract allows sharing of de-identified data but limits new product uses. Check each path against the actual clauses rather than assuming one answer covers both.
| Factor | Build an AI product | License records out |
|---|---|---|
| Upfront cost | Engineering, model and infrastructure spend before launch | Rights review and privacy preparation, mostly people time |
| Time to cash | After launch, adoption and any pricing change | On contract terms with a buyer, once preparation and approval are complete |
| Rights needed | Right to use records to provide or improve the company's own service | Right to disclose prepared records to a third party for a stated permitted use |
| Team needed | Product, data and machine learning staff, plus support for the feature | A project owner, IT for exports, counsel for the rights review |
| Main risks | Cost overruns, weak adoption, model errors in front of customers | Misread contract rights, customer perception, exclusivity limits |
| What the company keeps | The product, the model and the records | Ownership of the records and any rights the license reserves |
| How it looks at exit | Product capability and recurring revenue | License revenue plus obligations diligence will review |
When should licensing come first?#
Licensing should come first when a company's most valuable asset is its history rather than its software. That is the common case for operators such as trades, distribution and manufacturing businesses, and for software companies whose AI roadmap is unfunded.
Rule of thumb: if a company cannot name the people who would build and support the product, license first and reserve the right to build. Revisit the build question once the license has shown which record families outside buyers actually value, because that answer also tells a product team where the depth is.
- The company holds several years of linked records, such as tickets with resolutions or jobs with callbacks, that are no longer used day to day.
- No product team could ship and support an AI feature within the hold period.
- The records would be more useful to a general model developer than inside the company's own narrow product.
- Customer and vendor terms allow disclosure of prepared records, or the company-owned internal records are rich enough on their own.
- The board wants evidence of what the records are worth to outside buyers before committing capital to a build.
When does building come first?#
Building comes first when a company already owns a workflow its customers pay for and AI would make that workflow better or harder to replace. A vertical software company whose product is the system of record for its customers is the clearest case, because the AI feature rides on distribution the company already has.
Building also comes first when the records' value depends on exclusivity. If a dataset is what would let a competitor catch up, licensing it, even non-exclusively, may weaken the company's position. That judgment belongs to the CEO and the board and should be made before any licensing conversation starts.
Be honest about the team. Many portfolio companies can switch on a vendor's AI feature but cannot train, evaluate and support their own models. Buying a feature is not building one, and it does not turn the company's records into an asset.
Can a portfolio company do both?#
A portfolio company can do both if the license is drafted with the build in mind. Licensing older history while the product team builds on current data keeps the two paths from colliding. Where a license would cover the record families the product depends on, counsel should review whether a restriction on competitive use is needed.
| License term | How it protects a later build |
|---|---|
| Non-exclusive grant | The company can use or license the same records again |
| Field-of-use limit | The buyer's use stays outside the company's own product market |
| Fixed term with end-of-term handling | Obligations do not run open-ended into a future exit |
| Reserved rights clause | Confirms the company keeps every right not expressly licensed |
| Defined record scope | Live product data and current customer data can stay out of the package |
Illustrative: two portfolio companies, two answers#
Illustrative: a fictional sponsor owns a field service software vendor serving elevator maintenance contractors and a regional HVAC equipment distributor. The operating partner asks both CEOs the same question: build an AI product, or license records out?
The software vendor has a product team, owns its customers' service workflow and holds years of maintenance tickets linked to parts and outcomes. Its customer contracts allow aggregated use to improve the service but say little about disclosure. It builds a fault triage feature first and parks licensing until counsel has reviewed those contracts.
The distributor has no product team and no plan to hire one. Its NetSuite records and inside sales email hold years of order exceptions, substitutions and technical questions answered by experienced staff. It runs a metadata-only fit check, completes a rights review, and licenses a prepared subset non-exclusively, reserving the right to use the same records in any future internal tool.
How SourceX approaches the build-or-license question#
SourceX works only on the licensing side; it does not build AI products, and its own rights in a deidentified dataset are set out in the signed supplier agreement. The Supply and Rights steps of the SourceX five-step transaction establish which record families a company holds and whether they can be disclosed, which is often enough for a CEO to see whether licensing sits comfortably beside a build plan.
The SourceX Enterprise Data Value Framework, a SourceX-developed methodology with qualitative ratings, compares record families on drivers such as uniqueness, domain expertise, human-generated signal, recency, rights and AI utility, while preparation cost and privacy burden count against net value. The supplier approves every step, and a license scope can exclude any record family the product team wants to keep for itself.
Frequently asked questions
Will licensing records make it harder to build our own AI later?
Not if the license is scoped carefully. A non-exclusive grant with reserved rights leaves the company free to use the same records in its own products. The terms to watch are exclusivity, broad field-of-use language and any clause limiting competition with the buyer, which counsel should review before signing.
Who decides between building and licensing: the sponsor or the portfolio company?
The portfolio company decides, because it holds the records and signs any license or product commitment. The sponsor shapes the choice through the board, the value creation plan and capital allocation. In practice the operating partner frames the options, and the company's CEO and board choose.
Is switching on a vendor's AI feature the same as building on our own data?
No. A vendor's AI feature can improve operations, but the vendor's model is the asset, not the company's records. The feature may also send records to the vendor under terms that permit training, so check the settings before treating a vendor feature as the company's AI strategy.
Can records used to train our own model still be licensed?
Usually yes. Training an internal model does not use up the records, and buyers license prepared records, not the company's model. The rights review still applies, and the company should decide whether its model and the license overlap in a way that weakens its competitive position.
What if neither path fits a company today?
Park it and fix the basics: a system register, company-controlled admin access and a retention schedule. Circumstances change. A system retirement, an add-on that deepens history or a new product hire can move a company onto one path later, and the groundwork is the same for both.
Related resources
See if your company qualifies
A short company assessment. No data uploads are needed.