A senior AI leader at a large private-markets firm recently explained why they were reluctant to work with an external AI partner. Simply put, their investment process was too strategically important to outsource. The firm needed to retain its own judgment, ways of working, and intellectual property.

That instinct is right. But in the same conversation, they acknowledged that only one person is focused on AI per asset class, and most of the AI workflows were built by people without software engineering backgrounds. They had no systematic way to test whether those agents produced reliable work as the workflow instructions, data, or underlying models changed.

This is the tension many financial institutions face today. Their investment process is too valuable to surrender, but turning it into reliable, scalable AI requires capabilities they cannot sensibly build everywhere at once.

The mistake is treating “build versus buy” as a single decision.

Ownership and operation are different questions

For a private-markets firm, the differentiating assets are not the underlying models. Every competitor can access the same intelligence broadly.

Institutions should own where their Intellectual Property sits:

• The firm’s data and institutional history

• Its workflow logic and decision rules

• Its definition of what "good" looks like

• The exceptions and judgment accumulated by experienced investors

• Its ability to change or move the resulting workflows

But ownership does not mean the firm must also build and maintain every piece of infrastructure required to make that knowledge usable. Model selection, workflow versioning, testing infrastructure, permissions, and deployment machinery are necessary but rarely the source of an investment firm’s advantage. Rebuilding them internally may create more technical ownership while consuming the scarce people who should be encoding what actually differentiates the firm.

A credit investor should know how the firm underwrites risk. It does not follow that the firm should build the machinery required to run every step of that process reliably.

The fear of outsourcing is justified

External dependence creates real risk. A vendor can learn how the firm works, become embedded in important processes, and make switching "ways of working" difficult. Conversely, a vendor may agree that the institution owns its data, workflow logic and resulting intellectual property while still creating practical dependence. If the workflow only runs on the vendor’s infrastructure, only the vendor’s specialists can modify it, and changing providers requires rebuilding it from scratch, the institution may own the asset legally without controlling it operationally.

The Bank of England and FCA found that one-third of financial services AI use cases were third-party implementations. At the same time, 46% of responding firms said they had only a partial understanding of the AI technologies they used. Respondents expected third-party dependencies to be among the fastest-growing risks over the following three years. Bank of England and FCA, November 2024)

The answer, however, is not to reject external capability. It is to design partnerships around institutional ownership, which requires more than an IP clause. The institution should be able to:

• Understand how its workflow and decision rules have been encoded

• Retain its data, instructions and definition of good

• Develop enough internal capability to supervise and change the system

• Export the relevant workflow assets in a usable form

• Move to another platform without reconstructing its institutional memory from scratch

The PRA already applies a similar principle to outsourcing more broadly: firms can use third parties, but they cannot outsource their accountability. They must understand their respective responsibilities, protect their data, and retain enough capability to supervise and, where necessary, exit the arrangement. PRA Supervisory Statement SS2/21, updated November 2024

That is a better model for enterprise AI than a binary choice between buying a black box and rebuilding the entire stack.

The third option is co-building

A credible co-build model has four parts.

The institution owns the judgment. Its data, decision rules, workflow logic, and definition of good remain its property. Ideally, it can also inspect, modify, and move the resulting workflows.

The partner supplies scarce execution capability. Domain-technical experts work alongside the firm’s people to observe how work is performed, capture undocumented exceptions, and translate that knowledge into reliable software.

Commodity infrastructure remains shared. The institution avoids rebuilding machinery that does not differentiate it, while retaining enough transparency and portability to prevent dependency from becoming captivity.

Dependence falls rather than rises. The institution develops the ability to understand, supervise and modify what has been built. Its workflows and institutional memory remain portable enough that changing providers does not mean starting again.

This is not the same as outsourcing the investment process. The investment team continues to determine how decisions are made. The partner helps convert that judgment into additional operating capacity.

A credible co-build partner has to accept the uncomfortable test if the relationship is becoming more valuable over time because the institution’s capability is compounding, not because leaving is becoming progressively harder.

This does not require every component to be immediately replaceable, complex systems always have switching costs. But those costs should come from the genuine difficulty of the work, not from deliberately trapping the institution’s knowledge inside proprietary infrastructure.

Even the foundation-model companies are moving toward this structure. In February 2026, OpenAI paired its engineers with major consultancies to help enterprises embed AI into core workflows. OpenAI said the objective was for customers to eventually become self-sufficient, rather than remain permanently dependent on external implementation teams. Reuters, February 23, 2026

Building a Lasting AI Partnership

Investment leaders should ask more than whether a solution is internally built or externally supplied.

• Will our judgment remain ours?

• Will our institution become more capable through the engagement?

• Can we improve the workflow without waiting indefinitely for the vendor?

• Could we move it without reconstructing our institutional memory from scratch?

• Are our scarce internal people working on our differentiation or rebuilding generic infrastructure?

The best partner should make itself less indispensable at the infrastructure level while making the institution’s accumulated knowledge more valuable over time.

The strategic question is: what must we own because it differentiates us, and what are we operating ourselves only because we have confused ownership with control?