Build vs. Buy AI: What Institutional Investors Should Build, and What They Should Buy

Key Takeaways

• The AI models are a rental, available to every allocator at roughly the same time and price, so building the model layer is a treadmill rather than a moat
• The durable advantage is the data, specifically the ability to normalize and reconcile private-markets data to an audited record, and that layer is the clearest buy in the entire stack
• What is worth building is judgment: the portfolio-specific questions, pacing assumptions, and scenario logic no vendor can replicate
• Four objections stall most buy decisions: lock-in, security, uniqueness, and the belief that your data is your IP. Each has a direct answer
• Two questions settle most cases: is this capability genuinely proprietary to us, and can we keep it current for as long as we depend on it

The question usually arrives from the board, phrased the way boards now phrase it: what is our AI strategy? Behind it sits a decision with real budget and real governance exposure. Build an AI capability in-house, or buy one from a specialist. Most teams treat that as the whole decision, and it is the wrong place to start.

Start instead with a distinction that resolves most of the question on its own. The AI models are a rental, leased by every allocator on the same terms. Your data is the asset, and turning it into something an AI system can reason over is where the real work sits. Build vs. buy for AI is really a question about which layer of that data problem is yours to own.

The Model Is a Rental

Consider what commoditizes. Large models improve on a cycle measured in months, and each improvement reaches everyone at roughly the same time and price. A capability you build around a specific model this year will sit behind a better one by the time it reaches production, and you will have paid for the scaffolding twice. Owning the model layer is a treadmill, and every peer is running on equivalent equipment. The same is true for Solovis as for any institution: no one, including us, owns a durable edge at the model layer. The advantage lives in what surrounds it.

Your data does not commoditize: no one else has it, and getting it into a form a model can actually use is hard. That scarcity is where an edge can exist. The question worth arguing about is which part of the data problem an institution should own, and for a portfolio with meaningful private-market exposure, the answer is clearer than most build proposals make it look.

The Private-Markets Problem Is the Whole Argument

Look at what a private-markets book does to the data question. Valuations arrive quarterly, often a quarter or two late. Capital calls and distributions land on no fixed schedule. The underlying detail comes from dozens of general partners, each reporting in its own format, none of which lines up with the others or with how public managers and custodians report. Before any AI layer can say something useful about total portfolio risk or pacing, all of that has to be ingested, normalized, reconciled, and tied back to a record you can audit.

Picture a lean team at a pension fund with roughly $10 billion in assets and 40% in private markets, illustrative but familiar. They decide to build their own data-integration layer, because the data is theirs and the logic feels specific to them. A senior hire and 18 months later, it works. Then two general partners change their reporting templates, a third is acquired and re-papers its statements, and the reconciliation logic breaks quietly during a quarter-end close. No one notices until a private-markets figure in the board book does not tie. The build was never the hard part. Keeping it correct as the inputs shift underneath it is the hard part, and it is a job that does not end.

Institutions misjudge this piece when they scope a build. The model is the visible piece, so it draws the attention and the budget, while the integration underneath gets filed under plumbing and deferred. In a private-markets portfolio it is most of the work, most of the cost, and most of the risk. A build also creates a silo. The layer lives apart from the rest of the portfolio view, so even when it works, its output has to be reconciled by hand against everything else, and the institution ends up maintaining a second version of the truth. Building it buys no advantage: it is a capability every allocator needs and none competes on, so a vendor specialist spreads the maintenance across its whole client base while a single institution carries it alone, usually with a team already struggling to stay staffed. Across the entire stack, this is the clearest buy, and the one most often mistaken for something worth building in-house.

This is not a contrarian read. In a 2026 survey of roughly 200 asset managers, pension funds, and insurers, the two highest-ranked technology priorities on the buy side were consolidating vendors and modernizing data infrastructure, at 58% and 54% respectively. The market is settling on the blend this framework argues for.

What Is Worth Building

So what should an institution build? The judgment, and nothing beneath it. The questions a particular CIO asks of that portfolio. The pacing assumptions calibrated to a specific spending policy or liability stream. The scenario logic that reflects how a given investment committee weighs risk. None of that comes from a vendor, because it is specific to you, and it is the layer where an internal team's time compounds into something a competitor cannot copy.

Building here rarely means writing models from scratch. It means keeping ownership of your assumptions and of how the outputs get interpreted, so the answers reflect your view rather than a vendor's defaults.

 The Objections Worth Answering

Allocators do not resist buying because the logic is unclear. They resist for four reasons, and each deserves a direct answer.

The first is lock-in. Handing a core capability to a vendor can feel like ceding control, and it becomes exactly that if you cannot get your data out when the relationship ends. The protection is contractual: insist you own your data and can take it with you.

The second is security and fiduciary exposure. A breach carries financial and governance consequences that land on the CIO, and a specialist whose entire business depends on institutional-grade security will invest more in it than a lean team can. Press any vendor on auditability: every figure it produces should trace to a source your compliance function can stand behind.

The third is the belief that your situation is unique. At the judgment layer, it usually is. At the data-integration layer, it almost never is: the mechanics of normalizing a capital call notice are the same everywhere, and treating them as distinctive is how institutions talk themselves into building the one thing they should buy.

The fourth is that your data is your intellectual property. It is, and that is the case for guarding it, not for building the machinery around it. Your IP is the reconciled record and the judgment you apply to it, not the pipeline that ingests a capital call notice, which is cost. Conflating the two is expensive.

 The Test, in One View

When the decision comes up, two questions settle most of it. Is this capability genuinely proprietary to us, or is it something every allocator needs and none differentiates on? And if we build it, can we keep it current, funded, and staffed for as long as we rely on it? Plot any capability against those two axes and the call becomes visible.

Most of the AI stack lands in the bottom row. Nearly all of the private-markets data layer lands in the bottom-right cell, the clearest buy on the board.

Institutional investors navigating this decision need a trusted data foundation before they commit to an AI strategy, because the strategy is only as good as the record beneath it. Solovis gives institutional investors a unified view across public and private holdings, normalizing inputs from custodians, fund administrators, and general partners into a single audited record: the foundation any AI capability has to sit on, and the part of the stack that returns the most time to a lean team. That single record is what a home-built silo cannot deliver: one view across the portfolio rather than one more system to reconcile against the others. The institutions that get the most from AI are the ones honest about where their advantage lives. They buy everything below the judgment layer and put scarce internal talent into the judgment that makes the portfolio theirs.

If your own data foundation looks more like the silo than the single record, we welcome a conversation.

Sources

SimCorp, InvestOps 2026 report (buy-side technology priorities: vendor consolidation 58% and data infrastructure modernization 54%; 200 executives at asset managers, pension funds, and insurers, each managing at least $10 billion in AUM; more than two-thirds using AI in the front office). https://www.simcorp.com/about-us/news/2026/two-thirds-managers-adopt-AI

 Talk to a Solovis expert.