The Evidentiary Questions Missing from AI Build-or-Buy Decisions

Organizations that decide to rent or own their AI models based on cost, speed, and IP risk leave out a critical factor: what they can prove when something goes wrong.

The biggest and most conservative institutions in business are forgoing big contracts with artificial intelligence (AI) companies in favor of building their own AI infrastructure.

BNP Paribas built a large language model (LLM) as a service hosted on servers in its own data center. Kirkland & Ellis committed $500 million over three to four years to build its own AI platform. Latham & Watkins purchased Nvidia servers on which it plans to build in-house AI models.

Some of this stems from the falling price and rising quality of open-weight models (not to be confused with open-source models); building from scratch has become a viable option for organizations.

Most of the conversation around these “rent versus own” decisions has focused on cost, scaling internal use, and intellectual property (IP). While those are critical factors, hardly anyone is discussing another consideration: litigation risk.

When you license an off-the-shelf AI model, your records—prompts, logs, access history, configuration—will be stored on someone else’s servers, under their control, as governed by their contract terms. When you build your own AI model, you gain full control over your data but also take on the obligation to protect and preserve the record and defend the ways you handle and retain the data.

When a fraud surfaces, cybersecurity breach occurs, regulator knocks, or dispute lands in court, where you landed on rent versus own will help determine what you can know—and what you can prove.

What Do You Give Up When You Rent Your Model?

You can and likely would build a governance layer around a rented AI stack, which can offer some control over records and retention but not over areas like model routing, explainability, or training provenance. These distinctions will matter in an audit or investigation.

• You don’t control access. Renting your AI stack involves more than a single vendor. You pay a subscription to an AI lab, which buys its compute from a data-center operator. The model you rent will connect to your existing software layer. You might contract with a developer to build and configure the application your team uses. Each vendor touches your data at various points. None are responsible for what others do. When misconduct surfaces, you might have to rely on an outside party to provide a record—if you can determine where in the chain the misconduct happened.
• You don’t set contract terms. To handle access, your best option would be to draw clear, specific lines in the contract. But that’s highly unlikely when your vendor is a frontier AI lab. These labs have their own policies and procedures around data access, which likely are negotiable only with the largest corporate customers. That includes terms on data retention, which can have profound implications for investigations and disputes.

What Do You Take on When You Own?

The hard costs of owning your AI model are obvious: infrastructure up front, then maintenance, staffing, and engineering. Less obvious is accountability. While you’ll have full control over your AI model, data, and retention policies, you won’t necessarily have full visibility into your data.

• You must defend the model. You own the technology and every decision made in building it—including which data you trained it on and its ability to replicate results. When you rent a frontier model, often times you adopt an industry standard that provides a certain degree of transparency and disclosures. When you own, you’re on the hook for everything.
• You can’t see everything. To be useful, an AI agent running on your own infrastructure will interface with application programming interfaces (APIs), software-as-a-service (SaaS) tools, vendor platforms, and other companies’ agents. Much of what happens in those interactions will live on outside infrastructure. Owning will make it easier to know and prove what happened at the model layer, but it won’t give you perfect visibility into everything that happened.

How Long Are Your AI Audit Logs Retained?

Every rent-versus-own decision should involve some consideration of retention: the policies and procedures that will define which records are available in an investigation.

In the pre-AI digital age, investigators could piece together who did what, when, and why by analyzing the logs an organization kept on its servers. The audit log could reveal actions and behavior down to the keystroke.

AI audit logs serve a similar purpose for investigators. They can establish intent, prove illicit behavior, or show how a fraudulent result was achieved. But reviewing an AI audit log is more complicated than reviewing traditional logs. Because AI is nondeterministic, because its outputs vary by context, and because models regularly evolve and sometimes degrade, the same prompt could produce different outputs from one user to another or one week to the next. It is critical to ensure appropriate retention periods are established for audit logs.

The accepted standard for retaining logs has long been ninety days. But under pressure to protect customers’ data privacy, frontier labs have been shrinking the window as a competitive tactic. Many frontier model providers

One of the earliest questions we ask a client in an investigation is: “What’s your retention policy?” They often don’t have an answer. In other cases, they tell us it’s specified as thirty or sixty days but has never been confirmed in practice.

An organization that chooses to build its own AI model should establish a thoughtful, consistent, defensible retention policy from the start. Those who choose to rent need to understand that your AI provider effectively will set your retention policy—and that your evidentiary considerations aren’t their priorities. In that situation, you must manually save prompts and responses after every interaction—a practice that is both onerous and error-prone.

Whether you rent or own, you need to understand the impact your choices will have on what you will be able to prove.

“Good Enough,” Eventually?

As quickly as AI is changing, making predictions is a fraught business. But planning must go on, investments must be made, and some degree of tea-leaf reading will be necessary. Here’s one guess at where this could be headed:

In the past, new technology tended to settle on an industry standard: a product or system that costs less and performs worse than the highest-quality option but is deemed acceptable by the industry and then by the courts. In eDiscovery, that standard was proportionality: holding the scope and cost of discovery to a scale matching what’s at stake in the litigation. Courts have accepted a few unturned stones in the name of keeping matters moving.

A similar quality-floor consensus could form around AI models. Once companies broadly settle on common technology, they may also settle on the degree of opacity that comes with those ubiquitous models. At some point courts may begin to view widespread adoption as its own defense.

How the plaintiffs’ bar views it is a different question entirely. We would expect them to push back against treating any corporate decision as acceptable by default, especially if it curtails their ability to gather evidence.

Again, this is just a prediction. Renting your AI model eventually may lower the bar for retention and data protection. But we’re not there yet, and there’s no telling how far that protection may go.

What to Do Now

In BRG’s disputes and investigations work, we recommend four steps for general counsel and compliance leaders who have a voice in significant AI investment decisions:

Bottom Line

Cost, speed, and IP risk will drive the rent-versus-own decision for most organizations. These factors dictate how AI will work for the company but not what will be provable when the work is questioned in court. That should be part of the decision as well.