Third-Party LLMs and AI Escrow...

During negotiations for an AI-enabled software license, a beneficiary’s attorney may hear a confident argument from the depositor’s side of the table: “The intelligence in this product comes from a large language model we license from another company. We can’t deposit it, so escrow wouldn’t protect you anyway.”

The argument sounds reasonable, and it is incomplete. Counsel without a technical background can be talked out of meaningful protection by it. The model itself may well stay out of the deposit. Nearly everything that makes the product work with that model belongs in it.

This guide explains why commercially available LLMs usually remain outside escrow, what should be deposited instead, and how Technical Verification confirms that a release would give the beneficiary a workable path forward. 

What the Windsurf Episode Revealed About LLM Dependencies

In June 2025, Windsurf, a popular AI coding assistant, announced that Anthropic had cut off nearly all of its first-party access to Claude 3.x models with less than five days’ notice. Anthropic’s leadership publicly linked the decision to reports that OpenAI was acquiring Windsurf. Windsurf responded by shifting users toward alternative models and a bring-your-own-key option, where customers connected their own Anthropic API keys to keep using Claude.

For escrow purposes, the episode is instructive for one simple reason. Claude was never Windsurf’s asset. It was a licensed service, available on commercial terms, and continued access depended on the relationship between two companies. When that relationship changed, the path back to Claude ran through a license held directly by the user.

That is how most commercially available LLMs function inside enterprise software, and it shapes what an escrow deposit can and should contain.

Why Commercially Available LLMs Usually Stay Out of the Deposit

The depositor does not own the model

A software vendor that calls a hosted model through an API holds usage rights under the provider’s terms. Those rights rarely include redistribution, so the vendor typically has nothing it is permitted to deposit.

The model weights are not in the vendor’s hands

Leading commercial models run on the provider’s own infrastructure. The vendor sends requests and receives responses, and never possesses the underlying model files.

The beneficiary can license the model directly

If a release occurs, the beneficiary can generally open its own account with the model provider and procure access on standard commercial terms. The model is replaceable through an ordinary purchase.

In this respect, a hosted LLM resembles third-party dependencies that escrow practitioners already know well, such as a commercial database, a cloud platform, or a payment gateway. Escrow has never required the deposit of those products. It has required the deposit to identify them clearly enough that a beneficiary can obtain them. 

What the Depositor Does Own, and What Escrow Should Capture

The value in an AI-enabled application lives largely in the layer the vendor builds around the model. That layer is proprietary, difficult to reconstruct, and exactly what a beneficiary would need after a release. Deposit materials for an LLM-powered product commonly include:

  • Integration and orchestration code that sends requests to the model, handles responses, and manages errors, retries, and fallbacks
  • Prompt templates and system prompts, which often reflect a long history of product refinement
  • Agent and workflow logic, including tool definitions and multi-step processing chains
  • Retrieval pipelines, such as document ingestion, chunking rules, embedding configuration, and vector database schemas
  • Model configuration, including the provider, model name, version identifier, and parameters such as temperature and token limits
  • Guardrails and output validation, including content filters and business rules applied to model responses
  • Evaluation suites and test datasets used to confirm the product behaves as expected
  • Fine-tuning datasets and related artifacts, where the vendor owns them and the agreement permits deposit
  • Build and deployment documentation, including a list of every third-party service and license required to run the application

When a Model Might Belong in Escrow

Not every AI arrangement follows the “license it on release” pattern. Three situations deserve closer attention.

Proprietary models. When a vendor trains its own model, the weights and training pipeline are the vendor’s intellectual property and are strong candidates for deposit.

Self-hosted open-weight models. When a vendor runs an open-weight model on its own infrastructure, the applicable license may permit deposit of the weights or, at minimum, a precise reference to the exact model and version. Any vendor modifications or adapters should be captured either way.

Provider-hosted fine-tunes. A model fine-tuned on a provider’s platform may not be exportable, and a beneficiary’s new account may not have access to it. In that situation, the training data and fine-tuning configuration become critical, since they may be the only way to recreate the customization.

These distinctions are where an attorney benefits from technical input before agreeing that escrow does not apply.

The Real Release Risk: Undocumented Dependencies

If the model can be licensed on release, the remaining question is whether the beneficiary would know what to license, how to configure it, and how to confirm the result works. Gaps tend to appear in four places.

Model version changes. Providers retire older model versions on their own schedules. An application tuned against one version may behave differently on its successor, so the deposit should record the tested version and include evaluation sets a new team can use to validate a replacement.

Credentials. The depositor’s API keys belong to the depositor’s accounts and may be revoked at release. The beneficiary will need its own credentials, and the deposit should explain exactly where they are configured.

Supporting services. Embedding models, vector databases, moderation tools, and monitoring platforms often sit alongside the primary LLM, each with its own license.

Configuration held outside the codebase. Prompts and parameters stored only in a production environment or a provider dashboard never reach the deposit unless someone asks for them.

None of these are reasons to forgo escrow. They are reasons to verify it.

How Technical Verification Confirms Release Readiness

Technical Verification tests deposit materials against the question that matters at release: could a qualified team rebuild and run this application from what is in escrow? PRAXIS treats a full rebuild as the standard, and that standard brings third-party requirements to the surface early, while the depositor is still engaged and able to close gaps. 

For an AI-enabled application, verification can confirm that the depositor identifies each external model and service the application relies on, records the model versions and configuration needed to reproduce its behavior, and documents the licenses a beneficiary would need to procure upon release.

At Level 3, PRAXIS witnesses, evaluates, and video-captures the depositor building the software from source code, identifying the steps, resources, and artifacts the build requires. The recording is placed alongside the source code, credentials, documentation, and other deposit materials, and PRAXIS provides a comprehensive findings report once the deposit audit is complete. When an application calls a third-party LLM, the build makes that dependency visible, along with the configuration and licensing steps a beneficiary would need to repeat. While a Level 3 test does not prove the completeness, accuracy, or functionality of the deposit, it gives beneficiaries far deeper qualitative insight into what they hold. 

Questions Attorneys Should Ask in AI Escrow Negotiations

  1. Which LLMs and AI services does the application use, and who holds the license for each?
  2. Are any models proprietary, self-hosted, or fine-tuned, and who owns the resulting weights or adapters?
  3. Where are prompts, model configuration, and parameters stored, and will they be included in the deposit?
  4. Which model versions has the application been tested against, and what evaluation materials exist?
  5. What would a beneficiary need to license, configure, and validate to operate the application after a release?
  6. How will the deposit be updated when the vendor changes models or providers?
  7. What level of Technical Verification will confirm the deposit supports a rebuild?

Contract Provisions Worth Considering

Every agreement is different, and counsel should tailor terms to the specific engagement. Provisions that often strengthen AI escrow arrangements include:

  • A third-party dependency schedule listing each model provider, model, version, and license type, updated whenever the integration changes
  • Deposit update obligations triggered by a model or provider change, in addition to regular release cycles
  • Verification rights, including the verification level and how frequently verification occurs
  • Clear treatment of fine-tuning data, adapters, and other customization artifacts

How PRAXIS Supports AI Escrow

Automated Escrow keeps deposits current with the depositor’s development work, so changes to prompts, orchestration code, and model configuration reach escrow without depending on manual uploads.

Technical Verification applies a full rebuild standard, confirming that the deposit, including its documentation of third-party AI services, supports recovery.

Infinite Retention preserves prior deposits, so earlier configurations tied to retired model versions remain available if a beneficiary needs to trace how the application evolved.

SOC 2 certification reflects audited security controls, which matters when deposits contain proprietary prompts, training data, and system architecture.

U.S.-based jurisdiction gives beneficiaries a clear and predictable legal framework for the escrow relationship.

All-inclusive pricing keeps costs predictable for AI products, where frequent model and configuration changes can mean frequent deposit updates.

Keeping the Conversation Grounded

A third-party LLM is one dependency among many. Its absence from the deposit is expected, and it does not make escrow irrelevant. What protects a beneficiary is a deposit that captures everything the vendor built around the model and documents everything the beneficiary would need to license. Attorneys who understand that distinction are far better positioned to keep meaningful protection in place, and Technical Verification gives them the evidence to support it. Contact PRAXIS about AI escrow.

FAQs

Usually not. Commercially available LLMs are licensed services that run on the provider’s infrastructure, and the depositor rarely has the right to redistribute them. A beneficiary can generally license the model directly upon release, so the deposit should focus on documenting the dependency rather than including the model itself.

Escrow protects the proprietary layer the vendor builds around the model, including integration code, prompt templates, agent logic, retrieval pipelines, model configuration, evaluation suites, and build documentation. These materials are difficult to recreate and are what a beneficiary needs to operate the application after a release.

Providers retire model versions on their own schedules, and a successor model may behave differently. A deposit that records the tested model version and includes evaluation datasets allows a new team to validate a replacement model and adjust prompts or configuration as needed.

Technical Verification tests whether the deposit supports a rebuild of the application. For AI-enabled software, it can confirm that external models and services are identified, that model versions and configuration are documented, and that the licenses a beneficiary would need upon release are clearly referenced.

It depends on ownership and portability. A model fine-tuned on a provider’s platform may not be exportable, so the training data and fine-tuning configuration become important deposit materials. Where the vendor owns its model weights or adapters, those are strong candidates for deposit.

Generally no. API keys are tied to the depositor’s account and may be revoked. The beneficiary will typically need to obtain its own credentials from each provider, and the deposit should document where those credentials are configured within the application.

Glossary of Terms

An AI model trained on large volumes of text that generates and interprets language, often accessed by software applications through an API.

A model offered by a provider to the public or to businesses on standard commercial terms, typically hosted on the provider’s infrastructure.

The learned parameters that define how a model behaves. For hosted commercial models, weights remain with the provider.

A model whose weights are published for download and use under a specific license, allowing organizations to run it on their own infrastructure.

The process of further training an existing model on specialized data to adapt its behavior for a particular use case.

Structured instructions an application sends to a model to shape its responses. These are often proprietary and central to product behavior.

An approach in which an application retrieves relevant documents or data and supplies them to the model alongside a request.

An arrangement in which a user supplies their own API credentials to access a third-party model through an application.

Any external software, service, or license an application requires to build or run, which a beneficiary may need to procure separately upon release.

A structured evaluation of escrow deposit materials that confirms whether they are sufficient to rebuild and support the application.

Praxis Editorial Team

Praxis Editorial Team Author

Chris Smith is the Founder and CEO of PRAXIS Technology Escrow and a recognized leader in software and SaaS escrow with more than 20 years of industry experience. He pioneered the first automated escrow solution in 2016, transforming how escrow supports Agile development, SaaS platforms, and emerging technologies.

Leave a Comment