AI Exit Clauses and Model-Weight Escrow: What Happens When You Need to Leave

See also (wiki): ai-vendor-lock-in · ai-vendor-contracts · ai-platform-selection

Executive Summary

• Most enterprise AI contracts lack meaningful exit provisions. OpenAI, for example, does not include fine-tuned model portability, termination for convenience, or model-change notification in its standard enterprise terms — all must be negotiated separately.
• Once customer data is embedded in model weights through fine-tuning, clean extraction is technically impossible. The legal and practical answer is escrow: quarterly deposits of fine-tuned weights and training datasets with a neutral third party.
• Morgan Lewis identifies four categories of customer-owned AI artifacts that contracts must address at termination: input data, customer-developed artifacts (prompts, embeddings, retrieval indexes, guardrails), outputs, and deletion/return obligations with written certification.
• Vendor bankruptcy creates a worst case: court-appointed trustees can auction fine-tuned models containing customer data to competitors. At least one documented case involved a fintech losing six terabytes of training data when a vendor shut down mid-quarter with no backup provisions.
• The mitigation stack is straightforward but rarely implemented: escrow with quarterly updates, data segregation into customer-specific cloud buckets, survivability clauses with perpetual licenses, and a 90–180-day transition window as a contractual floor.

The Lock-In Mechanics Most Buyers Miss

Enterprise AI vendor lock-in operates through three channels that compound over time. The first is explicit contractual lock-in: multi-year minimum annual commitments (MACs) where early termination means paying the full remaining balance. The second is auto-renewal traps requiring 90-day advance notice — miss the window and the contract renews automatically. The third — and hardest to undo — is integration lock-in: once business logic calls vendor-specific APIs with vendor-specific prompt formats, switching costs become a rewrite, not a configuration change.

OpenAI’s standard enterprise agreement illustrates the pattern. According to a 2026 procurement analysis, the platform does not automatically provide model-change notification with adequate notice, audit rights over billing and usage data, price protection for committed discounts if pricing changes, termination for convenience rights, portability of fine-tuned models, or commitments restricting competitive intelligence use of aggregated usage data. Every one of these must be negotiated as an additional provision.

The market is responding. Menlo Ventures data (late 2025) shows OpenAI dropping to 27% of enterprise LLM API spend from approximately 50% in 2023, with Anthropic capturing roughly 40%. Thirty-seven percent of firms now run five or more models — a multi-vendor strategy driven less by capability comparison than by exit optionality.

Key Data Points

These case studies are vendor-published and represent selected wins with no control group and no independent verification.

Four Categories of Customer-Owned AI Artifacts

Morgan Lewis’s February 2026 guidance identifies four categories that enterprise AI contracts must address at termination. Getting these wrong means losing assets that took months to build.

1. Customer data and inputs. Provider use must be limited to service performance, with express constraints on training rights. The contract should specify that the vendor cannot use customer inputs to improve its models without separate written consent.

2. Customer-developed artifacts. This is the category most buyers underestimate. It includes prompt libraries, workflows, evaluation datasets, retrieval indexes, vector embeddings, and guardrail configurations. These represent months of iteration by internal teams. Without explicit ownership language, they may be treated as part of the vendor’s platform.

3. Outputs. Customer rights to AI-generated outputs for business purposes — internal use and downstream workflows — must be stated explicitly. Many standard vendor agreements disclaim warranties on output accuracy and leave ownership ambiguous.

4. Deletion and return obligations. The vendor must return or delete all customer data and artifacts from its systems, including backups and subcontractor environments, with written certification of completion.

The Escrow Question: When You Can’t Take the Model

Fine-tuned model weights present a unique portability challenge. As one legal analysis puts it: once customer text, code, or interaction data is part of model weights, clean extraction is impossible — comparable to “retrieving sourdough starter after it’s baked into 200 loaves.”

For proprietary models (GPT-4, Claude, Gemini), vendors will rarely allow weights to leave their platform. The practical alternatives fall into two tiers:

Tier 1 — Direct escrow (where available):

• Deposit fine-tuned model weights with a neutral third party, updated quarterly
• Survivability clauses granting perpetual licenses to vendor-derived model artifacts
• Segregate customer data into customer-specific cloud buckets rather than pooled training environments

Tier 2 — Portability substitutes (when weights aren’t exportable):

• Export of all training and fine-tuning datasets (to the extent customer-owned)
• Export of prompt libraries and system instructions
• Evaluation datasets and testing results
• Performance and compliance logs and telemetry
• Workflow, guardrail, and configuration documentation

The gap between Tier 1 and Tier 2 is significant. With Tier 1, the customer can resume operations with the same model on a different platform. With Tier 2, the customer must rebuild the fine-tuned model from exported ingredients — a process that can take weeks to months and may not replicate the original performance.

The Bankruptcy Scenario Nobody Plans For

When an AI vendor enters insolvency, the standard SaaS playbook breaks down. Court-appointed trustees control vendor assets, and the results are predictable: APIs shut off overnight, models and datasets get auctioned to the highest bidder (potentially a competitor), and confidential training data becomes mixed into liquidated assets.

At least one documented case involved a fintech that lost six terabytes of training data when its text-analysis vendor shut down mid-quarter with no backup copies. The data was not recoverable.

Six contractual safeguards address this scenario before it occurs:

1. Escrow fine-tuned models with quarterly updates to a neutral third party
2. Segregate customer data into separate cloud tenants — not pooled
3. Include survivability clauses with perpetual licenses for model artifacts
4. Require transparency on hosting infrastructure (AWS region, tenant ID)
5. Maintain local quarterly exports of fine-tuning datasets
6. Request twelve-month cash-flow statements as part of vendor due diligence

The insurance gap compounds the risk. Typical cyber liability policies cover approximately $1 million — inadequate for most AI-dependent operations. Standard indemnities commonly exclude vendor business failure and cap coverage at pilot-phase fees.

What This Means for Your Organization

The 90–180-day vendor switching benchmark that Morgan Lewis recommends is a useful stress test. Ask your legal and procurement teams: if your primary AI vendor shut down in 90 days, what would survive? If the answer is “our prompts and some CSV exports,” the contract needs work.

Three actions for this quarter:

First, audit existing AI vendor contracts for the four Morgan Lewis artifact categories. Most mid-market companies signed their first AI agreements during 2024–2025 pilot programs, when leverage was low and urgency was high. Those terms deserve a second look now that the relationship is operational.

Second, negotiate escrow or quarterly export provisions before the next renewal. The leverage window is open: with OpenAI’s enterprise market share dropping from 50% to 27% and five-plus-model strategies becoming standard, vendors are more willing to negotiate retention terms than they were twelve months ago.

Third, build a vendor-failure scenario into the AI governance playbook alongside the data-breach and model-hallucination scenarios that already get attention. The controls are the same ones that protect against any exit — escrow, segregation, export automation — but framing them as business-continuity measures tends to unlock budget faster than framing them as contract hygiene.

If this raised questions specific to your organization’s AI vendor contracts, I’d welcome the conversation — brandon@brandonsneider.com

Sources

1. Morgan Lewis, “Building Exit Rights and Portability into AI Deals,” Tech & Sourcing @ Morgan Lewis, February 2026. https://www.morganlewis.com/blogs/sourcingatmorganlewis/2026/02/building-exit-rights-and-portability-into-ai-deals — Credibility: HIGH (AmLaw 50 technology sourcing practice; practitioner guidance, not vendor marketing)

2. Morgan Lewis, “Negotiating AI Provisions in Commercial and Technology Contracts: Where the Market Is Heading,” Tech & Sourcing @ Morgan Lewis, April 2026. https://www.morganlewis.com/blogs/sourcingatmorganlewis/2026/04/negotiating-ai-provisions-in-commercial-and-technology-contracts-where-the-market-is-heading — Credibility: HIGH

3. Redress Compliance, “OpenAI Enterprise Pricing 2026: Benchmarks & Negotiation Tactics,” 2026. https://redresscompliance.com/openai-enterprise-procurement-negotiation-playbook/ — Credibility: MEDIUM (industry benchmarking firm; single vendor focus)

4. OnlyLawyer.io, “What Happens When Your AI Provider Goes Bankrupt (and Takes Your Data With It),” 2026. https://www.onlylawyer.io/p/what-happens-when-your-ai-provider — Credibility: MEDIUM (legal analysis with case examples; no named parties in bankruptcy scenario)

5. Kai Waehner, “Enterprise Agentic AI Landscape 2026: Trust, Flexibility, and Vendor Lock-in,” April 6, 2026. https://www.kai-waehner.de/blog/2026/04/06/enterprise-agentic-ai-landscape-2026-trust-flexibility-and-vendor-lock-in/ — Credibility: MEDIUM (industry analyst; Menlo Ventures data cited is HIGH)

6. CloudPro, “AI Vendor Lock-In: How to Manage Enterprise Risk in 2026,” April 8, 2026. https://www.cloudproinc.com.au/index.php/2026/04/08/anthropic-openai-and-google-are-all-locking-in-enterprise-customers-how-to-manage-vendor-risk/ — Credibility: MEDIUM (aggregated industry data; survey methodology not disclosed)

Brandon Sneider | brandon@brandonsneider.com April 2026